init

README.md

@@ -1,8 +1,8 @@
 ---
 title: FreeTraj
-emoji: 🌖
-colorFrom: gray
-colorTo: red
+emoji: 🔥
+colorFrom: green
+colorTo: indigo
 sdk: gradio
 sdk_version: 4.37.2
 app_file: app.py

app.py

@@ -0,0 +1,682 @@
+import gradio as gr
+
+import sys
+import os
+import argparse
+import random
+from omegaconf import OmegaConf
+import torch
+import torchvision
+from pytorch_lightning import seed_everything
+from huggingface_hub import hf_hub_download
+
+sys.path.insert(0, "scripts/evaluation")
+from funcs import (
+    batch_ddim_sampling,
+    batch_ddim_sampling_freetraj,
+    load_model_checkpoint,
+)
+from utils.utils import instantiate_from_config
+
+MAX_KEYS = 5
+
+def check_move(trajectory, video_length=16):
+    traj_len = len(trajectory)
+    if traj_len < 2:
+        return False
+    prev_pos = trajectory[0]
+    for i in range(1, traj_len):
+        cur_pos = trajectory[i]
+        if cur_pos[0] > video_length - 1:
+            return False
+        if (cur_pos[0] - prev_pos[0]) * ((cur_pos[1] - prev_pos[1]) ** 2 + (cur_pos[2] - prev_pos[2]) ** 2) ** 0.5 < 0.02:
+            print("Too small movement, please use ori mode.")
+            return False
+        prev_pos = cur_pos
+
+    return True
+
+def infer(*user_args):
+    prompt_in = user_args[0]
+    target_indices = user_args[1]
+    ddim_edit = user_args[2]
+    seed = user_args[3]
+    ddim_steps = user_args[4]
+    unconditional_guidance_scale = user_args[5]
+    video_fps = user_args[6]
+    save_fps = user_args[7]
+    height_ratio = user_args[8]
+    width_ratio = user_args[9]
+    radio_mode = user_args[10]
+    dropdown_diy = user_args[11]
+    frame_indices = user_args[-3 * MAX_KEYS: -2 * MAX_KEYS]
+    h_positions = user_args[-2 * MAX_KEYS: -MAX_KEYS]
+    w_positions = user_args[-MAX_KEYS:]
+    print(user_args)
+
+    video_length = 16
+    width = 512 
+    height = 320
+    ckpt_dir_512 = "checkpoints/base_512_v2"
+    ckpt_path_512 = "checkpoints/base_512_v2/model.ckpt"
+    if radio_mode == 'ori':
+        config_512 = "configs/inference_t2v_512_v2.0.yaml"
+    else:
+        config_512 = "configs/inference_t2v_freetraj_512_v2.0.yaml"
+
+    trajectory = []
+    for i in range(dropdown_diy):
+        trajectory.append([int(frame_indices[i]), h_positions[i], w_positions[i]])
+    trajectory.sort()
+    print(trajectory)
+
+    if not check_move(trajectory):
+        print("Error trajectory.")
+
+    input_traj = []
+    h_remain = 1 - height_ratio
+    w_remain = 1 - width_ratio
+    for i in trajectory:
+        h_relative = i[1] * h_remain
+        w_relative = i[2] * w_remain
+        input_traj.append([i[0], h_relative, h_relative+height_ratio, w_relative, w_relative+width_ratio])
+
+    indices_list = target_indices.split(',')
+    idx_list = []
+    for i in indices_list:
+        idx_list.append(int(i))
+
+    config_512 = OmegaConf.load(config_512)
+    model_config_512 = config_512.pop("model", OmegaConf.create())
+    model = instantiate_from_config(model_config_512)
+    model = model.cuda()
+    if not os.path.exists(ckpt_path_512):
+        os.makedirs(ckpt_dir_512, exist_ok=True)
+        hf_hub_download(repo_id="VideoCrafter/VideoCrafter2", filename="model.ckpt", local_dir=ckpt_dir_512)
+    try:
+        model = load_model_checkpoint(model, ckpt_path_512)
+    except:
+        hf_hub_download(repo_id="VideoCrafter/VideoCrafter2", filename="model.ckpt", local_dir=ckpt_dir_512, force_download=True)
+        model = load_model_checkpoint(model, ckpt_path_512)
+    model.eval()
+
+    if seed is None:
+        seed = int.from_bytes(os.urandom(2), "big")
+    print(f"Using seed: {seed}")
+    seed_everything(seed)
+
+    args = argparse.Namespace(
+        mode="base",
+        savefps=save_fps,
+        n_samples=1,
+        ddim_steps=ddim_steps,
+        ddim_eta=0.0,
+        bs=1,
+        height=height,
+        width=width,
+        frames=video_length,
+        fps=video_fps,
+        unconditional_guidance_scale=unconditional_guidance_scale,
+        unconditional_guidance_scale_temporal=None,
+        cond_input=None,
+        ddim_edit = ddim_edit,
+    )
+
+    ## latent noise shape
+    h, w = args.height // 8, args.width // 8
+    frames = model.temporal_length if args.frames < 0 else args.frames
+    channels = model.channels
+
+    batch_size = 1
+    noise_shape = [batch_size, channels, frames, h, w]
+    fps = torch.tensor([args.fps] * batch_size).to(model.device).long()
+    prompts = [prompt_in]
+    text_emb = model.get_learned_conditioning(prompts)
+
+    cond = {"c_crossattn": [text_emb], "fps": fps}
+
+    ## inference
+    if radio_mode == 'ori':
+        batch_samples = batch_ddim_sampling(
+            model,
+            cond,
+            noise_shape,
+            args.n_samples,
+            args.ddim_steps,
+            args.ddim_eta,
+            args.unconditional_guidance_scale,
+            args=args,
+        )
+    else:
+        batch_samples = batch_ddim_sampling_freetraj(
+            model,
+            cond,
+            noise_shape,
+            args.n_samples,
+            args.ddim_steps,
+            args.ddim_eta,
+            args.unconditional_guidance_scale,
+            idx_list = idx_list,
+            input_traj = input_traj,
+            args=args,
+        )
+
+    vid_tensor = batch_samples[0]
+    video = vid_tensor.detach().cpu()
+    video = torch.clamp(video.float(), -1.0, 1.0)
+    video = video.permute(2, 0, 1, 3, 4)  # t,n,c,h,w
+
+    if radio_mode == 'ori':
+        video_path = "output.mp4"
+        video_bbox_path = "output.mp4"
+        frame_grids = [
+            torchvision.utils.make_grid(framesheet, nrow=int(args.n_samples))
+            for framesheet in video
+        ]  # [3, 1*h, n*w]
+        grid = torch.stack(frame_grids, dim=0)  # stack in temporal dim [t, 3, n*h, w]
+        grid = (grid + 1.0) / 2.0
+        grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+
+        torchvision.io.write_video(
+            video_path,
+            grid,
+            fps=args.savefps,
+            video_codec="h264",
+            options={"crf": "10"},
+        )
+    else:
+        video_path = "output_freetraj.mp4"
+        video_bbox_path = "output_freetraj_bbox.mp4"
+        frame_grids = [
+            torchvision.utils.make_grid(framesheet, nrow=int(args.n_samples))
+            for framesheet in video
+        ]  # [3, 1*h, n*w]
+        grid = torch.stack(frame_grids, dim=0)  # stack in temporal dim [t, 3, n*h, w]
+        grid = (grid + 1.0) / 2.0
+        grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+
+        torchvision.io.write_video(
+            video_path,
+            grid,
+            fps=args.savefps,
+            video_codec="h264",
+            options={"crf": "10"},
+        )
+
+        BOX_SIZE_H = input_traj[0][2] - input_traj[0][1]
+        BOX_SIZE_W = input_traj[0][4] - input_traj[0][3]
+        PATHS = plan_path(input_traj)
+        h_len = grid.shape[1]
+        w_len = grid.shape[2]
+        sub_h = int(BOX_SIZE_H * h_len) 
+        sub_w = int(BOX_SIZE_W * w_len)
+        for j in range(grid.shape[0]):
+            h_start = int(PATHS[j][0] * h_len)
+            h_end = h_start + sub_h
+            w_start = int(PATHS[j][2] * w_len)
+            w_end = w_start + sub_w
+
+            h_start = max(1, h_start)
+            h_end = min(h_len-1, h_end)
+            w_start = max(1, w_start)
+            w_end = min(w_len-1, w_end)
+
+            grid[j, h_start-1:h_end+1, w_start-1:w_start+2, :] = torch.ones_like(grid[j, h_start-1:h_end+1, w_start-1:w_start+2, :]) * torch.Tensor([127, 255, 127]).view(1, 1, 3)
+            grid[j, h_start-1:h_end+1, w_end-2:w_end+1, :] = torch.ones_like(grid[j, h_start-1:h_end+1, w_end-2:w_end+1, :]) * torch.Tensor([127, 255, 127]).view(1, 1, 3)
+            grid[j, h_start-1:h_start+2, w_start-1:w_end+1, :] = torch.ones_like(grid[j, h_start-1:h_start+2, w_start-1:w_end+1, :]) * torch.Tensor([127, 255, 127]).view(1, 1, 3)
+            grid[j, h_end-2:h_end+1, w_start-1:w_end+1, :] = torch.ones_like(grid[j, h_end-2:h_end+1, w_start-1:w_end+1, :]) * torch.Tensor([127, 255, 127]).view(1, 1, 3)
+
+        torchvision.io.write_video(
+            video_bbox_path,
+            grid,
+            fps=args.savefps,
+            video_codec="h264",
+            options={"crf": "10"},
+        )
+
+    return video_path, video_bbox_path
+
+
+examples = [
+    ["A squirrel jumping from one tree to another.",],
+    ["A bear climbing down a tree after spotting a threat.",],
+    ["A deer walking in a snowy field.",],
+    ["A lion walking in the savanna grass.",],
+    ["A kangaroo jumping in the Australian outback.",],
+    ["A corgi running on the grassland on the grassland.",],
+    ["A horse galloping on a street.",],
+    ["A majestic eagle soaring high above the treetops, surveying its territory.",],
+]
+
+css = """
+#col-container {max-width: 1024px; margin-left: auto; margin-right: auto;}
+a {text-decoration-line: underline; font-weight: 600;}
+.animate-spin {
+  animation: spin 1s linear infinite;
+}
+#share-btn-container {
+  display: flex; 
+  padding-left: 0.5rem !important; 
+  padding-right: 0.5rem !important; 
+  background-color: #000000; 
+  justify-content: center; 
+  align-items: center; 
+  border-radius: 9999px !important; 
+  max-width: 15rem;
+  height: 36px;
+}
+div#share-btn-container > div {
+    flex-direction: row;
+    background: black;
+    align-items: center;
+}
+#share-btn-container:hover {
+  background-color: #060606;
+}
+#share-btn {
+  all: initial; 
+  color: #ffffff;
+  font-weight: 600; 
+  cursor:pointer; 
+  font-family: 'IBM Plex Sans', sans-serif; 
+  margin-left: 0.5rem !important; 
+  padding-top: 0.5rem !important; 
+  padding-bottom: 0.5rem !important;
+  right:0;
+}
+#share-btn * {
+  all: unset;
+}
+#share-btn-container div:nth-child(-n+2){
+  width: auto !important;
+  min-height: 0px !important;
+}
+#share-btn-container .wrap {
+  display: none !important;
+}
+#share-btn-container.hidden {
+  display: none!important;
+}
+img[src*='#center'] { 
+    display: inline-block;
+    margin: unset;
+}
+.footer {
+        margin-bottom: 45px;
+        margin-top: 10px;
+        text-align: center;
+        border-bottom: 1px solid #e5e5e5;
+    }
+    .footer>p {
+        font-size: .8rem;
+        display: inline-block;
+        padding: 0 10px;
+        transform: translateY(10px);
+        background: white;
+    }
+    .dark .footer {
+        border-color: #303030;
+    }
+    .dark .footer>p {
+        background: #0b0f19;
+    }
+"""
+
+def mode_update(mode):
+    if mode == 'demo':
+        trajectories_mode = [gr.Row(visible=True), gr.Row(visible=False)]
+    elif mode == 'diy':
+        trajectories_mode = [gr.Row(visible=False), gr.Row(visible=True)]
+    else:
+        trajectories_mode = [gr.Row(visible=False), gr.Row(visible=False)]
+    return trajectories_mode
+
+def keyframe_update(num):
+    keyframes = []
+    if type(num) != int:
+        num = 0
+
+    for i in range(num):
+        keyframes.append(gr.Row(visible=True))
+    for i in range(MAX_KEYS - num):
+        keyframes.append(gr.Row(visible=False))
+    return keyframes
+
+def demo_update(mode):
+    if mode == 'topleft->bottomright':
+        num = 2
+    elif mode == 'bottomleft->topright':
+        num = 2
+    elif mode == 'topleft->bottomleft->bottomright':
+        num = 3
+    elif mode == 'bottomright->topright->topleft':
+        num = 3
+    elif mode == '"V"':
+        num = 4
+    elif mode == '"^"':
+        num = 4
+    elif mode == 'left->right->left->right':
+        num = 4
+    elif mode == 'triangle':
+        num = 4
+    else:
+        num = 0
+
+    return num
+
+def demo_update_frame(mode):
+    frame_indices = []
+    if mode == 'topleft->bottomright':
+        num = 2
+        frame_indices.append(gr.Text(value=0))
+        frame_indices.append(gr.Text(value=15))
+    elif mode == 'bottomleft->topright':
+        num = 2
+        frame_indices.append(gr.Text(value=0))
+        frame_indices.append(gr.Text(value=15))
+    elif mode == 'topleft->bottomleft->bottomright':
+        num = 3
+        frame_indices.append(gr.Text(value=0))
+        frame_indices.append(gr.Text(value=9))
+        frame_indices.append(gr.Text(value=15))
+    elif mode == 'bottomright->topright->topleft':
+        num = 3
+        frame_indices.append(gr.Text(value=0))
+        frame_indices.append(gr.Text(value=6))
+        frame_indices.append(gr.Text(value=15))
+    elif mode == '"V"':
+        num = 4
+        frame_indices.append(gr.Text(value=0))
+        frame_indices.append(gr.Text(value=7))
+        frame_indices.append(gr.Text(value=8))
+        frame_indices.append(gr.Text(value=15))
+    elif mode == '"^"':
+        num = 4
+        frame_indices.append(gr.Text(value=0))
+        frame_indices.append(gr.Text(value=7))
+        frame_indices.append(gr.Text(value=8))
+        frame_indices.append(gr.Text(value=15))
+    elif mode == 'left->right->left->right':
+        num = 4
+        frame_indices.append(gr.Text(value=0))
+        frame_indices.append(gr.Text(value=5))
+        frame_indices.append(gr.Text(value=10))
+        frame_indices.append(gr.Text(value=15))
+    elif mode == 'triangle':
+        num = 4
+        frame_indices.append(gr.Text(value=0))
+        frame_indices.append(gr.Text(value=5))
+        frame_indices.append(gr.Text(value=10))
+        frame_indices.append(gr.Text(value=15))
+    else:
+        num = 0
+
+    for i in range(MAX_KEYS - num):
+        frame_indices.append(gr.Text())
+    return frame_indices
+
+def demo_update_h(mode):
+    h_positions = []
+    if mode == 'topleft->bottomright':
+        num = 2
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.9))
+    elif mode == 'bottomleft->topright':
+        num = 2
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.1))
+    elif mode == 'topleft->bottomleft->bottomright':
+        num = 3
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.9))
+    elif mode == 'bottomright->topright->topleft':
+        num = 3
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.1))
+    elif mode == '"V"':
+        num = 4
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.1))
+    elif mode == '"^"':
+        num = 4
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.9))
+    elif mode == 'left->right->left->right':
+        num = 4
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.9))
+    elif mode == 'triangle':
+        num = 4
+        h_positions.append(gr.Slider(value=0.1))
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.9))
+        h_positions.append(gr.Slider(value=0.1))
+    else:
+        num = 0
+
+    for i in range(MAX_KEYS - num):
+        h_positions.append(gr.Slider())
+    return h_positions
+
+def demo_update_w(mode):
+    w_positions = []
+    if mode == 'topleft->bottomright':
+        num = 2
+        w_positions.append(gr.Slider(value=0.1))
+        w_positions.append(gr.Slider(value=0.9))
+    elif mode == 'bottomleft->topright':
+        num = 2
+        w_positions.append(gr.Slider(value=0.1))
+        w_positions.append(gr.Slider(value=0.9))
+    elif mode == 'topleft->bottomleft->bottomright':
+        num = 3
+        w_positions.append(gr.Slider(value=0.1))
+        w_positions.append(gr.Slider(value=0.1))
+        w_positions.append(gr.Slider(value=0.9))
+    elif mode == 'bottomright->topright->topleft':
+        num = 3
+        w_positions.append(gr.Slider(value=0.9))
+        w_positions.append(gr.Slider(value=0.9))
+        w_positions.append(gr.Slider(value=0.1))
+    elif mode == '"V"':
+        num = 4
+        w_positions.append(gr.Slider(value=0.1))
+        w_positions.append(gr.Slider(value=0.8/15*7 + 0.1))
+        w_positions.append(gr.Slider(value=0.8/15*8 + 0.1))
+        w_positions.append(gr.Slider(value=0.9))
+    elif mode == '"^"':
+        num = 4
+        w_positions.append(gr.Slider(value=0.9))
+        w_positions.append(gr.Slider(value=0.8/15*8 + 0.1))
+        w_positions.append(gr.Slider(value=0.8/15*7 + 0.1))
+        w_positions.append(gr.Slider(value=0.1))
+    elif mode == 'left->right->left->right':
+        num = 4
+        w_positions.append(gr.Slider(value=0.5))
+        w_positions.append(gr.Slider(value=0.5))
+        w_positions.append(gr.Slider(value=0.5))
+        w_positions.append(gr.Slider(value=0.5))
+    elif mode == 'triangle':
+        num = 4
+        w_positions.append(gr.Slider(value=0.5))
+        w_positions.append(gr.Slider(value=0.9))
+        w_positions.append(gr.Slider(value=0.1))
+        w_positions.append(gr.Slider(value=0.5))
+    else:
+        num = 0
+
+    for i in range(MAX_KEYS - num):
+        w_positions.append(gr.Slider())
+    return w_positions
+
+with gr.Blocks(css=css) as demo:
+    with gr.Column(elem_id="col-container"):
+        gr.Markdown(
+            """
+            <h1 style="text-align: center;">FreeTraj</h1>
+            <p style="text-align: center;">
+            Tuning-Free Trajectory Control in Video Diffusion Models
+            </p>
+            <p style="text-align: center;">
+            <a href="https://arxiv.org/abs/2406.16863" target="_blank"><b>[arXiv]</b></a> &nbsp;&nbsp;&nbsp;&nbsp;
+            <a href="http://haonanqiu.com/projects/FreeTraj.html" target="_blank"><b>[Project Page]</b></a> &nbsp;&nbsp;&nbsp;&nbsp;
+            <a href="https://github.com/arthur-qiu/FreeTraj" target="_blank"><b>[Code]</b></a>
+            </p>         
+            """
+        )
+
+        keyframes = []
+        frame_indices = []
+        h_positions = []
+        w_positions = []
+
+        with gr.Row():
+            video_result = gr.Video(label="Video Output")
+            video_result_bbox = gr.Video(label="Video Output with BBox")
+
+        with gr.Row():
+            prompt_in = gr.Textbox(label="Prompt", placeholder="A corgi running on the grassland on the grassland.", scale = 3)
+            target_indices = gr.Textbox(label="Target Indices", placeholder="1,2", scale = 1)
+
+        with gr.Row():
+            radio_mode = gr.Radio(label='Trajectory Mode', choices = ['demo', 'diy', 'ori'], scale = 1)
+            height_ratio = gr.Slider(label='Height Ratio of BBox',
+                              minimum=0.2,
+                              maximum=0.4,
+                              step=0.01,
+                              value=0.3,
+                              scale = 1)
+            width_ratio = gr.Slider(label='Width Ratio of BBox',
+                              minimum=0.2,
+                              maximum=0.4,
+                              step=0.01,
+                              value=0.3, 
+                              scale = 1)
+          
+        with gr.Row(visible=False) as row_demo:
+            dropdown_demo = gr.Dropdown(
+                label="Demo Trajectory",
+                choices= ['topleft->bottomright', 'bottomleft->topright', 'topleft->bottomleft->bottomright', 'bottomright->topright->topleft', '"V"', '"^"', 'left->right->left->right', 'triangle']
+            )
+            
+        with gr.Row(visible=False) as row_diy:
+            dropdown_diy = gr.Dropdown(
+                label="Number of keyframes",
+                choices=range(2, MAX_KEYS+1),
+            )
+            
+        for i in range(MAX_KEYS):
+            with gr.Row(visible=False) as row:
+                text = f"Keyframe #{i}"
+                text = gr.HTML(text, visible=True)
+                frame_ids = gr.Textbox(
+                    None,
+                    label=f"Frame Indices #{i}",
+                    visible=True,
+                    interactive=True,
+                    scale=1
+                )
+                h_position = gr.Slider(label='Position in Height',
+                    minimum=0.0,
+                    maximum=1.0,
+                    step=0.01,
+                    scale=1)
+                w_position = gr.Slider(label='Position in Width',
+                    minimum=0.0,
+                    maximum=1.0,
+                    step=0.01,
+                    scale=1)
+                
+            frame_indices.append(frame_ids)
+            h_positions.append(h_position)
+            w_positions.append(w_position)
+            keyframes.append(row)
+
+
+        dropdown_demo.change(demo_update, dropdown_demo, dropdown_diy)
+        dropdown_diy.change(keyframe_update, dropdown_diy, keyframes)
+        dropdown_demo.change(demo_update_frame, dropdown_demo, frame_indices)
+        dropdown_demo.change(demo_update_h, dropdown_demo, h_positions)
+        dropdown_demo.change(demo_update_w, dropdown_demo, w_positions)
+        radio_mode.change(mode_update, radio_mode, [row_demo, row_diy])
+
+
+        with gr.Row():
+            with gr.Accordion('Useful FreeTraj Parameters (feel free to adjust these parameters based on your prompt): ', open=True):
+                with gr.Row():
+                    ddim_edit = gr.Slider(label='Editing Steps (larger for better control while losing some quality)',
+                             minimum=0,
+                             maximum=12,
+                             step=1,
+                             value=6)
+                    seed = gr.Slider(label='Random Seed',
+                             minimum=0,
+                             maximum=10000,
+                             step=1,
+                             value=123)
+                    
+        with gr.Row():
+            with gr.Accordion('Useless FreeTraj Parameters (mostly no need to adjust): ', open=False):
+                with gr.Row():
+                    ddim_steps = gr.Slider(label='DDIM Steps',
+                              minimum=5,
+                              maximum=200,
+                              step=1,
+                              value=50)
+                    unconditional_guidance_scale = gr.Slider(label='Unconditional Guidance Scale',
+                              minimum=1.0,
+                              maximum=20.0,
+                              step=0.1,
+                              value=12.0)
+                with gr.Row():
+                    video_fps = gr.Slider(label='Video FPS (larger for quicker motion)',
+                              minimum=8,
+                              maximum=36,
+                              step=4,
+                              value=16)
+                    save_fps = gr.Slider(label='Save FPS',
+                              minimum=1,
+                              maximum=30,
+                              step=1,
+                              value=10)
+                
+        with gr.Row():
+            submit_btn = gr.Button("Generate", variant='primary')
+
+        with gr.Row():
+            gr.Examples(label='Sample Prompts', examples=examples, inputs=[prompt_in, target_indices, ddim_edit, seed, ddim_steps, unconditional_guidance_scale, video_fps, save_fps, height_ratio, width_ratio, radio_mode, dropdown_diy, *frame_indices, *h_positions, *w_positions])
+
+        with gr.Row():
+            gr.Markdown(
+                """
+                <h2 style="text-align: center;">Hints</h2>
+                <p style="text-align: center;">
+                1. Choose trajectory mode <b>"ori"</b> to see whether the prompt works on the pre-trained model. 
+                </p>    
+                <p style="text-align: center;">
+                2. Adjust the prompt or random seed to get a qualified video.
+                </p>  
+                <p style="text-align: center;">
+                3. Choose trajectory mode <b>"demo"</b> to see whether <b>FreeTraj</b> works or not.
+                </p>  
+                <p style="text-align: center;">
+                4. Choose trajectory mode <b>"diy"</b> to plan new trajectory. It may fail in some extreme cases.
+                </p>       
+                """
+            )
+            
+
+    submit_btn.click(fn=infer,
+            inputs=[prompt_in, target_indices, ddim_edit, seed, ddim_steps, unconditional_guidance_scale, video_fps, save_fps, height_ratio, width_ratio, radio_mode, dropdown_diy, *frame_indices, *h_positions, *w_positions],
+            outputs=[video_result, video_result_bbox],
+            api_name="zrscp")
+
+demo.queue(max_size=12).launch(show_api=True)

configs/inference_i2v_512_v1.0.yaml

@@ -0,0 +1,83 @@
+model:
+  target: lvdm.models.ddpm3d.LatentVisualDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    image_size:
+    - 40
+    - 64
+    channels: 4
+    scale_by_std: false
+    scale_factor: 0.18215
+    use_ema: false
+    uncond_type: empty_seq
+    use_scale: true
+    scale_b: 0.7
+    finegrained: true
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d.UNetModel
+      params:
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: true
+        temporal_conv: true
+        temporal_attention: true
+        temporal_selfatt_only: true
+        use_relative_position: false
+        use_causal_attention: false
+        use_image_attention: true
+        temporal_length: 16
+        addition_attention: true
+        fps_cond: true
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 512
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: penultimate
+    cond_img_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPImageEmbedderV2
+      params:
+        freeze: true

configs/inference_t2v_1024_v1.0.yaml

@@ -0,0 +1,77 @@
+model:
+  target: lvdm.models.ddpm3d.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    image_size:
+    - 72
+    - 128
+    channels: 4
+    scale_by_std: false
+    scale_factor: 0.18215
+    use_ema: false
+    uncond_type: empty_seq
+    use_scale: true
+    fix_scale_bug: true
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d.UNetModel
+      params:
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: true
+        temporal_conv: false
+        temporal_attention: true
+        temporal_selfatt_only: true
+        use_relative_position: true
+        use_causal_attention: false
+        temporal_length: 16
+        addition_attention: true
+        fps_cond: true
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 512
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: penultimate

configs/inference_t2v_512_v1.0.yaml

@@ -0,0 +1,74 @@
+model:
+  target: lvdm.models.ddpm3d.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    image_size:
+    - 40
+    - 64
+    channels: 4
+    scale_by_std: false
+    scale_factor: 0.18215
+    use_ema: false
+    uncond_type: empty_seq
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d.UNetModel
+      params:
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: true
+        temporal_conv: false
+        temporal_attention: true
+        temporal_selfatt_only: true
+        use_relative_position: true
+        use_causal_attention: false
+        temporal_length: 16
+        addition_attention: true
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 512
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: penultimate

configs/inference_t2v_512_v2.0.yaml

@@ -0,0 +1,77 @@
+model:
+  target: lvdm.models.ddpm3d.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    image_size:
+    - 40
+    - 64
+    channels: 4
+    scale_by_std: false
+    scale_factor: 0.18215
+    use_ema: false
+    uncond_type: empty_seq
+    use_scale: true
+    scale_b: 0.7
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d.UNetModel
+      params:
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: true
+        temporal_conv: true
+        temporal_attention: true
+        temporal_selfatt_only: true
+        use_relative_position: false
+        use_causal_attention: false
+        temporal_length: 16
+        addition_attention: true
+        fps_cond: true
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 512
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: penultimate

configs/inference_t2v_freetraj_512_v2.0.yaml

@@ -0,0 +1,77 @@
+model:
+  target: lvdm.models.ddpm3d.LatentDiffusion
+  params:
+    linear_start: 0.00085
+    linear_end: 0.012
+    num_timesteps_cond: 1
+    timesteps: 1000
+    first_stage_key: video
+    cond_stage_key: caption
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    image_size:
+    - 40
+    - 64
+    channels: 4
+    scale_by_std: false
+    scale_factor: 0.18215
+    use_ema: false
+    uncond_type: empty_seq
+    use_scale: true
+    scale_b: 0.7
+    unet_config:
+      target: lvdm.modules.networks.openaimodel3d_freetraj.UNetModel
+      params:
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 4
+        - 2
+        - 1
+        num_res_blocks: 2
+        channel_mult:
+        - 1
+        - 2
+        - 4
+        - 4
+        num_head_channels: 64
+        transformer_depth: 1
+        context_dim: 1024
+        use_linear: true
+        use_checkpoint: true
+        temporal_conv: true
+        temporal_attention: true
+        temporal_selfatt_only: true
+        use_relative_position: false
+        use_causal_attention: false
+        temporal_length: 16
+        addition_attention: true
+        fps_cond: true
+    first_stage_config:
+      target: lvdm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 512
+          in_channels: 3
+          out_ch: 3
+          ch: 128
+          ch_mult:
+          - 1
+          - 2
+          - 4
+          - 4
+          num_res_blocks: 2
+          attn_resolutions: []
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+    cond_stage_config:
+      target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
+      params:
+        freeze: true
+        layer: penultimate

lvdm/basics.py

@@ -0,0 +1,100 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+import torch.nn as nn
+from utils.utils import instantiate_from_config
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def nonlinearity(type='silu'):
+    if type == 'silu':
+        return nn.SiLU()
+    elif type == 'leaky_relu':
+        return nn.LeakyReLU()
+
+
+class GroupNormSpecific(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+
+def normalization(channels, num_groups=32):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNormSpecific(num_groups, channels)
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}

lvdm/common.py

@@ -0,0 +1,95 @@
+import math
+from inspect import isfunction
+import torch
+from torch import nn
+import torch.distributed as dist
+
+
+def gather_data(data, return_np=True):
+    ''' gather data from multiple processes to one list '''
+    data_list = [torch.zeros_like(data) for _ in range(dist.get_world_size())]
+    dist.all_gather(data_list, data)  # gather not supported with NCCL
+    if return_np:
+        data_list = [data.cpu().numpy() for data in data_list]
+    return data_list
+
+def autocast(f):
+    def do_autocast(*args, **kwargs):
+        with torch.cuda.amp.autocast(enabled=True,
+                                     dtype=torch.get_autocast_gpu_dtype(),
+                                     cache_enabled=torch.is_autocast_cache_enabled()):
+            return f(*args, **kwargs)
+    return do_autocast
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+def exists(val):
+    return val is not None
+
+def identity(*args, **kwargs):
+    return nn.Identity()
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+def ismap(x):
+    if not isinstance(x, torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] > 3)
+
+def isimage(x):
+    if not isinstance(x,torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+def shape_to_str(x):
+    shape_str = "x".join([str(x) for x in x.shape])
+    return shape_str
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+ckpt = torch.utils.checkpoint.checkpoint
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        return ckpt(func, *inputs)
+    else:
+        return func(*inputs)
+

lvdm/distributions.py

@@ -0,0 +1,95 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self, noise=None):
+        if noise is None:
+            noise = torch.randn(self.mean.shape)
+        
+        x = self.mean + self.std * noise.to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

lvdm/ema.py

@@ -0,0 +1,76 @@
+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError('Decay must be between 0 and 1')
+
+        self.m_name2s_name = {}
+        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates
+                             else torch.tensor(-1,dtype=torch.int))
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                #remove as '.'-character is not allowed in buffers
+                s_name = name.replace('.','')
+                self.m_name2s_name.update({name:s_name})
+                self.register_buffer(s_name,p.clone().detach().data)
+
+        self.collected_params = []
+
+    def forward(self,model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

lvdm/models/autoencoder.py

@@ -0,0 +1,219 @@
+import os
+from contextlib import contextmanager
+import torch
+import numpy as np
+from einops import rearrange
+import torch.nn.functional as F
+import pytorch_lightning as pl
+from lvdm.modules.networks.ae_modules import Encoder, Decoder
+from lvdm.distributions import DiagonalGaussianDistribution
+from utils.utils import instantiate_from_config
+
+
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 test=False,
+                 logdir=None,
+                 input_dim=4,
+                 test_args=None,
+                 ):
+        super().__init__()
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+        self.input_dim = input_dim
+        self.test = test
+        self.test_args = test_args
+        self.logdir = logdir
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        if self.test:
+            self.init_test()
+    
+    def init_test(self,):
+        self.test = True
+        save_dir = os.path.join(self.logdir, "test")
+        if 'ckpt' in self.test_args:
+            ckpt_name = os.path.basename(self.test_args.ckpt).split('.ckpt')[0] + f'_epoch{self._cur_epoch}'
+            self.root = os.path.join(save_dir, ckpt_name)
+        else:
+            self.root = save_dir
+        if 'test_subdir' in self.test_args:
+            self.root = os.path.join(save_dir, self.test_args.test_subdir)
+
+        self.root_zs = os.path.join(self.root, "zs")
+        self.root_dec = os.path.join(self.root, "reconstructions")
+        self.root_inputs = os.path.join(self.root, "inputs")
+        os.makedirs(self.root, exist_ok=True)
+
+        if self.test_args.save_z:
+            os.makedirs(self.root_zs, exist_ok=True)
+        if self.test_args.save_reconstruction:
+            os.makedirs(self.root_dec, exist_ok=True)
+        if self.test_args.save_input:
+            os.makedirs(self.root_inputs, exist_ok=True)
+        assert(self.test_args is not None)
+        self.test_maximum = getattr(self.test_args, 'test_maximum', None) 
+        self.count = 0
+        self.eval_metrics = {}
+        self.decodes = []
+        self.save_decode_samples = 2048
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")
+        try:
+            self._cur_epoch = sd['epoch']
+            sd = sd["state_dict"]
+        except:
+            self._cur_epoch = 'null'
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        self.load_state_dict(sd, strict=False)
+        # self.load_state_dict(sd, strict=True)
+        print(f"Restored from {path}")
+
+    def encode(self, x, **kwargs):
+        
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z, **kwargs):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        return dec
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        return dec, posterior
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if x.dim() == 5 and self.input_dim == 4:
+            b,c,t,h,w = x.shape
+            self.b = b
+            self.t = t 
+            x = rearrange(x, 'b c t h w -> (b t) c h w')
+
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+
+        if optimizer_idx == 0:
+            # train encoder+decoder+logvar
+            aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # train the discriminator
+            discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                                last_layer=self.get_last_layer(), split="train")
+
+            self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+        aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
+                                        last_layer=self.get_last_layer(), split="val")
+
+        discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
+                                            last_layer=self.get_last_layer(), split="val")
+
+        self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+    
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr, betas=(0.5, 0.9))
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    @torch.no_grad()
+    def log_images(self, batch, only_inputs=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if not only_inputs:
+            xrec, posterior = self(x)
+            if x.shape[1] > 3:
+                # colorize with random projection
+                assert xrec.shape[1] > 3
+                x = self.to_rgb(x)
+                xrec = self.to_rgb(xrec)
+            log["samples"] = self.decode(torch.randn_like(posterior.sample()))
+            log["reconstructions"] = xrec
+        log["inputs"] = x
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+class IdentityFirstStage(torch.nn.Module):
+    def __init__(self, *args, vq_interface=False, **kwargs):
+        self.vq_interface = vq_interface  # TODO: Should be true by default but check to not break older stuff
+        super().__init__()
+
+    def encode(self, x, *args, **kwargs):
+        return x
+
+    def decode(self, x, *args, **kwargs):
+        return x
+
+    def quantize(self, x, *args, **kwargs):
+        if self.vq_interface:
+            return x, None, [None, None, None]
+        return x
+
+    def forward(self, x, *args, **kwargs):
+        return x

lvdm/models/ddpm3d.py

@@ -0,0 +1,763 @@
+"""
+wild mixture of
+https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
+https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+https://github.com/CompVis/taming-transformers
+-- merci
+"""
+
+from functools import partial
+from contextlib import contextmanager
+import numpy as np
+from tqdm import tqdm
+from einops import rearrange, repeat
+import logging
+mainlogger = logging.getLogger('mainlogger')
+import torch
+import torch.nn as nn
+from torchvision.utils import make_grid
+import pytorch_lightning as pl
+from utils.utils import instantiate_from_config
+from lvdm.ema import LitEma
+from lvdm.distributions import DiagonalGaussianDistribution
+from lvdm.models.utils_diffusion import make_beta_schedule
+from lvdm.modules.encoders.ip_resampler import ImageProjModel, Resampler
+from lvdm.basics import disabled_train
+from lvdm.common import (
+    extract_into_tensor,
+    noise_like,
+    exists,
+    default
+)
+
+
+__conditioning_keys__ = {'concat': 'c_concat',
+                         'crossattn': 'c_crossattn',
+                         'adm': 'y'}
+
+class DDPM(pl.LightningModule):
+    # classic DDPM with Gaussian diffusion, in image space
+    def __init__(self,
+                 unet_config,
+                 timesteps=1000,
+                 beta_schedule="linear",
+                 loss_type="l2",
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 load_only_unet=False,
+                 monitor=None,
+                 use_ema=True,
+                 first_stage_key="image",
+                 image_size=256,
+                 channels=3,
+                 log_every_t=100,
+                 clip_denoised=True,
+                 linear_start=1e-4,
+                 linear_end=2e-2,
+                 cosine_s=8e-3,
+                 given_betas=None,
+                 original_elbo_weight=0.,
+                 v_posterior=0.,  # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
+                 l_simple_weight=1.,
+                 conditioning_key=None,
+                 parameterization="eps",  # all assuming fixed variance schedules
+                 scheduler_config=None,
+                 use_positional_encodings=False,
+                 learn_logvar=False,
+                 logvar_init=0.
+                 ):
+        super().__init__()
+        assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
+        self.parameterization = parameterization
+        mainlogger.info(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
+        self.cond_stage_model = None
+        self.clip_denoised = clip_denoised
+        self.log_every_t = log_every_t
+        self.first_stage_key = first_stage_key
+        self.channels = channels
+        self.temporal_length = unet_config.params.temporal_length
+        self.image_size = image_size 
+        if isinstance(self.image_size, int):
+            self.image_size = [self.image_size, self.image_size]
+        self.use_positional_encodings = use_positional_encodings
+        self.model = DiffusionWrapper(unet_config, conditioning_key)
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self.model)
+            mainlogger.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        self.use_scheduler = scheduler_config is not None
+        if self.use_scheduler:
+            self.scheduler_config = scheduler_config
+
+        self.v_posterior = v_posterior
+        self.original_elbo_weight = original_elbo_weight
+        self.l_simple_weight = l_simple_weight
+
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
+
+        self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
+                               linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
+
+        self.loss_type = loss_type
+
+        self.learn_logvar = learn_logvar
+        self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
+        if self.learn_logvar:
+            self.logvar = nn.Parameter(self.logvar, requires_grad=True)
+
+
+    def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        if exists(given_betas):
+            betas = given_betas
+        else:
+            betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                       cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+        # calculations for posterior q(x_{t-1} | x_t, x_0)
+        posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
+                    1. - alphas_cumprod) + self.v_posterior * betas
+        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
+        self.register_buffer('posterior_variance', to_torch(posterior_variance))
+        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
+        self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
+        self.register_buffer('posterior_mean_coef1', to_torch(
+            betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
+        self.register_buffer('posterior_mean_coef2', to_torch(
+            (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
+
+        if self.parameterization == "eps":
+            lvlb_weights = self.betas ** 2 / (
+                        2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
+        elif self.parameterization == "x0":
+            lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
+        else:
+            raise NotImplementedError("mu not supported")
+        # TODO how to choose this term
+        lvlb_weights[0] = lvlb_weights[1]
+        self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
+        assert not torch.isnan(self.lvlb_weights).all()
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.model.parameters())
+            self.model_ema.copy_to(self.model)
+            if context is not None:
+                mainlogger.info(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.model.parameters())
+                if context is not None:
+                    mainlogger.info(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    mainlogger.info("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        mainlogger.info(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            mainlogger.info(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            mainlogger.info(f"Unexpected Keys: {unexpected}")
+
+    def q_mean_variance(self, x_start, t):
+        """
+        Get the distribution q(x_t | x_0).
+        :param x_start: the [N x C x ...] tensor of noiseless inputs.
+        :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
+        :return: A tuple (mean, variance, log_variance), all of x_start's shape.
+        """
+        mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
+        variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
+        log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
+        return mean, variance, log_variance
+
+    def predict_start_from_noise(self, x_t, t, noise):
+        return (
+                extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
+                extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
+        )
+
+    def q_posterior(self, x_start, x_t, t):
+        posterior_mean = (
+                extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
+                extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
+        )
+        posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
+        posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
+        return posterior_mean, posterior_variance, posterior_log_variance_clipped
+
+    def p_mean_variance(self, x, t, clip_denoised: bool):
+        model_out = self.model(x, t)
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
+        b, *_, device = *x.shape, x.device
+        model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
+        noise = noise_like(x.shape, device, repeat_noise)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, shape, return_intermediates=False):
+        device = self.betas.device
+        b = shape[0]
+        img = torch.randn(shape, device=device)
+        intermediates = [img]
+        for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
+            img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
+                                clip_denoised=self.clip_denoised)
+            if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
+                intermediates.append(img)
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, batch_size=16, return_intermediates=False):
+        image_size = self.image_size
+        channels = self.channels
+        return self.p_sample_loop((batch_size, channels, image_size, image_size),
+                                  return_intermediates=return_intermediates)
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start *
+                extract_into_tensor(self.scale_arr, t, x_start.shape) +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def _get_rows_from_list(self, samples):
+        n_imgs_per_row = len(samples)
+        denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.first_stage_key)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        x = x.to(self.device)[:N]
+        log["inputs"] = x
+
+        # get diffusion row
+        diffusion_row = list()
+        x_start = x[:n_row]
+
+        for t in range(self.num_timesteps):
+            if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                t = t.to(self.device).long()
+                noise = torch.randn_like(x_start)
+                x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+                diffusion_row.append(x_noisy)
+
+        log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
+
+            log["samples"] = samples
+            log["denoise_row"] = self._get_rows_from_list(denoise_row)
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+
+class LatentDiffusion(DDPM):
+    """main class"""
+    def __init__(self,
+                 first_stage_config,
+                 cond_stage_config,
+                 num_timesteps_cond=None,
+                 cond_stage_key="caption",
+                 cond_stage_trainable=False,
+                 cond_stage_forward=None,
+                 conditioning_key=None,
+                 uncond_prob=0.2,
+                 uncond_type="empty_seq",
+                 scale_factor=1.0,
+                 scale_by_std=False,
+                 encoder_type="2d",
+                 only_model=False,
+                 use_scale=False,
+                 scale_a=1,
+                 scale_b=0.3,
+                 mid_step=400,
+                 fix_scale_bug=False,
+                 *args, **kwargs):
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs['timesteps']
+        # for backwards compatibility after implementation of DiffusionWrapper
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        conditioning_key = default(conditioning_key, 'crossattn')
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+
+        self.cond_stage_trainable = cond_stage_trainable
+        self.cond_stage_key = cond_stage_key
+
+        # scale factor
+        self.use_scale=use_scale
+        if self.use_scale:
+            self.scale_a=scale_a
+            self.scale_b=scale_b
+            if fix_scale_bug:
+                scale_step=self.num_timesteps-mid_step
+            else: #bug
+                scale_step = self.num_timesteps
+
+            scale_arr1 = np.linspace(scale_a, scale_b, mid_step)
+            scale_arr2 = np.full(scale_step, scale_b)
+            scale_arr = np.concatenate((scale_arr1, scale_arr2))
+            scale_arr_prev = np.append(scale_a, scale_arr[:-1])
+            to_torch = partial(torch.tensor, dtype=torch.float32)
+            self.register_buffer('scale_arr', to_torch(scale_arr))
+
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer('scale_factor', torch.tensor(scale_factor))
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.first_stage_config = first_stage_config
+        self.cond_stage_config = cond_stage_config        
+        self.clip_denoised = False
+
+        self.cond_stage_forward = cond_stage_forward
+        self.encoder_type = encoder_type
+        assert(encoder_type in ["2d", "3d"])
+        self.uncond_prob = uncond_prob
+        self.classifier_free_guidance = True if uncond_prob > 0 else False
+        assert(uncond_type in ["zero_embed", "empty_seq"])
+        self.uncond_type = uncond_type
+
+
+        self.restarted_from_ckpt = False
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys, only_model=only_model)
+            self.restarted_from_ckpt = True
+                
+
+    def make_cond_schedule(self, ):
+        self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
+        ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
+        self.cond_ids[:self.num_timesteps_cond] = ids
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        if self.use_scale:  
+            return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start *
+                extract_into_tensor(self.scale_arr, t, x_start.shape) +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+        else:
+            return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+
+    def _freeze_model(self):
+        for name, para in self.model.diffusion_model.named_parameters():
+            para.requires_grad = False
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model.eval()
+            self.cond_stage_model.train = disabled_train
+            for param in self.cond_stage_model.parameters():
+                param.requires_grad = False
+        else:
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+    
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+    def get_first_stage_encoding(self, encoder_posterior, noise=None):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample(noise=noise)
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+        return self.scale_factor * z
+   
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        if self.encoder_type == "2d" and x.dim() == 5:
+            b, _, t, _, _ = x.shape
+            x = rearrange(x, 'b c t h w -> (b t) c h w')
+            reshape_back = True
+        else:
+            reshape_back = False
+        
+        encoder_posterior = self.first_stage_model.encode(x)
+        results = self.get_first_stage_encoding(encoder_posterior).detach()
+        
+        if reshape_back:
+            results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t)
+        
+        return results
+    
+    @torch.no_grad()
+    def encode_first_stage_2DAE(self, x):
+
+        b, _, t, _, _ = x.shape
+        results = torch.cat([self.get_first_stage_encoding(self.first_stage_model.encode(x[:,:,i])).detach().unsqueeze(2) for i in range(t)], dim=2)
+        
+        return results
+    
+    def decode_core(self, z, **kwargs):
+        if self.encoder_type == "2d" and z.dim() == 5:
+            b, _, t, _, _ = z.shape
+            z = rearrange(z, 'b c t h w -> (b t) c h w')
+            reshape_back = True
+        else:
+            reshape_back = False
+            
+        z = 1. / self.scale_factor * z
+
+        results = self.first_stage_model.decode(z, **kwargs)
+            
+        if reshape_back:
+            results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t)
+        return results
+
+    @torch.no_grad()
+    def decode_first_stage(self, z, **kwargs):
+        return self.decode_core(z, **kwargs)
+
+    def apply_model(self, x_noisy, t, cond, **kwargs):
+        if isinstance(cond, dict):
+            # hybrid case, cond is exptected to be a dict
+            pass
+        else:
+            if not isinstance(cond, list):
+                cond = [cond]
+            key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
+            cond = {key: cond}
+
+        x_recon = self.model(x_noisy, t, **cond, **kwargs)
+
+        if isinstance(x_recon, tuple):
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def _get_denoise_row_from_list(self, samples, desc=''):
+        denoise_row = []
+        for zd in tqdm(samples, desc=desc):
+            denoise_row.append(self.decode_first_stage(zd.to(self.device)))
+        n_log_timesteps = len(denoise_row)
+
+        denoise_row = torch.stack(denoise_row)  # n_log_timesteps, b, C, H, W
+        
+        if denoise_row.dim() == 5:
+            # img, num_imgs= n_log_timesteps * bs, grid_size=[bs,n_log_timesteps]
+            denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
+            denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+            denoise_grid = make_grid(denoise_grid, nrow=n_log_timesteps)
+        elif denoise_row.dim() == 6:
+            # video, grid_size=[n_log_timesteps*bs, t]
+            video_length = denoise_row.shape[3]
+            denoise_grid = rearrange(denoise_row, 'n b c t h w -> b n c t h w')
+            denoise_grid = rearrange(denoise_grid, 'b n c t h w -> (b n) c t h w')
+            denoise_grid = rearrange(denoise_grid, 'n c t h w -> (n t) c h w')
+            denoise_grid = make_grid(denoise_grid, nrow=video_length)
+        else:
+            raise ValueError
+
+        return denoise_grid
+ 
+
+    @torch.no_grad()
+    def decode_first_stage_2DAE(self, z, **kwargs):
+
+        b, _, t, _, _ = z.shape
+        z = 1. / self.scale_factor * z
+        results = torch.cat([self.first_stage_model.decode(z[:,:,i], **kwargs).unsqueeze(2) for i in range(t)], dim=2)
+
+        return results
+
+
+    def p_mean_variance(self, x, c, t, clip_denoised: bool, return_x0=False, score_corrector=None, corrector_kwargs=None, **kwargs):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, **kwargs)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+
+        if return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, return_x0=False, \
+                 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, **kwargs):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, return_x0=return_x0, \
+                                       score_corrector=score_corrector, corrector_kwargs=corrector_kwargs, **kwargs)
+        if return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+
+        if return_x0:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, cond, shape, return_intermediates=False, x_T=None, verbose=True, callback=None, \
+                      timesteps=None, mask=None, x0=None, img_callback=None, start_T=None, log_every_t=None, **kwargs):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]        
+        # sample an initial noise
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(range(0, timesteps))
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(img, cond, ts, clip_denoised=self.clip_denoised, **kwargs)
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+
+class LatentVisualDiffusion(LatentDiffusion):
+    def __init__(self, cond_img_config, finegrained=False, random_cond=False, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.random_cond = random_cond
+        self.instantiate_img_embedder(cond_img_config, freeze=True)
+        num_tokens = 16 if finegrained else 4
+        self.image_proj_model = self.init_projector(use_finegrained=finegrained, num_tokens=num_tokens, input_dim=1024,\
+                                            cross_attention_dim=1024, dim=1280)    
+
+    def instantiate_img_embedder(self, config, freeze=True):
+        embedder = instantiate_from_config(config)
+        if freeze:
+            self.embedder = embedder.eval()
+            self.embedder.train = disabled_train
+            for param in self.embedder.parameters():
+                param.requires_grad = False
+
+    def init_projector(self, use_finegrained, num_tokens, input_dim, cross_attention_dim, dim):
+        if not use_finegrained:
+            image_proj_model = ImageProjModel(clip_extra_context_tokens=num_tokens, cross_attention_dim=cross_attention_dim,
+                clip_embeddings_dim=input_dim
+            )
+        else:
+            image_proj_model = Resampler(dim=input_dim, depth=4, dim_head=64, heads=12, num_queries=num_tokens,
+                embedding_dim=dim, output_dim=cross_attention_dim, ff_mult=4
+            )
+        return image_proj_model
+
+    ## Never delete this func: it is used in log_images() and inference stage
+    def get_image_embeds(self, batch_imgs):
+        ## img: b c h w
+        img_token = self.embedder(batch_imgs)
+        img_emb = self.image_proj_model(img_token)
+        return img_emb
+
+
+class DiffusionWrapper(pl.LightningModule):
+    def __init__(self, diff_model_config, conditioning_key):
+        super().__init__()
+        self.diffusion_model = instantiate_from_config(diff_model_config)
+        self.conditioning_key = conditioning_key
+
+    def forward(self, x, t, c_concat: list = None, c_crossattn: list = None,
+                c_adm=None, s=None, mask=None, **kwargs):
+        # temporal_context = fps is foNone
+        if self.conditioning_key is None:
+            out = self.diffusion_model(x, t)
+        elif self.conditioning_key == 'concat':
+            xc = torch.cat([x] + c_concat, dim=1)
+            out = self.diffusion_model(xc, t, **kwargs)
+        elif self.conditioning_key == 'crossattn':
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(x, t, context=cc, **kwargs)
+        elif self.conditioning_key == 'hybrid':
+            ## it is just right [b,c,t,h,w]: concatenate in channel dim
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc)
+        elif self.conditioning_key == 'resblockcond':
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, context=cc)
+        elif self.conditioning_key == 'adm':
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, y=cc)
+        elif self.conditioning_key == 'hybrid-adm':
+            assert c_adm is not None
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc, y=c_adm)
+        elif self.conditioning_key == 'hybrid-time':
+            assert s is not None
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc, s=s)
+        elif self.conditioning_key == 'concat-time-mask':
+            # assert s is not None
+            # mainlogger.info('x & mask:',x.shape,c_concat[0].shape)
+            xc = torch.cat([x] + c_concat, dim=1)
+            out = self.diffusion_model(xc, t, context=None, s=s, mask=mask)
+        elif self.conditioning_key == 'concat-adm-mask':
+            # assert s is not None
+            # mainlogger.info('x & mask:',x.shape,c_concat[0].shape)
+            if c_concat is not None:
+                xc = torch.cat([x] + c_concat, dim=1)
+            else:
+                xc = x
+            out = self.diffusion_model(xc, t, context=None, y=s, mask=mask)
+        elif self.conditioning_key == 'hybrid-adm-mask':
+            cc = torch.cat(c_crossattn, 1)
+            if c_concat is not None:
+                xc = torch.cat([x] + c_concat, dim=1)
+            else:
+                xc = x
+            out = self.diffusion_model(xc, t, context=cc, y=s, mask=mask)
+        elif self.conditioning_key == 'hybrid-time-adm': # adm means y, e.g., class index
+            # assert s is not None
+            assert c_adm is not None
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc, s=s, y=c_adm)
+        else:
+            raise NotImplementedError()
+
+        return out

lvdm/models/samplers/ddim.py

@@ -0,0 +1,336 @@
+import numpy as np
+from tqdm import tqdm
+import torch
+from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps
+from lvdm.common import noise_like
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+        self.counter = 0
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+        self.use_scale = self.model.use_scale
+        print('DDIM scale', self.use_scale)
+
+        if self.use_scale:
+            self.register_buffer('scale_arr', to_torch(self.model.scale_arr))
+            ddim_scale_arr = self.scale_arr.cpu()[self.ddim_timesteps]
+            self.register_buffer('ddim_scale_arr', ddim_scale_arr)
+            ddim_scale_arr = np.asarray([self.scale_arr.cpu()[0]] + self.scale_arr.cpu()[self.ddim_timesteps[:-1]].tolist())
+            self.register_buffer('ddim_scale_arr_prev', ddim_scale_arr)
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               schedule_verbose=False,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        
+        # check condition bs
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                try:
+                    cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                except:
+                    cbs = conditioning[list(conditioning.keys())[0]][0].shape[0]
+
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=schedule_verbose)
+        
+        # make shape
+        if len(shape) == 3:
+            C, H, W = shape
+            size = (batch_size, C, H, W)
+        elif len(shape) == 4:
+            C, T, H, W = shape
+            size = (batch_size, C, T, H, W)
+        # print(f'Data shape for DDIM sampling is {size}, eta {eta}')
+        
+        samples, intermediates = self.ddim_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    verbose=verbose,
+                                                    **kwargs)
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,
+                      cond_tau=1., target_size=None, start_timesteps=None,
+                      **kwargs):
+        device = self.model.betas.device        
+        print('ddim device', device)
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+        
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+            
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        if verbose:
+            iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+        else:
+            iterator = time_range
+
+        init_x0 = False
+        clean_cond = kwargs.pop("clean_cond", False)
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+            if start_timesteps is not None:
+                assert x0 is not None
+                if step > start_timesteps*time_range[0]:
+                    continue
+                elif not init_x0:
+                    img = self.model.q_sample(x0, ts) 
+                    init_x0 = True
+
+            # use mask to blend noised original latent (img_orig) & new sampled latent (img)
+            if mask is not None:
+                assert x0 is not None
+                if clean_cond:
+                    img_orig = x0
+                else:
+                    img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass? <ddim inversion>
+                img = img_orig * mask + (1. - mask) * img # keep original & modify use img
+            
+            index_clip =  int((1 - cond_tau) * total_steps)
+            if index <= index_clip and target_size is not None:
+                target_size_ = [target_size[0], target_size[1]//8, target_size[2]//8]
+                img = torch.nn.functional.interpolate(
+                img,
+                size=target_size_,
+                mode="nearest",
+                )
+            outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      x0=x0,
+                                      **kwargs)
+            
+            img, pred_x0 = outs
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,
+                      uc_type=None, conditional_guidance_scale_temporal=None, **kwargs):
+        b, *_, device = *x.shape, x.device
+        if x.dim() == 5:
+            is_video = True
+        else:
+            is_video = False
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+            e_t = self.model.apply_model(x, t, c, **kwargs) # unet denoiser
+        else:
+            # with unconditional condition
+            if isinstance(c, torch.Tensor):
+                e_t = self.model.apply_model(x, t, c, **kwargs)
+                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
+            elif isinstance(c, dict):
+                e_t = self.model.apply_model(x, t, c, **kwargs)
+                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
+            else:
+                raise NotImplementedError
+            # text cfg
+            if uc_type is None:
+                e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+            else:
+                if uc_type == 'cfg_original':
+                    e_t = e_t + unconditional_guidance_scale * (e_t - e_t_uncond)
+                elif uc_type == 'cfg_ours':
+                    e_t = e_t + unconditional_guidance_scale * (e_t_uncond - e_t)
+                else:
+                    raise NotImplementedError
+            # temporal guidance
+            if conditional_guidance_scale_temporal is not None:
+                e_t_temporal = self.model.apply_model(x, t, c, **kwargs)
+                e_t_image = self.model.apply_model(x, t, c, no_temporal_attn=True, **kwargs)
+                e_t = e_t + conditional_guidance_scale_temporal * (e_t_temporal - e_t_image)
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps"
+            e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        # select parameters corresponding to the currently considered timestep
+        
+        if is_video:
+            size = (b, 1, 1, 1, 1)
+        else:
+            size = (b, 1, 1, 1)
+        a_t = torch.full(size, alphas[index], device=device)
+        a_prev = torch.full(size, alphas_prev[index], device=device)
+        sigma_t = torch.full(size, sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full(size, sqrt_one_minus_alphas[index],device=device)
+
+        # current prediction for x_0
+        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        # direction pointing to x_t
+        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        if self.use_scale:
+            scale_arr = self.model.scale_arr if use_original_steps else self.ddim_scale_arr
+            scale_t = torch.full(size, scale_arr[index], device=device)
+            scale_arr_prev = self.model.scale_arr_prev if use_original_steps else self.ddim_scale_arr_prev
+            scale_t_prev = torch.full(size, scale_arr_prev[index], device=device)
+            pred_x0 /= scale_t 
+            x_prev = a_prev.sqrt() * scale_t_prev * pred_x0 + dir_xt + noise
+        else:
+            x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+
+        return x_prev, pred_x0
+
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+
+        def extract_into_tensor(a, t, x_shape):
+            b, *_ = t.shape
+            out = a.gather(-1, t)
+            return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+        return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
+                extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
+
+    @torch.no_grad()
+    def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
+               use_original_steps=False):
+
+        timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
+            x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
+                                          unconditional_guidance_scale=unconditional_guidance_scale,
+                                          unconditional_conditioning=unconditional_conditioning)
+        return x_dec
+

lvdm/models/samplers/ddim_freetraj.py

@@ -0,0 +1,340 @@
+import numpy as np
+from tqdm import tqdm
+import torch
+from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps
+from lvdm.common import noise_like
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+        self.counter = 0
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+        self.use_scale = self.model.use_scale
+        print('DDIM scale', self.use_scale)
+
+        if self.use_scale:
+            self.register_buffer('scale_arr', to_torch(self.model.scale_arr))
+            ddim_scale_arr = self.scale_arr.cpu()[self.ddim_timesteps]
+            self.register_buffer('ddim_scale_arr', ddim_scale_arr)
+            ddim_scale_arr = np.asarray([self.scale_arr.cpu()[0]] + self.scale_arr.cpu()[self.ddim_timesteps[:-1]].tolist())
+            self.register_buffer('ddim_scale_arr_prev', ddim_scale_arr)
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               schedule_verbose=False,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        
+        # check condition bs
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                try:
+                    cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                except:
+                    cbs = conditioning[list(conditioning.keys())[0]][0].shape[0]
+
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=schedule_verbose)
+        
+        # make shape
+        if len(shape) == 3:
+            C, H, W = shape
+            size = (batch_size, C, H, W)
+        elif len(shape) == 4:
+            C, T, H, W = shape
+            size = (batch_size, C, T, H, W)
+        # print(f'Data shape for DDIM sampling is {size}, eta {eta}')
+        
+        samples, intermediates = self.ddim_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    verbose=verbose,
+                                                    **kwargs)
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,
+                      cond_tau=1., target_size=None, start_timesteps=None,
+                      **kwargs):
+        device = self.model.betas.device        
+        print('ddim device', device)
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+        
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+            
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        if verbose:
+            iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+        else:
+            iterator = time_range
+
+        init_x0 = False
+        clean_cond = kwargs.pop("clean_cond", False)
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+            if start_timesteps is not None:
+                assert x0 is not None
+                if step > start_timesteps*time_range[0]:
+                    continue
+                elif not init_x0:
+                    img = self.model.q_sample(x0, ts) 
+                    init_x0 = True
+
+            # use mask to blend noised original latent (img_orig) & new sampled latent (img)
+            if mask is not None:
+                assert x0 is not None
+                if clean_cond:
+                    img_orig = x0
+                else:
+                    img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass? <ddim inversion>
+                img = img_orig * mask + (1. - mask) * img # keep original & modify use img
+            
+            index_clip =  int((1 - cond_tau) * total_steps)
+            if index <= index_clip and target_size is not None:
+                target_size_ = [target_size[0], target_size[1]//8, target_size[2]//8]
+                img = torch.nn.functional.interpolate(
+                img,
+                size=target_size_,
+                mode="nearest",
+                )
+            outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      x0=x0,
+                                      step=i,
+                                      **kwargs)
+            
+            img, pred_x0 = outs
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,
+                      uc_type=None, conditional_guidance_scale_temporal=None, step=0, ddim_edit=0, **kwargs):
+        b, *_, device = *x.shape, x.device
+        if x.dim() == 5:
+            is_video = True
+        else:
+            is_video = False
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+            e_t = self.model.apply_model(x, t, c, **kwargs) # unet denoiser
+        else:
+            # with unconditional condition
+            if step < ddim_edit:
+                e_t = self.model.apply_model(x, t, c, use_freetraj=True, **kwargs)
+                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, use_freetraj=True, **kwargs)
+            elif isinstance(c, torch.Tensor):
+                e_t = self.model.apply_model(x, t, c, **kwargs)
+                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
+            elif isinstance(c, dict):
+                e_t = self.model.apply_model(x, t, c, **kwargs)
+                e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
+            else:
+                raise NotImplementedError
+            # text cfg
+            if uc_type is None:
+                e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+            else:
+                if uc_type == 'cfg_original':
+                    e_t = e_t + unconditional_guidance_scale * (e_t - e_t_uncond)
+                elif uc_type == 'cfg_ours':
+                    e_t = e_t + unconditional_guidance_scale * (e_t_uncond - e_t)
+                else:
+                    raise NotImplementedError
+            # temporal guidance
+            if conditional_guidance_scale_temporal is not None:
+                e_t_temporal = self.model.apply_model(x, t, c, **kwargs)
+                e_t_image = self.model.apply_model(x, t, c, no_temporal_attn=True, **kwargs)
+                e_t = e_t + conditional_guidance_scale_temporal * (e_t_temporal - e_t_image)
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps"
+            e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        # select parameters corresponding to the currently considered timestep
+        
+        if is_video:
+            size = (b, 1, 1, 1, 1)
+        else:
+            size = (b, 1, 1, 1)
+        a_t = torch.full(size, alphas[index], device=device)
+        a_prev = torch.full(size, alphas_prev[index], device=device)
+        sigma_t = torch.full(size, sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full(size, sqrt_one_minus_alphas[index],device=device)
+
+        # current prediction for x_0
+        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        # direction pointing to x_t
+        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        if self.use_scale:
+            scale_arr = self.model.scale_arr if use_original_steps else self.ddim_scale_arr
+            scale_t = torch.full(size, scale_arr[index], device=device)
+            scale_arr_prev = self.model.scale_arr_prev if use_original_steps else self.ddim_scale_arr_prev
+            scale_t_prev = torch.full(size, scale_arr_prev[index], device=device)
+            pred_x0 /= scale_t 
+            x_prev = a_prev.sqrt() * scale_t_prev * pred_x0 + dir_xt + noise
+        else:
+            x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+
+        return x_prev, pred_x0
+
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+
+        def extract_into_tensor(a, t, x_shape):
+            b, *_ = t.shape
+            out = a.gather(-1, t)
+            return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+        return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
+                extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
+
+    @torch.no_grad()
+    def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
+               use_original_steps=False):
+
+        timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
+            x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
+                                          unconditional_guidance_scale=unconditional_guidance_scale,
+                                          unconditional_conditioning=unconditional_conditioning)
+        return x_dec
+

lvdm/models/utils_diffusion.py

@@ -0,0 +1,104 @@
+import math
+import numpy as np
+from einops import repeat
+import torch
+import torch.nn.functional as F
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    # print(f'ddim_timesteps={ddim_timesteps}, len_alphacums={len(alphacums)}')
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        print(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)

lvdm/modules/attention.py

@@ -0,0 +1,475 @@
+from functools import partial
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+from einops import rearrange, repeat
+try:
+    import xformers
+    import xformers.ops
+    XFORMERS_IS_AVAILBLE = True
+except:
+    XFORMERS_IS_AVAILBLE = False
+from lvdm.common import (
+    checkpoint,
+    exists,
+    default,
+)
+from lvdm.basics import (
+    zero_module,
+)
+
+class RelativePosition(nn.Module):
+    """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """
+
+    def __init__(self, num_units, max_relative_position):
+        super().__init__()
+        self.num_units = num_units
+        self.max_relative_position = max_relative_position
+        self.embeddings_table = nn.Parameter(torch.Tensor(max_relative_position * 2 + 1, num_units))
+        nn.init.xavier_uniform_(self.embeddings_table)
+
+    def forward(self, length_q, length_k):
+        device = self.embeddings_table.device
+        range_vec_q = torch.arange(length_q, device=device)
+        range_vec_k = torch.arange(length_k, device=device)
+        distance_mat = range_vec_k[None, :] - range_vec_q[:, None]
+        distance_mat_clipped = torch.clamp(distance_mat, -self.max_relative_position, self.max_relative_position)
+        final_mat = distance_mat_clipped + self.max_relative_position
+        final_mat = final_mat.long()
+        embeddings = self.embeddings_table[final_mat]
+        return embeddings
+
+
+class CrossAttention(nn.Module):
+
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., 
+                 relative_position=False, temporal_length=None, img_cross_attention=False):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head**-0.5
+        self.heads = heads
+        self.dim_head = dim_head
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
+
+        self.image_cross_attention_scale = 1.0
+        self.text_context_len = 77
+        self.img_cross_attention = img_cross_attention
+        if self.img_cross_attention:
+            self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
+            self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
+        
+        self.relative_position = relative_position
+        if self.relative_position:
+            assert(temporal_length is not None)
+            self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
+            self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
+        else:
+            ## only used for spatial attention, while NOT for temporal attention
+            if XFORMERS_IS_AVAILBLE and temporal_length is None:
+                self.forward = self.efficient_forward
+
+    def forward(self, x, context=None, mask=None):
+        h = self.heads
+
+        q = self.to_q(x)
+        context = default(context, x)
+        ## considering image token additionally
+        if context is not None and self.img_cross_attention:
+            context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
+            k = self.to_k(context)
+            v = self.to_v(context)
+            k_ip = self.to_k_ip(context_img)
+            v_ip = self.to_v_ip(context_img)
+        else:
+            k = self.to_k(context)
+            v = self.to_v(context)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+        sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
+        if self.relative_position:
+            len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
+            k2 = self.relative_position_k(len_q, len_k)
+            sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check 
+            sim += sim2
+        del k
+
+        if exists(mask):
+            ## feasible for causal attention mask only
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b i j -> (b h) i j', h=h)
+            sim.masked_fill_(~(mask>0.5), max_neg_value)
+
+        # attention, what we cannot get enough of
+        sim = sim.softmax(dim=-1)
+        out = torch.einsum('b i j, b j d -> b i d', sim, v)
+        if self.relative_position:
+            v2 = self.relative_position_v(len_q, len_v)
+            out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
+            out += out2
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+
+        ## considering image token additionally
+        if context is not None and self.img_cross_attention:
+            k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip))
+            sim_ip =  torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale
+            del k_ip
+            sim_ip = sim_ip.softmax(dim=-1)
+            out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
+            out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
+            out = out + self.image_cross_attention_scale * out_ip
+        del q
+
+        return self.to_out(out)
+    
+    def efficient_forward(self, x, context=None, mask=None):
+        q = self.to_q(x)
+        context = default(context, x)
+
+        ## considering image token additionally
+        if context is not None and self.img_cross_attention:
+            context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
+            k = self.to_k(context)
+            v = self.to_v(context)
+            k_ip = self.to_k_ip(context_img)
+            v_ip = self.to_v_ip(context_img)
+        else:
+            k = self.to_k(context)
+            v = self.to_v(context)
+
+        b, _, _ = q.shape
+        q, k, v = map(
+            lambda t: t.unsqueeze(3)
+            .reshape(b, t.shape[1], self.heads, self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b * self.heads, t.shape[1], self.dim_head)
+            .contiguous(),
+            (q, k, v),
+        )
+        # actually compute the attention, what we cannot get enough of
+        out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None)
+
+        ## considering image token additionally
+        if context is not None and self.img_cross_attention:
+            k_ip, v_ip = map(
+                lambda t: t.unsqueeze(3)
+                .reshape(b, t.shape[1], self.heads, self.dim_head)
+                .permute(0, 2, 1, 3)
+                .reshape(b * self.heads, t.shape[1], self.dim_head)
+                .contiguous(),
+                (k_ip, v_ip),
+            )
+            out_ip = xformers.ops.memory_efficient_attention(q, k_ip, v_ip, attn_bias=None, op=None)
+            out_ip = (
+                out_ip.unsqueeze(0)
+                .reshape(b, self.heads, out.shape[1], self.dim_head)
+                .permute(0, 2, 1, 3)
+                .reshape(b, out.shape[1], self.heads * self.dim_head)
+            )
+
+        if exists(mask):
+            raise NotImplementedError
+        out = (
+            out.unsqueeze(0)
+            .reshape(b, self.heads, out.shape[1], self.dim_head)
+            .permute(0, 2, 1, 3)
+            .reshape(b, out.shape[1], self.heads * self.dim_head)
+        )
+        if context is not None and self.img_cross_attention:
+            out = out + self.image_cross_attention_scale * out_ip
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
+                disable_self_attn=False, attention_cls=None, img_cross_attention=False):
+        super().__init__()
+        attn_cls = CrossAttention if attention_cls is None else attention_cls
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+            context_dim=context_dim if self.disable_self_attn else None)
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+            img_cross_attention=img_cross_attention)
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None, mask=None):
+        ## implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments
+        input_tuple = (x,)      ## should not be (x), otherwise *input_tuple will decouple x into multiple arguments
+        if context is not None:
+            input_tuple = (x, context)
+        if mask is not None:
+            forward_mask = partial(self._forward, mask=mask)
+            return checkpoint(forward_mask, (x,), self.parameters(), self.checkpoint)
+        if context is not None and mask is not None:
+            input_tuple = (x, context, mask)
+        return checkpoint(self._forward, input_tuple, self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context=None, mask=None):
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None, mask=mask) + x
+        x = self.attn2(self.norm2(x), context=context, mask=mask) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data in spatial axis.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    NEW: use_linear for more efficiency instead of the 1x1 convs
+    """
+
+    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
+                 use_checkpoint=True, disable_self_attn=False, use_linear=False, img_cross_attention=False):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        if not use_linear:
+            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList([
+            BasicTransformerBlock(
+                inner_dim,
+                n_heads,
+                d_head,
+                dropout=dropout,
+                context_dim=context_dim,
+                img_cross_attention=img_cross_attention,
+                disable_self_attn=disable_self_attn,
+                checkpoint=use_checkpoint) for d in range(depth)
+        ])
+        if not use_linear:
+            self.proj_out = zero_module(nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
+        else:
+            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
+        self.use_linear = use_linear
+
+
+    def forward(self, x, context=None):
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            x = block(x, context=context)
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = self.proj_out(x)
+        return x + x_in
+    
+    
+class TemporalTransformer(nn.Module):
+    """
+    Transformer block for image-like data in temporal axis.
+    First, reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
+                 use_checkpoint=True, use_linear=False, only_self_att=True, causal_attention=False,
+                 relative_position=False, temporal_length=None):
+        super().__init__()
+        self.only_self_att = only_self_att
+        self.relative_position = relative_position
+        self.causal_attention = causal_attention
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        if not use_linear:
+            self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        if relative_position:
+            assert(temporal_length is not None)
+            attention_cls = partial(CrossAttention, relative_position=True, temporal_length=temporal_length)
+        else:
+            attention_cls = None
+        if self.causal_attention:
+            assert(temporal_length is not None)
+            self.mask = torch.tril(torch.ones([1, temporal_length, temporal_length]))
+
+        if self.only_self_att:
+            context_dim = None
+        self.transformer_blocks = nn.ModuleList([
+            BasicTransformerBlock(
+                inner_dim,
+                n_heads,
+                d_head,
+                dropout=dropout,
+                context_dim=context_dim,
+                attention_cls=attention_cls,
+                checkpoint=use_checkpoint) for d in range(depth)
+        ])
+        if not use_linear:
+            self.proj_out = zero_module(nn.Conv1d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
+        else:
+            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
+        self.use_linear = use_linear
+
+    def forward(self, x, context=None):
+        b, c, t, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous()
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'bhw c t -> bhw t c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+
+        if self.causal_attention:
+            mask = self.mask.to(x.device)
+            mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b*h*w)
+        else:
+            mask = None
+
+        if self.only_self_att:
+            ## note: if no context is given, cross-attention defaults to self-attention
+            for i, block in enumerate(self.transformer_blocks):
+                x = block(x, mask=mask)
+            x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
+        else:
+            x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
+            context = rearrange(context, '(b t) l con -> b t l con', t=t).contiguous()
+            for i, block in enumerate(self.transformer_blocks):
+                # calculate each batch one by one (since number in shape could not greater then 65,535 for some package)
+                for j in range(b):
+                    context_j = repeat(
+                        context[j],
+                        't l con -> (t r) l con', r=(h * w) // t, t=t).contiguous()
+                    ## note: causal mask will not applied in cross-attention case
+                    x[j] = block(x[j], context=context_j)
+        
+        if self.use_linear:
+            x = self.proj_out(x)
+            x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous()
+            x = self.proj_out(x)
+            x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h, w=w).contiguous()
+
+        return x + x_in
+    
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_

lvdm/modules/attention_freetraj.py

@@ -0,0 +1,648 @@
+from functools import partial
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+from einops import rearrange, repeat
+try:
+    import xformers
+    import xformers.ops
+    XFORMERS_IS_AVAILBLE = True
+except:
+    XFORMERS_IS_AVAILBLE = False
+from lvdm.common import (
+    checkpoint,
+    exists,
+    default,
+)
+from lvdm.basics import (
+    zero_module,
+)
+
+from utils.utils_freetraj import get_path, plan_path
+import math
+
+def gaussian_2d(x=0, y=0, mx=0, my=0, sx=1, sy=1):
+    """ 2d Gaussian weight function
+    """
+    gaussian_map = (
+        1
+        / (2 * math.pi * sx * sy)
+        * torch.exp(-((x - mx) ** 2 / (2 * sx**2) + (y - my) ** 2 / (2 * sy**2)))
+    )
+    gaussian_map.div_(gaussian_map.max())
+    return gaussian_map
+
+def gaussian_weight(height=32, width=32, KERNEL_DIVISION=3.0):
+
+    x = torch.linspace(0, height, height)
+    y = torch.linspace(0, width, width)
+    x, y = torch.meshgrid(x, y, indexing="ij")
+    noise_patch = (
+                    gaussian_2d(
+                        x,
+                        y,
+                        mx=int(height / 2),
+                        my=int(width / 2),
+                        sx=float(height / KERNEL_DIVISION),
+                        sy=float(width / KERNEL_DIVISION),
+                    )
+                ).half()
+    return noise_patch
+
+class RelativePosition(nn.Module):
+    """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """
+
+    def __init__(self, num_units, max_relative_position):
+        super().__init__()
+        self.num_units = num_units
+        self.max_relative_position = max_relative_position
+        self.embeddings_table = nn.Parameter(torch.Tensor(max_relative_position * 2 + 1, num_units))
+        nn.init.xavier_uniform_(self.embeddings_table)
+
+    def forward(self, length_q, length_k):
+        device = self.embeddings_table.device
+        range_vec_q = torch.arange(length_q, device=device)
+        range_vec_k = torch.arange(length_k, device=device)
+        distance_mat = range_vec_k[None, :] - range_vec_q[:, None]
+        distance_mat_clipped = torch.clamp(distance_mat, -self.max_relative_position, self.max_relative_position)
+        final_mat = distance_mat_clipped + self.max_relative_position
+        final_mat = final_mat.long()
+        embeddings = self.embeddings_table[final_mat]
+        return embeddings
+
+
+class CrossAttention(nn.Module):
+
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., 
+                 relative_position=False, temporal_length=None, img_cross_attention=False):
+        super().__init__()
+        inner_dim = dim_head * heads
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head**-0.5
+        self.heads = heads
+        self.dim_head = dim_head
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
+
+        self.image_cross_attention_scale = 1.0
+        self.text_context_len = 77
+        self.img_cross_attention = img_cross_attention
+        if self.img_cross_attention:
+            self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
+            self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
+        
+        self.relative_position = relative_position
+        if self.relative_position:
+            assert(temporal_length is not None)
+            self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
+            self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
+        else:
+            ## only used for spatial attention, while NOT for temporal attention
+            if XFORMERS_IS_AVAILBLE and temporal_length is None:
+                self.forward = self.space_forward
+
+    def forward(self, x, context=None, mask=None, use_freetraj=False, idx_list=[], input_traj=[]):
+        h = self.heads
+
+        q = self.to_q(x)
+        context = default(context, x)
+        ## considering image token additionally
+        if context is not None and self.img_cross_attention:
+            context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
+            k = self.to_k(context)
+            v = self.to_v(context)
+            k_ip = self.to_k_ip(context_img)
+            v_ip = self.to_v_ip(context_img)
+        else:
+            k = self.to_k(context)
+            v = self.to_v(context)
+
+        hw = q.shape[0]
+        w_base = 64
+        h_base = 40
+        w_len = int((hw / w_base / h_base) ** 0.5 * h_base)
+        h_len = int(hw / w_len)
+        BOX_SIZE_H = input_traj[0][2] - input_traj[0][1]
+        BOX_SIZE_W = input_traj[0][4] - input_traj[0][3]
+        PATHS = plan_path(input_traj)
+        sub_h = int(BOX_SIZE_H * h_len) 
+        sub_w = int(BOX_SIZE_W * w_len)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+        sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
+        if self.relative_position:
+            len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
+            k2 = self.relative_position_k(len_q, len_k)
+            sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check 
+            sim += sim2
+        del k
+
+        if use_freetraj:
+            sim = rearrange(sim, '(y x h) i j -> y x h i j', h=h, y=h_len)
+            sim_mask = torch.zeros_like(sim)
+            for i in range(sim.shape[3]):
+                h_start1 = int(PATHS[i][0] * h_len)
+                h_end1 = h_start1 + sub_h
+                w_start1 = int(PATHS[i][2] * w_len)
+                w_end1 = w_start1 + sub_w
+
+                h_fg1 = list(range(h_start1, h_end1))
+                h_fg_tensor1 = torch.zeros(h_len, device=sim.device)
+                h_fg_tensor1[h_fg1] = 1
+                w_fg1 = list(range(w_start1, w_end1))
+                w_fg_tensor1 = torch.zeros(w_len, device=sim.device)
+                w_fg_tensor1[w_fg1] = 1
+                fg_tensor1 = h_fg_tensor1.view(-1, 1) * w_fg_tensor1.view(1, -1)
+                bg_tensor1 = 1 - fg_tensor1
+
+                for j in range(sim.shape[4]):
+                    h_start2 = int(PATHS[j][0] * h_len)
+                    h_end2 = h_start2 + sub_h
+                    w_start2 = int(PATHS[j][2] * w_len)
+                    w_end2 = w_start2 + sub_w
+
+                    h_fg2 = list(range(h_start2, h_end2))
+                    h_fg_tensor2 = torch.zeros(h_len, device=sim.device)
+                    h_fg_tensor2[h_fg2] = 1
+                    w_fg2 = list(range(w_start2, w_end2))
+                    w_fg_tensor2 = torch.zeros(w_len, device=sim.device)
+                    w_fg_tensor2[w_fg2] = 1
+                    fg_tensor2 = h_fg_tensor2.view(-1, 1) * w_fg_tensor2.view(1, -1)
+                    bg_tensor2 = 1 - fg_tensor2
+                    fg_tensor = fg_tensor1 * fg_tensor2
+                    bg_tensor = bg_tensor1 * bg_tensor2
+
+                    coef = 0.01
+                    sim_mask[:, :, :, i, j] = coef * torch.ones_like(sim_mask[:, :, :, i, j])
+                    sim_mask[:, :, :, i, j] += (1 - coef) * torch.ones_like(sim_mask[:, :, :, i, j]) * (fg_tensor.view(h_len, w_len, 1) + bg_tensor.view(h_len, w_len, 1))
+
+            sim *= sim_mask
+            sim = rearrange(sim, 'y x h i j -> (y x h) i j')
+
+            del sim_mask
+
+        if exists(mask):
+            ## feasible for causal attention mask only
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b i j -> (b h) i j', h=h)
+            sim.masked_fill_(~(mask>0.5), max_neg_value)
+
+        # attention, what we cannot get enough of
+        sim = sim.softmax(dim=-1)
+        out = torch.einsum('b i j, b j d -> b i d', sim, v)
+        if self.relative_position:
+            v2 = self.relative_position_v(len_q, len_v)
+            out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
+            out += out2
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+
+        ## considering image token additionally
+        if context is not None and self.img_cross_attention:
+            k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip))
+            sim_ip =  torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale
+            del k_ip
+            sim_ip = sim_ip.softmax(dim=-1)
+            out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
+            out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
+            out = out + self.image_cross_attention_scale * out_ip
+        del q
+
+        return self.to_out(out)
+    
+    def space_forward(self, x, context=None, mask=None, use_freetraj=False, idx_list=[], input_traj=[]):
+        
+        if context is None:
+            SA_flag = True
+        else:
+            SA_flag = False
+
+        h = self.heads
+        
+        q = self.to_q(x)
+        context = default(context, x)
+
+        ## considering image token additionally
+        if context is not None and self.img_cross_attention:
+            context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
+            k = self.to_k(context)
+            v = self.to_v(context)
+            k_ip = self.to_k_ip(context_img)
+            v_ip = self.to_v_ip(context_img)
+        else:
+            k = self.to_k(context)
+            v = self.to_v(context)
+
+        hw = q.shape[1]
+        w_base = 64
+        h_base = 40
+        w_len = int((hw / h_base / w_base) ** 0.5 * w_base)
+        h_len = int(hw / w_len)
+        BOX_SIZE_H = input_traj[0][2] - input_traj[0][1]
+        BOX_SIZE_W = input_traj[0][4] - input_traj[0][3]
+        PATHS = plan_path(input_traj)
+        sub_h = int(BOX_SIZE_H * h_len) 
+        sub_w = int(BOX_SIZE_W * w_len)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+        sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
+        if self.relative_position:
+            len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
+            k2 = self.relative_position_k(len_q, len_k)
+            sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check 
+            sim += sim2
+        del k
+
+        if use_freetraj:
+            coef_a = 0.25 / (BOX_SIZE_H * BOX_SIZE_W) / len(idx_list)
+            weight = gaussian_weight(sub_h, sub_w).to(x.device)
+            if SA_flag:
+                weight_add = 0
+
+                sim = rearrange(sim, '(t h) (y x) (y0 x0) -> t h y x y0 x0', h=h, y=h_len, y0=h_len)
+                sim_mask = torch.zeros_like(sim)
+                for i in range(sim.shape[0]):
+                    h_start = int(PATHS[i][0] * h_len)
+                    h_end = h_start + sub_h
+                    w_start = int(PATHS[i][2] * w_len)
+                    w_end = w_start + sub_w
+
+                    h_fg = list(range(h_start, h_end))
+                    h_fg_tensor = torch.zeros(h_len, device=sim.device)
+                    h_fg_tensor[h_fg] = 1
+                    w_fg = list(range(w_start, w_end))
+                    w_fg_tensor = torch.zeros(w_len, device=sim.device)
+                    w_fg_tensor[w_fg] = 1
+                    fg_tensor = h_fg_tensor.view(-1, 1) * w_fg_tensor.view(1, -1)
+                    bg_tensor = 1 - fg_tensor
+
+                    coef = 0.01
+                    sim_mask[i] = coef * torch.ones_like(sim_mask[i])
+                    sim_mask[i] += (1-coef) * (torch.ones_like(sim_mask[i]) * fg_tensor.view(1, h_len, w_len, 1, 1) * fg_tensor.view(1, 1, 1, h_len, w_len) + torch.ones_like(sim_mask[i]) * bg_tensor.view(1, h_len, w_len, 1, 1) * bg_tensor.view(1, 1, 1, h_len, w_len))
+                
+                sim *= sim_mask
+                sim = rearrange(sim, 't h y x y0 x0 -> (t h) (y x) (y0 x0)')    
+
+            else:
+                sim = rearrange(sim, '(t h) (y x) d -> t h y x d', h=h, y=h_len)
+                sim_mask = torch.zeros_like(sim)
+                weight_add = torch.zeros_like(sim)
+                weight_map = torch.zeros([sim.shape[0], h_len, w_len], device=sim.device)
+                for i in range(sim.shape[0]):
+                    h_start = int(PATHS[i][0] * h_len)
+                    h_end = h_start + sub_h
+                    w_start = int(PATHS[i][2] * w_len)
+                    w_end = w_start + sub_w
+
+                    h_fg = list(range(h_start, h_end))
+                    h_fg_tensor = torch.zeros(h_len, device=sim.device)
+                    h_fg_tensor[h_fg] = 1
+                    w_fg = list(range(w_start, w_end))
+                    w_fg_tensor = torch.zeros(w_len, device=sim.device)
+                    w_fg_tensor[w_fg] = 1
+                    fg_tensor = h_fg_tensor.view(-1, 1) * w_fg_tensor.view(1, -1)
+                    bg_tensor = 1 - fg_tensor
+                    if idx_list == []:
+                        p_fg = [2]
+                    else:
+                        p_fg = idx_list
+                    p_bg = list(range(77))
+                    for j in p_fg:
+                        p_bg.remove(j)
+
+                    weight_map[i, h_start:h_end, w_start:w_end] = weight * coef_a
+                    sim_mask[i, :, :, :, p_bg] = torch.ones_like(sim_mask[i, :, :, :, p_bg]) * bg_tensor.view(1, h_len, w_len, 1)
+                    weight_add[i, :, :, :, p_fg] = torch.ones_like(sim_mask[i, :, :, :, p_fg]) * weight_map[i].view(1, h_len, w_len, 1)
+
+                max_neg_value = -torch.finfo(sim.dtype).max
+                sim.masked_fill_(~(sim_mask>0.5), max_neg_value)
+                sim = rearrange(sim, 't h y x d -> (t h) (y x) d')
+                weight_add = rearrange(weight_add, 't h y x d -> (t h) (y x) d')
+
+            del sim_mask
+
+        if exists(mask):
+            ## feasible for causal attention mask only
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b i j -> (b h) i j', h=h)
+            sim.masked_fill_(~(mask>0.5), max_neg_value)
+
+        # attention, what we cannot get enough of
+        sim = sim.softmax(dim=-1)
+        if use_freetraj:
+            sim += weight_add
+
+        out = torch.einsum('b i j, b j d -> b i d', sim, v)
+        if self.relative_position:
+            v2 = self.relative_position_v(len_q, len_v)
+            out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
+            out += out2
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+
+        ## considering image token additionally
+        if context is not None and self.img_cross_attention:
+            k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip))
+            sim_ip =  torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale
+            del k_ip
+            sim_ip = sim_ip.softmax(dim=-1)
+            out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
+            out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
+            out = out + self.image_cross_attention_scale * out_ip
+        del q
+
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
+                disable_self_attn=False, attention_cls=None, img_cross_attention=False):
+        super().__init__()
+        attn_cls = CrossAttention if attention_cls is None else attention_cls
+        self.disable_self_attn = disable_self_attn
+        self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+            context_dim=context_dim if self.disable_self_attn else None)
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
+            img_cross_attention=img_cross_attention)
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None, mask=None, use_freetraj=False, idx_list=[], input_traj=[], **kwargs):
+        ## implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments
+        input_tuple = (x,)      ## should not be (x), otherwise *input_tuple will decouple x into multiple arguments
+        if context is not None:
+            input_tuple = (x, context)
+        if mask is not None:
+            forward_mask = partial(self._forward, mask=mask)
+            return checkpoint(forward_mask, (x,), self.parameters(), self.checkpoint)
+        if context is not None and mask is not None:
+            input_tuple = (x, context, mask)
+        input_tuple = (x, context, mask, use_freetraj, idx_list, input_traj)
+        return checkpoint(self._forward, input_tuple, self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context=None, mask=None, use_freetraj=False, idx_list=[], input_traj=[]):
+        x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None, mask=mask, use_freetraj=use_freetraj, idx_list=idx_list, input_traj=input_traj) + x
+        x = self.attn2(self.norm2(x), context=context, mask=mask, use_freetraj=use_freetraj, idx_list=idx_list, input_traj=input_traj) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data in spatial axis.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    NEW: use_linear for more efficiency instead of the 1x1 convs
+    """
+
+    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
+                 use_checkpoint=True, disable_self_attn=False, use_linear=False, img_cross_attention=False):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        if not use_linear:
+            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList([
+            BasicTransformerBlock(
+                inner_dim,
+                n_heads,
+                d_head,
+                dropout=dropout,
+                context_dim=context_dim,
+                img_cross_attention=img_cross_attention,
+                disable_self_attn=disable_self_attn,
+                checkpoint=use_checkpoint) for d in range(depth)
+        ])
+        if not use_linear:
+            self.proj_out = zero_module(nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
+        else:
+            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
+        self.use_linear = use_linear
+
+
+    def forward(self, x, context=None, **kwargs):
+        b, c, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+        for i, block in enumerate(self.transformer_blocks):
+            x = block(x, context=context, **kwargs)
+        if self.use_linear:
+            x = self.proj_out(x)
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = self.proj_out(x)
+        return x + x_in
+    
+    
+class TemporalTransformer(nn.Module):
+    """
+    Transformer block for image-like data in temporal axis.
+    First, reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+    def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
+                 use_checkpoint=True, use_linear=False, only_self_att=True, causal_attention=False,
+                 relative_position=False, temporal_length=None):
+        super().__init__()
+        self.only_self_att = only_self_att
+        self.relative_position = relative_position
+        self.causal_attention = causal_attention
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        if not use_linear:
+            self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        else:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        if relative_position:
+            assert(temporal_length is not None)
+            attention_cls = partial(CrossAttention, relative_position=True, temporal_length=temporal_length)
+        else:
+            attention_cls = partial(CrossAttention, temporal_length=temporal_length)
+            # attention_cls = None
+        if self.causal_attention:
+            assert(temporal_length is not None)
+            self.mask = torch.tril(torch.ones([1, temporal_length, temporal_length]))
+
+        if self.only_self_att:
+            context_dim = None
+        self.transformer_blocks = nn.ModuleList([
+            BasicTransformerBlock(
+                inner_dim,
+                n_heads,
+                d_head,
+                dropout=dropout,
+                context_dim=context_dim,
+                attention_cls=attention_cls,
+                checkpoint=use_checkpoint) for d in range(depth)
+        ])
+        if not use_linear:
+            self.proj_out = zero_module(nn.Conv1d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
+        else:
+            self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
+        self.use_linear = use_linear
+
+    def forward(self, x, context=None, **kwargs):
+        b, c, t, h, w = x.shape
+        x_in = x
+        x = self.norm(x)
+        x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous()
+        if not self.use_linear:
+            x = self.proj_in(x)
+        x = rearrange(x, 'bhw c t -> bhw t c').contiguous()
+        if self.use_linear:
+            x = self.proj_in(x)
+
+        if self.causal_attention:
+            mask = self.mask.to(x.device)
+            mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b*h*w)
+        else:
+            mask = None
+
+        if self.only_self_att:
+            ## note: if no context is given, cross-attention defaults to self-attention
+            for i, block in enumerate(self.transformer_blocks):
+                x = block(x, mask=mask, **kwargs)
+            x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
+        else:
+            x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
+            context = rearrange(context, '(b t) l con -> b t l con', t=t).contiguous()
+            for i, block in enumerate(self.transformer_blocks):
+                # calculate each batch one by one (since number in shape could not greater then 65,535 for some package)
+                for j in range(b):
+                    context_j = repeat(
+                        context[j],
+                        't l con -> (t r) l con', r=(h * w) // t, t=t).contiguous()
+                    ## note: causal mask will not applied in cross-attention case
+                    x[j] = block(x[j], context=context_j, **kwargs)
+        
+        if self.use_linear:
+            x = self.proj_out(x)
+            x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous()
+        if not self.use_linear:
+            x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous()
+            x = self.proj_out(x)
+            x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h, w=w).contiguous()
+
+        return x + x_in
+    
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x, **kwargs):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_

lvdm/modules/encoders/condition.py

@@ -0,0 +1,392 @@
+import torch
+import torch.nn as nn
+from torch.utils.checkpoint import checkpoint
+import kornia
+import open_clip
+from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
+from lvdm.common import autocast
+from utils.utils import count_params
+
+class AbstractEncoder(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+class IdentityEncoder(AbstractEncoder):
+
+    def encode(self, x):
+        return x
+
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+        self.n_classes = n_classes
+        self.ucg_rate = ucg_rate
+
+    def forward(self, batch, key=None, disable_dropout=False):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        if self.ucg_rate > 0. and not disable_dropout:
+            mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate)
+            c = mask * c + (1 - mask) * torch.ones_like(c) * (self.n_classes - 1)
+            c = c.long()
+        c = self.embedding(c)
+        return c
+
+    def get_unconditional_conditioning(self, bs, device="cuda"):
+        uc_class = self.n_classes - 1  # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000)
+        uc = torch.ones((bs,), device=device) * uc_class
+        uc = {self.key: uc}
+        return uc
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class FrozenT5Embedder(AbstractEncoder):
+    """Uses the T5 transformer encoder for text"""
+
+    def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77,
+                 freeze=True):  # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
+        super().__init__()
+        self.tokenizer = T5Tokenizer.from_pretrained(version)
+        self.transformer = T5EncoderModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length  # TODO: typical value?
+        if freeze:
+            self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        # self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenCLIPEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from huggingface)"""
+    LAYERS = [
+        "last",
+        "pooled",
+        "hidden"
+    ]
+
+    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77,
+                 freeze=True, layer="last", layer_idx=None):  # clip-vit-base-patch32
+        super().__init__()
+        assert layer in self.LAYERS
+        self.tokenizer = CLIPTokenizer.from_pretrained(version)
+        self.transformer = CLIPTextModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        self.layer_idx = layer_idx
+        if layer == "hidden":
+            assert layer_idx is not None
+            assert 0 <= abs(layer_idx) <= 12
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        # self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer == "hidden")
+        if self.layer == "last":
+            z = outputs.last_hidden_state
+        elif self.layer == "pooled":
+            z = outputs.pooler_output[:, None, :]
+        else:
+            z = outputs.hidden_states[self.layer_idx]
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+
+class ClipImageEmbedder(nn.Module):
+    def __init__(
+            self,
+            model,
+            jit=False,
+            device='cuda' if torch.cuda.is_available() else 'cpu',
+            antialias=True,
+            ucg_rate=0.
+    ):
+        super().__init__()
+        from clip import load as load_clip
+        self.model, _ = load_clip(name=model, device=device, jit=jit)
+
+        self.antialias = antialias
+
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+        self.ucg_rate = ucg_rate
+
+    def preprocess(self, x):
+        # normalize to [0,1]
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic', align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # re-normalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def forward(self, x, no_dropout=False):
+        # x is assumed to be in range [-1,1]
+        out = self.model.encode_image(self.preprocess(x))
+        out = out.to(x.dtype)
+        if self.ucg_rate > 0. and not no_dropout:
+            out = torch.bernoulli((1. - self.ucg_rate) * torch.ones(out.shape[0], device=out.device))[:, None] * out
+        return out
+
+
+class FrozenOpenCLIPEmbedder(AbstractEncoder):
+    """
+    Uses the OpenCLIP transformer encoder for text
+    """
+    LAYERS = [
+        # "pooled",
+        "last",
+        "penultimate"
+    ]
+
+    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
+                 freeze=True, layer="last"):
+        super().__init__()
+        assert layer in self.LAYERS
+        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'))
+        del model.visual
+        self.model = model
+
+        self.device = device
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        if self.layer == "last":
+            self.layer_idx = 0
+        elif self.layer == "penultimate":
+            self.layer_idx = 1
+        else:
+            raise NotImplementedError()
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        self.device = self.model.positional_embedding.device
+        tokens = open_clip.tokenize(text)
+        z = self.encode_with_transformer(tokens.to(self.device))
+        return z
+
+    def encode_with_transformer(self, text):
+        x = self.model.token_embedding(text)  # [batch_size, n_ctx, d_model]
+        x = x + self.model.positional_embedding
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+        x = self.model.ln_final(x)
+        return x
+
+    def text_transformer_forward(self, x: torch.Tensor, attn_mask=None):
+        for i, r in enumerate(self.model.transformer.resblocks):
+            if i == len(self.model.transformer.resblocks) - self.layer_idx:
+                break
+            if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting():
+                x = checkpoint(r, x, attn_mask)
+            else:
+                x = r(x, attn_mask=attn_mask)
+        return x
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenOpenCLIPImageEmbedder(AbstractEncoder):
+    """
+    Uses the OpenCLIP vision transformer encoder for images
+    """
+
+    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
+                 freeze=True, layer="pooled", antialias=True, ucg_rate=0.):
+        super().__init__()
+        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
+                                                            pretrained=version, )
+        del model.transformer
+        self.model = model
+
+        self.device = device
+        self.max_length = max_length
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        if self.layer == "penultimate":
+            raise NotImplementedError()
+            self.layer_idx = 1
+
+        self.antialias = antialias
+
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+        self.ucg_rate = ucg_rate
+
+    def preprocess(self, x):
+        # normalize to [0,1]
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic', align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+
+    @autocast
+    def forward(self, image, no_dropout=False):
+        z = self.encode_with_vision_transformer(image)
+        if self.ucg_rate > 0. and not no_dropout:
+            z = torch.bernoulli((1. - self.ucg_rate) * torch.ones(z.shape[0], device=z.device))[:, None] * z
+        return z
+
+    def encode_with_vision_transformer(self, img):
+        img = self.preprocess(img)
+        x = self.model.visual(img)
+        return x
+
+    def encode(self, text):
+        return self(text)
+
+
+
+class FrozenOpenCLIPImageEmbedderV2(AbstractEncoder):
+    """
+    Uses the OpenCLIP vision transformer encoder for images
+    """
+
+    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda",
+                 freeze=True, layer="pooled", antialias=True):
+        super().__init__()
+        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
+                                                            pretrained=version, )
+        del model.transformer
+        self.model = model
+        self.device = device
+
+        if freeze:
+            self.freeze()
+        self.layer = layer
+        if self.layer == "penultimate":
+            raise NotImplementedError()
+            self.layer_idx = 1
+
+        self.antialias = antialias
+        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
+        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
+
+
+    def preprocess(self, x):
+        # normalize to [0,1]
+        x = kornia.geometry.resize(x, (224, 224),
+                                   interpolation='bicubic', align_corners=True,
+                                   antialias=self.antialias)
+        x = (x + 1.) / 2.
+        # renormalize according to clip
+        x = kornia.enhance.normalize(x, self.mean, self.std)
+        return x
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.model.parameters():
+            param.requires_grad = False
+
+    def forward(self, image, no_dropout=False):
+        ## image: b c h w
+        z = self.encode_with_vision_transformer(image)
+        return z
+
+    def encode_with_vision_transformer(self, x):
+        x = self.preprocess(x)
+
+        # to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1
+        if self.model.visual.input_patchnorm:
+            # einops - rearrange(x, 'b c (h p1) (w p2) -> b (h w) (c p1 p2)')
+            x = x.reshape(x.shape[0], x.shape[1], self.model.visual.grid_size[0], self.model.visual.patch_size[0], self.model.visual.grid_size[1], self.model.visual.patch_size[1])
+            x = x.permute(0, 2, 4, 1, 3, 5)
+            x = x.reshape(x.shape[0], self.model.visual.grid_size[0] * self.model.visual.grid_size[1], -1)
+            x = self.model.visual.patchnorm_pre_ln(x)
+            x = self.model.visual.conv1(x)
+        else:
+            x = self.model.visual.conv1(x)  # shape = [*, width, grid, grid]
+            x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+            x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+
+        # class embeddings and positional embeddings
+        x = torch.cat(
+            [self.model.visual.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device),
+             x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.model.visual.positional_embedding.to(x.dtype)
+
+        # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
+        x = self.model.visual.patch_dropout(x)
+        x = self.model.visual.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.model.visual.transformer(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        return x
+
+
+class FrozenCLIPT5Encoder(AbstractEncoder):
+    def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda",
+                 clip_max_length=77, t5_max_length=77):
+        super().__init__()
+        self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length)
+        self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length)
+        print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, "
+              f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params.")
+
+    def encode(self, text):
+        return self(text)
+
+    def forward(self, text):
+        clip_z = self.clip_encoder.encode(text)
+        t5_z = self.t5_encoder.encode(text)
+        return [clip_z, t5_z]

lvdm/modules/encoders/ip_resampler.py

@@ -0,0 +1,136 @@
+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+import math
+import torch
+import torch.nn as nn
+
+
+class ImageProjModel(nn.Module):
+    """Projection Model"""
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
+        super().__init__()        
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = nn.LayerNorm(cross_attention_dim)
+        
+    def forward(self, image_embeds):
+        #embeds = image_embeds
+        embeds = image_embeds.type(list(self.proj.parameters())[0].dtype)
+        clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim)
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+    
+    
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    #(bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+        
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+        
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+        
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+    ):
+        super().__init__()
+        
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+        
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+        
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        
+        latents = self.latents.repeat(x.size(0), 1, 1)
+        
+        x = self.proj_in(x)
+        
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+            
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)

lvdm/modules/networks/ae_modules.py

@@ -0,0 +1,845 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import numpy as np
+import torch.nn as nn
+from einops import rearrange
+from utils.utils import instantiate_from_config
+from lvdm.modules.attention import LinearAttention
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+
+class LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w) # bcl
+        q = q.permute(0,2,1)   # bcl -> blc l=hw
+        k = k.reshape(b,c,h*w) # bcl
+        
+        w_ = torch.bmm(q,k)    # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+def make_attn(in_channels, attn_type="vanilla"):
+    assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
+    #print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        return AttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+ 
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        self.in_channels = in_channels
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        self.in_channels = in_channels
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList([
+                torch.nn.Linear(self.ch,
+                                self.temb_ch),
+                torch.nn.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # print(f'encoder-input={x.shape}')
+        # downsampling
+        hs = [self.conv_in(x)]
+        # print(f'encoder-conv in feat={hs[0].shape}')
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                # print(f'encoder-down feat={h.shape}')
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                # print(f'encoder-downsample (input)={hs[-1].shape}')
+                hs.append(self.down[i_level].downsample(hs[-1]))
+                # print(f'encoder-downsample (output)={hs[-1].shape}')
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        # print(f'encoder-mid1 feat={h.shape}')
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+        # print(f'encoder-mid2 feat={h.shape}')
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        # print(f'end feat={h.shape}')
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 attn_type="vanilla", **ignorekwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        print("AE working on z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # print(f'decoder-input={z.shape}')
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+        # print(f'decoder-conv in feat={h.shape}')
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+        # print(f'decoder-mid feat={h.shape}')
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+                # print(f'decoder-up feat={h.shape}')
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+                # print(f'decoder-upsample feat={h.shape}')
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        # print(f'decoder-conv_out feat={h.shape}')
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
+                                     ResnetBlock(in_channels=in_channels,
+                                                 out_channels=2 * in_channels,
+                                                 temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=2 * in_channels,
+                                                out_channels=4 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=4 * in_channels,
+                                                out_channels=2 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     nn.Conv2d(2*in_channels, in_channels, 1),
+                                     Upsample(in_channels, with_conv=True)])
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(in_channels,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1,2,3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
+                 ch_mult=(2,2), dropout=0.0):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(in_channels,
+                                 mid_channels,
+                                 kernel_size=3,
+                                 stride=1,
+                                 padding=1)
+        self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+
+        self.conv_out = nn.Conv2d(mid_channels,
+                                  out_channels,
+                                  kernel_size=1,
+                                  )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
+        x = self.attn(x)
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True,
+                 ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
+                               z_channels=intermediate_chn, double_z=False, resolution=resolution,
+                               attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
+                               out_ch=None)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
+                                       mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
+                 dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        tmp_chn = z_channels*ch_mult[-1]
+        self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
+                               resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
+                               ch_mult=ch_mult, resolution=resolution, ch=ch)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
+                                       out_channels=tmp_chn, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size//in_size))+1
+        factor_up = 1.+ (out_size % in_size)
+        print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
+        self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
+                                       out_channels=in_channels)
+        self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
+                               attn_resolutions=[], in_channels=None, ch=in_channels,
+                               ch_mult=[ch_mult for _ in range(num_blocks)])
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=4,
+                                        stride=2,
+                                        padding=1)
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor==1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
+        return x
+
+class FirstStagePostProcessor(nn.Module):
+
+    def __init__(self, ch_mult:list, in_channels,
+                 pretrained_model:nn.Module=None,
+                 reshape=False,
+                 n_channels=None,
+                 dropout=0.,
+                 pretrained_config=None):
+        super().__init__()
+        if pretrained_config is None:
+            assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.pretrained_model = pretrained_model
+        else:
+            assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.instantiate_pretrained(pretrained_config)
+
+        self.do_reshape = reshape
+
+        if n_channels is None:
+            n_channels = self.pretrained_model.encoder.ch
+
+        self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
+        self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
+                            stride=1,padding=1)
+
+        blocks = []
+        downs = []
+        ch_in = n_channels
+        for m in ch_mult:
+            blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
+            ch_in = m * n_channels
+            downs.append(Downsample(ch_in, with_conv=False))
+
+        self.model = nn.ModuleList(blocks)
+        self.downsampler = nn.ModuleList(downs)
+
+
+    def instantiate_pretrained(self, config):
+        model = instantiate_from_config(config)
+        self.pretrained_model = model.eval()
+        # self.pretrained_model.train = False
+        for param in self.pretrained_model.parameters():
+            param.requires_grad = False
+
+
+    @torch.no_grad()
+    def encode_with_pretrained(self,x):
+        c = self.pretrained_model.encode(x)
+        if isinstance(c, DiagonalGaussianDistribution):
+            c = c.mode()
+        return  c
+
+    def forward(self,x):
+        z_fs = self.encode_with_pretrained(x)
+        z = self.proj_norm(z_fs)
+        z = self.proj(z)
+        z = nonlinearity(z)
+
+        for submodel, downmodel in zip(self.model,self.downsampler):
+            z = submodel(z,temb=None)
+            z = downmodel(z)
+
+        if self.do_reshape:
+            z = rearrange(z,'b c h w -> b (h w) c')
+        return z
+

lvdm/modules/networks/openaimodel3d.py

@@ -0,0 +1,577 @@
+from functools import partial
+from abc import abstractmethod
+import torch
+import torch.nn as nn
+from einops import rearrange
+import torch.nn.functional as F
+from lvdm.models.utils_diffusion import timestep_embedding
+from lvdm.common import checkpoint
+from lvdm.basics import (
+    zero_module,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    normalization
+)
+from lvdm.modules.attention import SpatialTransformer, TemporalTransformer
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None, batch_size=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb, batch_size)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            elif isinstance(layer, TemporalTransformer):
+                x = rearrange(x, '(b f) c h w -> b c f h w', b=batch_size)
+                x = layer(x, context)
+                x = rearrange(x, 'b c f h w -> (b f) c h w')
+            else:
+                x = layer(x,)
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode='nearest')
+        else:
+            x = F.interpolate(x, scale_factor=2, mode='nearest')
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        use_conv=False,
+        up=False,
+        down=False,
+        use_temporal_conv=False,
+        tempspatial_aware=False
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+        self.use_temporal_conv = use_temporal_conv
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(nn.Conv2d(self.out_channels, self.out_channels, 3, padding=1)),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 3, padding=1)
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+        if self.use_temporal_conv:
+            self.temopral_conv = TemporalConvBlock(
+                self.out_channels,
+                self.out_channels,
+                dropout=0.1,
+                spatial_aware=tempspatial_aware
+            )
+
+    def forward(self, x, emb,  batch_size=None):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        input_tuple = (x, emb,)
+        if batch_size:
+            forward_batchsize = partial(self._forward, batch_size=batch_size)
+            return checkpoint(forward_batchsize, input_tuple, self.parameters(), self.use_checkpoint)
+        return checkpoint(self._forward, input_tuple, self.parameters(), self.use_checkpoint)
+
+    def _forward(self, x, emb,  batch_size=None,):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = torch.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        h = self.skip_connection(x) + h
+        
+        if self.use_temporal_conv and batch_size:
+            h = rearrange(h, '(b t) c h w -> b c t h w', b=batch_size)
+            h = self.temopral_conv(h)
+            h = rearrange(h, 'b c t h w -> (b t) c h w')
+        return h
+
+
+class TemporalConvBlock(nn.Module):
+    """
+    Adapted from modelscope: https://github.com/modelscope/modelscope/blob/master/modelscope/models/multi_modal/video_synthesis/unet_sd.py
+    """
+
+    def __init__(self, in_channels, out_channels=None, dropout=0.0, spatial_aware=False):
+        super(TemporalConvBlock, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        kernel_shape = (3, 1, 1) if not spatial_aware else (3, 3, 3)
+        padding_shape = (1, 0, 0) if not spatial_aware else (1, 1, 1)
+
+        # conv layers
+        self.conv1 = nn.Sequential(
+            nn.GroupNorm(32, in_channels), nn.SiLU(),
+            nn.Conv3d(in_channels, out_channels, kernel_shape, padding=padding_shape))
+        self.conv2 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, kernel_shape, padding=padding_shape))
+        self.conv3 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, (3, 1, 1), padding=(1, 0, 0)))
+        self.conv4 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, (3, 1, 1), padding=(1, 0, 0)))
+
+        # zero out the last layer params,so the conv block is identity
+        nn.init.zeros_(self.conv4[-1].weight)
+        nn.init.zeros_(self.conv4[-1].bias)
+
+    def forward(self, x):
+        identity = x
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.conv3(x)
+        x = self.conv4(x)
+
+        return x + identity
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: in_channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 model_channels,
+                 out_channels,
+                 num_res_blocks,
+                 attention_resolutions,
+                 dropout=0.0,
+                 channel_mult=(1, 2, 4, 8),
+                 conv_resample=True,
+                 dims=2,
+                 context_dim=None,
+                 use_scale_shift_norm=False,
+                 resblock_updown=False,
+                 num_heads=-1,
+                 num_head_channels=-1,
+                 transformer_depth=1,
+                 use_linear=False,
+                 use_checkpoint=False,
+                 temporal_conv=False,
+                 tempspatial_aware=False,
+                 temporal_attention=True,
+                 temporal_selfatt_only=True,
+                 use_relative_position=True,
+                 use_causal_attention=False,
+                 temporal_length=None,
+                 use_fp16=False,
+                 addition_attention=False,
+                 use_image_attention=False,
+                 temporal_transformer_depth=1,
+                 fps_cond=False,
+                ):
+        super(UNetModel, self).__init__()
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.temporal_attention = temporal_attention
+        time_embed_dim = model_channels * 4
+        self.use_checkpoint = use_checkpoint
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+        self.addition_attention=addition_attention
+        self.use_image_attention = use_image_attention
+        self.fps_cond=fps_cond
+
+
+
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+        if self.fps_cond:
+            self.fps_embedding = nn.Sequential(
+                linear(model_channels, time_embed_dim),
+                nn.SiLU(),
+                linear(time_embed_dim, time_embed_dim),
+            )
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1))
+            ]
+        )
+        if self.addition_attention:
+            self.init_attn=TimestepEmbedSequential(
+                TemporalTransformer(
+                    model_channels,
+                    n_heads=8,
+                    d_head=num_head_channels,
+                    depth=transformer_depth,
+                    context_dim=context_dim,
+                    use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                    causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                    temporal_length=temporal_length))
+            
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(ch, time_embed_dim, dropout,
+                        out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                        use_temporal_conv=temporal_conv
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    layers.append(
+                        SpatialTransformer(ch, num_heads, dim_head, 
+                            depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                            use_checkpoint=use_checkpoint, disable_self_attn=False,
+                            img_cross_attention=self.use_image_attention
+                        )
+                    )
+                    if self.temporal_attention:
+                        layers.append(
+                            TemporalTransformer(ch, num_heads, dim_head,
+                                depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                                use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                                causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                                temporal_length=temporal_length
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(ch, time_embed_dim, dropout, 
+                            out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True
+                        )
+                        if resblock_updown
+                        else Downsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        layers = [
+            ResBlock(ch, time_embed_dim, dropout,
+                dims=dims, use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                use_temporal_conv=temporal_conv
+            ),
+            SpatialTransformer(ch, num_heads, dim_head, 
+                depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                use_checkpoint=use_checkpoint, disable_self_attn=False,
+                img_cross_attention=self.use_image_attention
+            )
+        ]
+        if self.temporal_attention:
+            layers.append(
+                TemporalTransformer(ch, num_heads, dim_head,
+                    depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                    use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                    causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                    temporal_length=temporal_length
+                )
+            )
+        layers.append(
+            ResBlock(ch, time_embed_dim, dropout,
+                dims=dims, use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                use_temporal_conv=temporal_conv
+                )
+        )
+        self.middle_block = TimestepEmbedSequential(*layers)
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(num_res_blocks + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(ch + ich, time_embed_dim, dropout,
+                        out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                        use_temporal_conv=temporal_conv
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    layers.append(
+                        SpatialTransformer(ch, num_heads, dim_head, 
+                            depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                            use_checkpoint=use_checkpoint, disable_self_attn=False,
+                            img_cross_attention=self.use_image_attention
+                        )
+                    )
+                    if self.temporal_attention:
+                        layers.append(
+                            TemporalTransformer(ch, num_heads, dim_head,
+                                depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                                use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                                causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                                temporal_length=temporal_length
+                            )
+                        )
+                if level and i == num_res_blocks:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(ch, time_embed_dim, dropout,
+                            out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+
+    def forward(self, x, timesteps, context=None, features_adapter=None, fps=16, **kwargs):
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        if self.fps_cond:
+            if type(fps) == int:
+                fps = torch.full_like(timesteps, fps)
+            fps_emb = timestep_embedding(fps,self.model_channels, repeat_only=False)
+            emb += self.fps_embedding(fps_emb)
+
+        b,_,t,_,_ = x.shape
+        ## repeat t times for context [(b t) 77 768] & time embedding
+        context = context.repeat_interleave(repeats=t, dim=0)
+        emb = emb.repeat_interleave(repeats=t, dim=0)
+
+        ## always in shape (b t) c h w, except for temporal layer
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+
+        h = x.type(self.dtype)
+        adapter_idx = 0
+        hs = []
+        for id, module in enumerate(self.input_blocks):
+            h = module(h, emb, context=context, batch_size=b)
+            if id ==0 and self.addition_attention:
+                h = self.init_attn(h, emb, context=context, batch_size=b)
+            ## plug-in adapter features
+            if ((id+1)%3 == 0) and features_adapter is not None:
+                h = h + features_adapter[adapter_idx]
+                adapter_idx += 1
+            hs.append(h)
+        if features_adapter is not None:
+            assert len(features_adapter)==adapter_idx, 'Wrong features_adapter'
+
+        h = self.middle_block(h, emb, context=context, batch_size=b)
+        for module in self.output_blocks:
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context=context, batch_size=b)
+        h = h.type(x.dtype)
+        y = self.out(h)
+        
+        # reshape back to (b c t h w)
+        y = rearrange(y, '(b t) c h w -> b c t h w', b=b)
+        return y
+    

lvdm/modules/networks/openaimodel3d_freetraj.py

@@ -0,0 +1,577 @@
+from functools import partial
+from abc import abstractmethod
+import torch
+import torch.nn as nn
+from einops import rearrange
+import torch.nn.functional as F
+from lvdm.models.utils_diffusion import timestep_embedding
+from lvdm.common import checkpoint
+from lvdm.basics import (
+    zero_module,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    normalization
+)
+from lvdm.modules.attention_freetraj import SpatialTransformer, TemporalTransformer
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None, batch_size=None, use_freetraj=False, **kwargs):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb, batch_size)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context, use_freetraj=use_freetraj, **kwargs)
+            elif isinstance(layer, TemporalTransformer):
+                x = rearrange(x, '(b f) c h w -> b c f h w', b=batch_size)
+                x = layer(x, context, use_freetraj=use_freetraj, **kwargs)
+                x = rearrange(x, 'b c f h w -> (b f) c h w')
+            else:
+                x = layer(x,)
+        return x
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode='nearest')
+        else:
+            x = F.interpolate(x, scale_factor=2, mode='nearest')
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        use_conv=False,
+        up=False,
+        down=False,
+        use_temporal_conv=False,
+        tempspatial_aware=False
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+        self.use_temporal_conv = use_temporal_conv
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(nn.Conv2d(self.out_channels, self.out_channels, 3, padding=1)),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 3, padding=1)
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+        if self.use_temporal_conv:
+            self.temopral_conv = TemporalConvBlock(
+                self.out_channels,
+                self.out_channels,
+                dropout=0.1,
+                spatial_aware=tempspatial_aware
+            )
+
+    def forward(self, x, emb,  batch_size=None):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        input_tuple = (x, emb,)
+        if batch_size:
+            forward_batchsize = partial(self._forward, batch_size=batch_size)
+            return checkpoint(forward_batchsize, input_tuple, self.parameters(), self.use_checkpoint)
+        return checkpoint(self._forward, input_tuple, self.parameters(), self.use_checkpoint)
+
+    def _forward(self, x, emb,  batch_size=None,):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = torch.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        h = self.skip_connection(x) + h
+        
+        if self.use_temporal_conv and batch_size:
+            h = rearrange(h, '(b t) c h w -> b c t h w', b=batch_size)
+            h = self.temopral_conv(h)
+            h = rearrange(h, 'b c t h w -> (b t) c h w')
+        return h
+
+
+class TemporalConvBlock(nn.Module):
+    """
+    Adapted from modelscope: https://github.com/modelscope/modelscope/blob/master/modelscope/models/multi_modal/video_synthesis/unet_sd.py
+    """
+
+    def __init__(self, in_channels, out_channels=None, dropout=0.0, spatial_aware=False):
+        super(TemporalConvBlock, self).__init__()
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        kernel_shape = (3, 1, 1) if not spatial_aware else (3, 3, 3)
+        padding_shape = (1, 0, 0) if not spatial_aware else (1, 1, 1)
+
+        # conv layers
+        self.conv1 = nn.Sequential(
+            nn.GroupNorm(32, in_channels), nn.SiLU(),
+            nn.Conv3d(in_channels, out_channels, kernel_shape, padding=padding_shape))
+        self.conv2 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, kernel_shape, padding=padding_shape))
+        self.conv3 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, (3, 1, 1), padding=(1, 0, 0)))
+        self.conv4 = nn.Sequential(
+            nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
+            nn.Conv3d(out_channels, in_channels, (3, 1, 1), padding=(1, 0, 0)))
+
+        # zero out the last layer params,so the conv block is identity
+        nn.init.zeros_(self.conv4[-1].weight)
+        nn.init.zeros_(self.conv4[-1].bias)
+
+    def forward(self, x):
+        identity = x
+        x = self.conv1(x)
+        x = self.conv2(x)
+        x = self.conv3(x)
+        x = self.conv4(x)
+
+        return x + identity
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: in_channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 model_channels,
+                 out_channels,
+                 num_res_blocks,
+                 attention_resolutions,
+                 dropout=0.0,
+                 channel_mult=(1, 2, 4, 8),
+                 conv_resample=True,
+                 dims=2,
+                 context_dim=None,
+                 use_scale_shift_norm=False,
+                 resblock_updown=False,
+                 num_heads=-1,
+                 num_head_channels=-1,
+                 transformer_depth=1,
+                 use_linear=False,
+                 use_checkpoint=False,
+                 temporal_conv=False,
+                 tempspatial_aware=False,
+                 temporal_attention=True,
+                 temporal_selfatt_only=True,
+                 use_relative_position=True,
+                 use_causal_attention=False,
+                 temporal_length=None,
+                 use_fp16=False,
+                 addition_attention=False,
+                 use_image_attention=False,
+                 temporal_transformer_depth=1,
+                 fps_cond=False,
+                ):
+        super(UNetModel, self).__init__()
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.temporal_attention = temporal_attention
+        time_embed_dim = model_channels * 4
+        self.use_checkpoint = use_checkpoint
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+        self.addition_attention=addition_attention
+        self.use_image_attention = use_image_attention
+        self.fps_cond=fps_cond
+
+
+
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+        if self.fps_cond:
+            self.fps_embedding = nn.Sequential(
+                linear(model_channels, time_embed_dim),
+                nn.SiLU(),
+                linear(time_embed_dim, time_embed_dim),
+            )
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1))
+            ]
+        )
+        if self.addition_attention:
+            self.init_attn=TimestepEmbedSequential(
+                TemporalTransformer(
+                    model_channels,
+                    n_heads=8,
+                    d_head=num_head_channels,
+                    depth=transformer_depth,
+                    context_dim=context_dim,
+                    use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                    causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                    temporal_length=temporal_length))
+            
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(ch, time_embed_dim, dropout,
+                        out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                        use_temporal_conv=temporal_conv
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    layers.append(
+                        SpatialTransformer(ch, num_heads, dim_head, 
+                            depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                            use_checkpoint=use_checkpoint, disable_self_attn=False,
+                            img_cross_attention=self.use_image_attention
+                        )
+                    )
+                    if self.temporal_attention:
+                        layers.append(
+                            TemporalTransformer(ch, num_heads, dim_head,
+                                depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                                use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                                causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                                temporal_length=temporal_length
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(ch, time_embed_dim, dropout, 
+                            out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True
+                        )
+                        if resblock_updown
+                        else Downsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        layers = [
+            ResBlock(ch, time_embed_dim, dropout,
+                dims=dims, use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                use_temporal_conv=temporal_conv
+            ),
+            SpatialTransformer(ch, num_heads, dim_head, 
+                depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                use_checkpoint=use_checkpoint, disable_self_attn=False,
+                img_cross_attention=self.use_image_attention
+            )
+        ]
+        if self.temporal_attention:
+            layers.append(
+                TemporalTransformer(ch, num_heads, dim_head,
+                    depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                    use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                    causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                    temporal_length=temporal_length
+                )
+            )
+        layers.append(
+            ResBlock(ch, time_embed_dim, dropout,
+                dims=dims, use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                use_temporal_conv=temporal_conv
+                )
+        )
+        self.middle_block = TimestepEmbedSequential(*layers)
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(num_res_blocks + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(ch + ich, time_embed_dim, dropout,
+                        out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
+                        use_temporal_conv=temporal_conv
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    layers.append(
+                        SpatialTransformer(ch, num_heads, dim_head, 
+                            depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                            use_checkpoint=use_checkpoint, disable_self_attn=False,
+                            img_cross_attention=self.use_image_attention
+                        )
+                    )
+                    if self.temporal_attention:
+                        layers.append(
+                            TemporalTransformer(ch, num_heads, dim_head,
+                                depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
+                                use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only, 
+                                causal_attention=use_causal_attention, relative_position=use_relative_position, 
+                                temporal_length=temporal_length
+                            )
+                        )
+                if level and i == num_res_blocks:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(ch, time_embed_dim, dropout,
+                            out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+
+    def forward(self, x, timesteps, context=None, features_adapter=None, fps=16, **kwargs):
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        if self.fps_cond:
+            if type(fps) == int:
+                fps = torch.full_like(timesteps, fps)
+            fps_emb = timestep_embedding(fps,self.model_channels, repeat_only=False)
+            emb += self.fps_embedding(fps_emb)
+
+        b,_,t,_,_ = x.shape
+        ## repeat t times for context [(b t) 77 768] & time embedding
+        context = context.repeat_interleave(repeats=t, dim=0)
+        emb = emb.repeat_interleave(repeats=t, dim=0)
+
+        ## always in shape (b t) c h w, except for temporal layer
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+
+        h = x.type(self.dtype)
+        adapter_idx = 0
+        hs = []
+        for id, module in enumerate(self.input_blocks):
+            h = module(h, emb, context=context, batch_size=b, **kwargs)
+            if id ==0 and self.addition_attention:
+                h = self.init_attn(h, emb, context=context, batch_size=b, **kwargs)
+            ## plug-in adapter features
+            if ((id+1)%3 == 0) and features_adapter is not None:
+                h = h + features_adapter[adapter_idx]
+                adapter_idx += 1
+            hs.append(h)
+        if features_adapter is not None:
+            assert len(features_adapter)==adapter_idx, 'Wrong features_adapter'
+
+        h = self.middle_block(h, emb, context=context, batch_size=b, **kwargs)
+        for module in self.output_blocks:
+            h = torch.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context=context, batch_size=b, **kwargs)
+        h = h.type(x.dtype)
+        y = self.out(h)
+        
+        # reshape back to (b c t h w)
+        y = rearrange(y, '(b t) c h w -> b c t h w', b=b)
+        return y
+    

lvdm/modules/x_transformer.py

@@ -0,0 +1,640 @@
+"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers"""
+from functools import partial
+from inspect import isfunction
+from collections import namedtuple
+from einops import rearrange, repeat
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+
+# constants
+DEFAULT_DIM_HEAD = 64
+
+Intermediates = namedtuple('Intermediates', [
+    'pre_softmax_attn',
+    'post_softmax_attn'
+])
+
+LayerIntermediates = namedtuple('Intermediates', [
+    'hiddens',
+    'attn_intermediates'
+])
+
+
+class AbsolutePositionalEmbedding(nn.Module):
+    def __init__(self, dim, max_seq_len):
+        super().__init__()
+        self.emb = nn.Embedding(max_seq_len, dim)
+        self.init_()
+
+    def init_(self):
+        nn.init.normal_(self.emb.weight, std=0.02)
+
+    def forward(self, x):
+        n = torch.arange(x.shape[1], device=x.device)
+        return self.emb(n)[None, :, :]
+
+
+class FixedPositionalEmbedding(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer('inv_freq', inv_freq)
+
+    def forward(self, x, seq_dim=1, offset=0):
+        t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset
+        sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq)
+        emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
+        return emb[None, :, :]
+
+
+# helpers
+
+def exists(val):
+    return val is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def always(val):
+    def inner(*args, **kwargs):
+        return val
+    return inner
+
+
+def not_equals(val):
+    def inner(x):
+        return x != val
+    return inner
+
+
+def equals(val):
+    def inner(x):
+        return x == val
+    return inner
+
+
+def max_neg_value(tensor):
+    return -torch.finfo(tensor.dtype).max
+
+
+# keyword argument helpers
+
+def pick_and_pop(keys, d):
+    values = list(map(lambda key: d.pop(key), keys))
+    return dict(zip(keys, values))
+
+
+def group_dict_by_key(cond, d):
+    return_val = [dict(), dict()]
+    for key in d.keys():
+        match = bool(cond(key))
+        ind = int(not match)
+        return_val[ind][key] = d[key]
+    return (*return_val,)
+
+
+def string_begins_with(prefix, str):
+    return str.startswith(prefix)
+
+
+def group_by_key_prefix(prefix, d):
+    return group_dict_by_key(partial(string_begins_with, prefix), d)
+
+
+def groupby_prefix_and_trim(prefix, d):
+    kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
+    kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
+    return kwargs_without_prefix, kwargs
+
+
+# classes
+class Scale(nn.Module):
+    def __init__(self, value, fn):
+        super().__init__()
+        self.value = value
+        self.fn = fn
+
+    def forward(self, x, **kwargs):
+        x, *rest = self.fn(x, **kwargs)
+        return (x * self.value, *rest)
+
+
+class Rezero(nn.Module):
+    def __init__(self, fn):
+        super().__init__()
+        self.fn = fn
+        self.g = nn.Parameter(torch.zeros(1))
+
+    def forward(self, x, **kwargs):
+        x, *rest = self.fn(x, **kwargs)
+        return (x * self.g, *rest)
+
+
+class ScaleNorm(nn.Module):
+    def __init__(self, dim, eps=1e-5):
+        super().__init__()
+        self.scale = dim ** -0.5
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(1))
+
+    def forward(self, x):
+        norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
+        return x / norm.clamp(min=self.eps) * self.g
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-8):
+        super().__init__()
+        self.scale = dim ** -0.5
+        self.eps = eps
+        self.g = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
+        return x / norm.clamp(min=self.eps) * self.g
+
+
+class Residual(nn.Module):
+    def forward(self, x, residual):
+        return x + residual
+
+
+class GRUGating(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.gru = nn.GRUCell(dim, dim)
+
+    def forward(self, x, residual):
+        gated_output = self.gru(
+            rearrange(x, 'b n d -> (b n) d'),
+            rearrange(residual, 'b n d -> (b n) d')
+        )
+
+        return gated_output.reshape_as(x)
+
+
+# feedforward
+
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+# attention.
+class Attention(nn.Module):
+    def __init__(
+            self,
+            dim,
+            dim_head=DEFAULT_DIM_HEAD,
+            heads=8,
+            causal=False,
+            mask=None,
+            talking_heads=False,
+            sparse_topk=None,
+            use_entmax15=False,
+            num_mem_kv=0,
+            dropout=0.,
+            on_attn=False
+    ):
+        super().__init__()
+        if use_entmax15:
+            raise NotImplementedError("Check out entmax activation instead of softmax activation!")
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+        self.causal = causal
+        self.mask = mask
+
+        inner_dim = dim_head * heads
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(dim, inner_dim, bias=False)
+        self.dropout = nn.Dropout(dropout)
+
+        # talking heads
+        self.talking_heads = talking_heads
+        if talking_heads:
+            self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads))
+            self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads))
+
+        # explicit topk sparse attention
+        self.sparse_topk = sparse_topk
+
+        # entmax
+        #self.attn_fn = entmax15 if use_entmax15 else F.softmax
+        self.attn_fn = F.softmax
+
+        # add memory key / values
+        self.num_mem_kv = num_mem_kv
+        if num_mem_kv > 0:
+            self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
+            self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
+
+        # attention on attention
+        self.attn_on_attn = on_attn
+        self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim)
+
+    def forward(
+            self,
+            x,
+            context=None,
+            mask=None,
+            context_mask=None,
+            rel_pos=None,
+            sinusoidal_emb=None,
+            prev_attn=None,
+            mem=None
+    ):
+        b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device
+        kv_input = default(context, x)
+
+        q_input = x
+        k_input = kv_input
+        v_input = kv_input
+
+        if exists(mem):
+            k_input = torch.cat((mem, k_input), dim=-2)
+            v_input = torch.cat((mem, v_input), dim=-2)
+
+        if exists(sinusoidal_emb):
+            # in shortformer, the query would start at a position offset depending on the past cached memory
+            offset = k_input.shape[-2] - q_input.shape[-2]
+            q_input = q_input + sinusoidal_emb(q_input, offset=offset)
+            k_input = k_input + sinusoidal_emb(k_input)
+
+        q = self.to_q(q_input)
+        k = self.to_k(k_input)
+        v = self.to_v(v_input)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))
+
+        input_mask = None
+        if any(map(exists, (mask, context_mask))):
+            q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool())
+            k_mask = q_mask if not exists(context) else context_mask
+            k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool())
+            q_mask = rearrange(q_mask, 'b i -> b () i ()')
+            k_mask = rearrange(k_mask, 'b j -> b () () j')
+            input_mask = q_mask * k_mask
+
+        if self.num_mem_kv > 0:
+            mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v))
+            k = torch.cat((mem_k, k), dim=-2)
+            v = torch.cat((mem_v, v), dim=-2)
+            if exists(input_mask):
+                input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True)
+
+        dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
+        mask_value = max_neg_value(dots)
+
+        if exists(prev_attn):
+            dots = dots + prev_attn
+
+        pre_softmax_attn = dots
+
+        if talking_heads:
+            dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous()
+
+        if exists(rel_pos):
+            dots = rel_pos(dots)
+
+        if exists(input_mask):
+            dots.masked_fill_(~input_mask, mask_value)
+            del input_mask
+
+        if self.causal:
+            i, j = dots.shape[-2:]
+            r = torch.arange(i, device=device)
+            mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j')
+            mask = F.pad(mask, (j - i, 0), value=False)
+            dots.masked_fill_(mask, mask_value)
+            del mask
+
+        if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]:
+            top, _ = dots.topk(self.sparse_topk, dim=-1)
+            vk = top[..., -1].unsqueeze(-1).expand_as(dots)
+            mask = dots < vk
+            dots.masked_fill_(mask, mask_value)
+            del mask
+
+        attn = self.attn_fn(dots, dim=-1)
+        post_softmax_attn = attn
+
+        attn = self.dropout(attn)
+
+        if talking_heads:
+            attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous()
+
+        out = einsum('b h i j, b h j d -> b h i d', attn, v)
+        out = rearrange(out, 'b h n d -> b n (h d)')
+
+        intermediates = Intermediates(
+            pre_softmax_attn=pre_softmax_attn,
+            post_softmax_attn=post_softmax_attn
+        )
+
+        return self.to_out(out), intermediates
+
+
+class AttentionLayers(nn.Module):
+    def __init__(
+            self,
+            dim,
+            depth,
+            heads=8,
+            causal=False,
+            cross_attend=False,
+            only_cross=False,
+            use_scalenorm=False,
+            use_rmsnorm=False,
+            use_rezero=False,
+            rel_pos_num_buckets=32,
+            rel_pos_max_distance=128,
+            position_infused_attn=False,
+            custom_layers=None,
+            sandwich_coef=None,
+            par_ratio=None,
+            residual_attn=False,
+            cross_residual_attn=False,
+            macaron=False,
+            pre_norm=True,
+            gate_residual=False,
+            **kwargs
+    ):
+        super().__init__()
+        ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs)
+        attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs)
+
+        dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
+
+        self.dim = dim
+        self.depth = depth
+        self.layers = nn.ModuleList([])
+
+        self.has_pos_emb = position_infused_attn
+        self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None
+        self.rotary_pos_emb = always(None)
+
+        assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
+        self.rel_pos = None
+
+        self.pre_norm = pre_norm
+
+        self.residual_attn = residual_attn
+        self.cross_residual_attn = cross_residual_attn
+
+        norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm
+        norm_class = RMSNorm if use_rmsnorm else norm_class
+        norm_fn = partial(norm_class, dim)
+
+        norm_fn = nn.Identity if use_rezero else norm_fn
+        branch_fn = Rezero if use_rezero else None
+
+        if cross_attend and not only_cross:
+            default_block = ('a', 'c', 'f')
+        elif cross_attend and only_cross:
+            default_block = ('c', 'f')
+        else:
+            default_block = ('a', 'f')
+
+        if macaron:
+            default_block = ('f',) + default_block
+
+        if exists(custom_layers):
+            layer_types = custom_layers
+        elif exists(par_ratio):
+            par_depth = depth * len(default_block)
+            assert 1 < par_ratio <= par_depth, 'par ratio out of range'
+            default_block = tuple(filter(not_equals('f'), default_block))
+            par_attn = par_depth // par_ratio
+            depth_cut = par_depth * 2 // 3  # 2 / 3 attention layer cutoff suggested by PAR paper
+            par_width = (depth_cut + depth_cut // par_attn) // par_attn
+            assert len(default_block) <= par_width, 'default block is too large for par_ratio'
+            par_block = default_block + ('f',) * (par_width - len(default_block))
+            par_head = par_block * par_attn
+            layer_types = par_head + ('f',) * (par_depth - len(par_head))
+        elif exists(sandwich_coef):
+            assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
+            layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
+        else:
+            layer_types = default_block * depth
+
+        self.layer_types = layer_types
+        self.num_attn_layers = len(list(filter(equals('a'), layer_types)))
+
+        for layer_type in self.layer_types:
+            if layer_type == 'a':
+                layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs)
+            elif layer_type == 'c':
+                layer = Attention(dim, heads=heads, **attn_kwargs)
+            elif layer_type == 'f':
+                layer = FeedForward(dim, **ff_kwargs)
+                layer = layer if not macaron else Scale(0.5, layer)
+            else:
+                raise Exception(f'invalid layer type {layer_type}')
+
+            if isinstance(layer, Attention) and exists(branch_fn):
+                layer = branch_fn(layer)
+
+            if gate_residual:
+                residual_fn = GRUGating(dim)
+            else:
+                residual_fn = Residual()
+
+            self.layers.append(nn.ModuleList([
+                norm_fn(),
+                layer,
+                residual_fn
+            ]))
+
+    def forward(
+            self,
+            x,
+            context=None,
+            mask=None,
+            context_mask=None,
+            mems=None,
+            return_hiddens=False
+    ):
+        hiddens = []
+        intermediates = []
+        prev_attn = None
+        prev_cross_attn = None
+
+        mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers
+
+        for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)):
+            is_last = ind == (len(self.layers) - 1)
+
+            if layer_type == 'a':
+                hiddens.append(x)
+                layer_mem = mems.pop(0)
+
+            residual = x
+
+            if self.pre_norm:
+                x = norm(x)
+
+            if layer_type == 'a':
+                out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos,
+                                   prev_attn=prev_attn, mem=layer_mem)
+            elif layer_type == 'c':
+                out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn)
+            elif layer_type == 'f':
+                out = block(x)
+
+            x = residual_fn(out, residual)
+
+            if layer_type in ('a', 'c'):
+                intermediates.append(inter)
+
+            if layer_type == 'a' and self.residual_attn:
+                prev_attn = inter.pre_softmax_attn
+            elif layer_type == 'c' and self.cross_residual_attn:
+                prev_cross_attn = inter.pre_softmax_attn
+
+            if not self.pre_norm and not is_last:
+                x = norm(x)
+
+        if return_hiddens:
+            intermediates = LayerIntermediates(
+                hiddens=hiddens,
+                attn_intermediates=intermediates
+            )
+
+            return x, intermediates
+
+        return x
+
+
+class Encoder(AttentionLayers):
+    def __init__(self, **kwargs):
+        assert 'causal' not in kwargs, 'cannot set causality on encoder'
+        super().__init__(causal=False, **kwargs)
+
+
+
+class TransformerWrapper(nn.Module):
+    def __init__(
+            self,
+            *,
+            num_tokens,
+            max_seq_len,
+            attn_layers,
+            emb_dim=None,
+            max_mem_len=0.,
+            emb_dropout=0.,
+            num_memory_tokens=None,
+            tie_embedding=False,
+            use_pos_emb=True
+    ):
+        super().__init__()
+        assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'
+
+        dim = attn_layers.dim
+        emb_dim = default(emb_dim, dim)
+
+        self.max_seq_len = max_seq_len
+        self.max_mem_len = max_mem_len
+        self.num_tokens = num_tokens
+
+        self.token_emb = nn.Embedding(num_tokens, emb_dim)
+        self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if (
+                    use_pos_emb and not attn_layers.has_pos_emb) else always(0)
+        self.emb_dropout = nn.Dropout(emb_dropout)
+
+        self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity()
+        self.attn_layers = attn_layers
+        self.norm = nn.LayerNorm(dim)
+
+        self.init_()
+
+        self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t()
+
+        # memory tokens (like [cls]) from Memory Transformers paper
+        num_memory_tokens = default(num_memory_tokens, 0)
+        self.num_memory_tokens = num_memory_tokens
+        if num_memory_tokens > 0:
+            self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim))
+
+            # let funnel encoder know number of memory tokens, if specified
+            if hasattr(attn_layers, 'num_memory_tokens'):
+                attn_layers.num_memory_tokens = num_memory_tokens
+
+    def init_(self):
+        nn.init.normal_(self.token_emb.weight, std=0.02)
+
+    def forward(
+            self,
+            x,
+            return_embeddings=False,
+            mask=None,
+            return_mems=False,
+            return_attn=False,
+            mems=None,
+            **kwargs
+    ):
+        b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens
+        x = self.token_emb(x)
+        x += self.pos_emb(x)
+        x = self.emb_dropout(x)
+
+        x = self.project_emb(x)
+
+        if num_mem > 0:
+            mem = repeat(self.memory_tokens, 'n d -> b n d', b=b)
+            x = torch.cat((mem, x), dim=1)
+
+            # auto-handle masking after appending memory tokens
+            if exists(mask):
+                mask = F.pad(mask, (num_mem, 0), value=True)
+
+        x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
+        x = self.norm(x)
+
+        mem, x = x[:, :num_mem], x[:, num_mem:]
+
+        out = self.to_logits(x) if not return_embeddings else x
+
+        if return_mems:
+            hiddens = intermediates.hiddens
+            new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens
+            new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems))
+            return out, new_mems
+
+        if return_attn:
+            attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
+            return out, attn_maps
+
+        return out
+

prompts/freetraj/idx.txt

@@ -0,0 +1,8 @@
+1,2
+1,2
+1,2
+1,2
+1,2
+1,2
+1,2
+1,2,3

prompts/freetraj/text.txt

@@ -0,0 +1,8 @@
+A squirrel jumping from one tree to another.
+A bear climbing down a tree after spotting a threat. 
+A deer walking in a snowy field.
+A lion walking in the savanna grass.
+A kangaroo jumping in the Australian outback.
+A corgi running on the grassland on the grassland.
+A horse galloping on a street.
+A majestic eagle soaring high above the treetops, surveying its territory.

prompts/freetraj/traj_d.txt

@@ -0,0 +1,2 @@
+0, 0, 0.3, 0, 0.3
+15, 0.7, 1, 0.7, 1

prompts/freetraj/traj_l.txt

@@ -0,0 +1,3 @@
+0, 0.1, 0.4, 0.1, 0.4
+6, 0.6, 0.9, 0.1, 0.4
+15, 0.6, 0.9, 0.6, 0.9

requirements.txt

@@ -0,0 +1,23 @@
+decord==0.6.0
+einops==0.3.0
+imageio==2.9.0
+numpy==1.24.2
+omegaconf==2.1.1
+opencv_python
+pandas==2.0.0
+Pillow==9.5.0
+pytorch_lightning==1.8.3
+PyYAML==6.0
+setuptools==65.6.3
+torch==2.0.0
+torchvision
+tqdm==4.65.0
+transformers==4.25.1
+moviepy
+av
+xformers
+gradio
+timm
+scikit-learn 
+open_clip_torch==2.22.0
+kornia

scripts/evaluation/ddp_wrapper.py

@@ -0,0 +1,46 @@
+import datetime
+import argparse, importlib
+from pytorch_lightning import seed_everything
+
+import torch
+import torch.distributed as dist
+
+def setup_dist(local_rank):
+    if dist.is_initialized():
+        return
+    torch.cuda.set_device(local_rank)
+    torch.distributed.init_process_group('nccl', init_method='env://')
+
+
+def get_dist_info():
+    if dist.is_available():
+        initialized = dist.is_initialized()
+    else:
+        initialized = False
+    if initialized:
+        rank = dist.get_rank()
+        world_size = dist.get_world_size()
+    else:
+        rank = 0
+        world_size = 1
+    return rank, world_size
+
+
+if __name__ == '__main__':
+    now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--module", type=str, help="module name", default="inference")
+    parser.add_argument("--local_rank", type=int, nargs="?", help="for ddp", default=0)
+    args, unknown = parser.parse_known_args()
+    inference_api = importlib.import_module(args.module, package=None)
+
+    inference_parser = inference_api.get_parser()
+    inference_args, unknown = inference_parser.parse_known_args()
+
+    seed_everything(inference_args.seed)
+    setup_dist(args.local_rank)
+    torch.backends.cudnn.benchmark = True
+    rank, gpu_num = get_dist_info()
+
+    print("@CoLVDM Inference [rank%d]: %s"%(rank, now))
+    inference_api.run_inference(inference_args, gpu_num, rank)

scripts/evaluation/funcs.py

@@ -0,0 +1,372 @@
+import os, sys, glob
+import numpy as np
+from collections import OrderedDict
+from decord import VideoReader, cpu
+import cv2
+
+import torch
+import torchvision
+sys.path.insert(1, os.path.join(sys.path[0], '..', '..'))
+from lvdm.models.samplers.ddim import DDIMSampler
+from lvdm.models.samplers.ddim_freetraj import DDIMSampler as DDIMFreeTrajSampler
+
+from utils.utils_freetraj import get_freq_filter, freq_mix_3d, get_path, plan_path
+
+def batch_ddim_sampling_freetraj(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
+                        cfg_scale=1.0, temporal_cfg_scale=None, idx_list=[], input_traj=[], x_T_total=None, args=None, **kwargs):
+    ddim_sampler = DDIMFreeTrajSampler(model)
+    uncond_type = model.uncond_type
+    batch_size, channels, frames, h, w = noise_shape
+
+    ## construct unconditional guidance
+    if cfg_scale != 1.0:
+        if uncond_type == "empty_seq":
+            prompts = batch_size * [""]
+            #prompts = N * T * [""]  ## if is_imgbatch=True
+            uc_emb = model.get_learned_conditioning(prompts)
+        elif uncond_type == "zero_embed":
+            c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
+            uc_emb = torch.zeros_like(c_emb)
+                
+        ## process image embedding token
+        if hasattr(model, 'embedder'):
+            uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
+            ## img: b c h w >> b l c
+            uc_img = model.get_image_embeds(uc_img)
+            uc_emb = torch.cat([uc_emb, uc_img], dim=1)
+        
+        if isinstance(cond, dict):
+            uc = {key:cond[key] for key in cond.keys()}
+            uc.update({'c_crossattn': [uc_emb]})
+        else:
+            uc = uc_emb
+    else:
+        uc = None
+    
+    total_shape = [args.n_samples, 1, channels, frames, h, w]
+    print('total_shape', total_shape)
+
+    if x_T_total is None:
+        x_T_total = torch.randn(total_shape, device=model.device).repeat(1, batch_size, 1, 1, 1, 1)
+
+        noise_flow = True
+        if noise_flow:
+            print('noise_flow')
+            BOX_SIZE_H = input_traj[0][2] - input_traj[0][1]
+            BOX_SIZE_W = input_traj[0][4] - input_traj[0][3]
+            PATHS = plan_path(input_traj)
+            sub_h = int(BOX_SIZE_H * h) 
+            sub_w = int(BOX_SIZE_W * w)
+            x_T_sub = torch.randn([args.n_samples, 1, channels, sub_h, sub_w], device=model.device)
+            for i in range(frames):
+                h_start = int(PATHS[i][0] * h)
+                h_end = h_start + sub_h
+                w_start = int(PATHS[i][2] * w)
+                w_end = w_start + sub_w
+
+                # no mix
+                x_T_total[:, :, :, i, h_start:h_end, w_start:w_end] = x_T_sub
+
+            filter_shape = [
+                1, 
+                channels, 
+                frames, 
+                h, 
+                w
+            ]
+
+            freq_filter = get_freq_filter(
+                filter_shape, 
+                device = model.device, 
+                filter_type='butterworth',
+                n=4,
+                d_s=0.25,
+                d_t=0.1
+            )
+
+            x_T_rand = torch.randn([1, 1, channels, frames, h, w], device=model.device)
+            x_T_total = freq_mix_3d(x_T_total.to(dtype=torch.float32), x_T_rand, LPF=freq_filter)
+
+        
+    # x_T = None
+    batch_variants = []
+    #batch_variants1, batch_variants2 = [], []
+    for _ in range(n_samples):
+        x_T = x_T_total[_]
+        if ddim_sampler is not None:
+            kwargs.update({"clean_cond": True})
+            samples, _ = ddim_sampler.sample(S=ddim_steps,
+                                            conditioning=cond,
+                                            batch_size=noise_shape[0],
+                                            shape=noise_shape[1:],
+                                            verbose=False,
+                                            unconditional_guidance_scale=cfg_scale,
+                                            unconditional_conditioning=uc,
+                                            eta=ddim_eta,
+                                            temporal_length=noise_shape[2],
+                                            conditional_guidance_scale_temporal=temporal_cfg_scale,
+                                            x_T=x_T,
+                                            idx_list=idx_list,
+                                            input_traj=input_traj,
+                                            ddim_edit = args.ddim_edit,
+                                            **kwargs
+                                            )
+        ## reconstruct from latent to pixel space
+        batch_images = model.decode_first_stage_2DAE(samples)
+        batch_variants.append(batch_images)
+    ## batch, <samples>, c, t, h, w
+    batch_variants = torch.stack(batch_variants, dim=1)
+    return batch_variants
+
+def batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
+                        cfg_scale=1.0, temporal_cfg_scale=None, **kwargs):
+    ddim_sampler = DDIMSampler(model)
+    uncond_type = model.uncond_type
+    batch_size = noise_shape[0]
+
+    ## construct unconditional guidance
+    if cfg_scale != 1.0:
+        if uncond_type == "empty_seq":
+            prompts = batch_size * [""]
+            #prompts = N * T * [""]  ## if is_imgbatch=True
+            uc_emb = model.get_learned_conditioning(prompts)
+        elif uncond_type == "zero_embed":
+            c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
+            uc_emb = torch.zeros_like(c_emb)
+                
+        ## process image embedding token
+        if hasattr(model, 'embedder'):
+            uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
+            ## img: b c h w >> b l c
+            uc_img = model.get_image_embeds(uc_img)
+            uc_emb = torch.cat([uc_emb, uc_img], dim=1)
+        
+        if isinstance(cond, dict):
+            uc = {key:cond[key] for key in cond.keys()}
+            uc.update({'c_crossattn': [uc_emb]})
+        else:
+            uc = uc_emb
+    else:
+        uc = None
+    
+    x_T = None
+    batch_variants = []
+    #batch_variants1, batch_variants2 = [], []
+    for _ in range(n_samples):
+        if ddim_sampler is not None:
+            kwargs.update({"clean_cond": True})
+            samples, _ = ddim_sampler.sample(S=ddim_steps,
+                                            conditioning=cond,
+                                            batch_size=noise_shape[0],
+                                            shape=noise_shape[1:],
+                                            verbose=False,
+                                            unconditional_guidance_scale=cfg_scale,
+                                            unconditional_conditioning=uc,
+                                            eta=ddim_eta,
+                                            temporal_length=noise_shape[2],
+                                            conditional_guidance_scale_temporal=temporal_cfg_scale,
+                                            x_T=x_T,
+                                            **kwargs
+                                            )
+        ## reconstruct from latent to pixel space
+        batch_images = model.decode_first_stage_2DAE(samples)
+        batch_variants.append(batch_images)
+    ## batch, <samples>, c, t, h, w
+    batch_variants = torch.stack(batch_variants, dim=1)
+    return batch_variants
+
+
+def get_filelist(data_dir, ext='*'):
+    file_list = glob.glob(os.path.join(data_dir, '*.%s'%ext))
+    file_list.sort()
+    return file_list
+
+def get_dirlist(path):
+    list = []
+    if (os.path.exists(path)):
+        files = os.listdir(path)
+        for file in files:
+            m = os.path.join(path,file)
+            if (os.path.isdir(m)):
+                list.append(m)
+    list.sort()
+    return list
+
+
+def load_model_checkpoint(model, ckpt):
+    def load_checkpoint(model, ckpt, full_strict):
+        state_dict = torch.load(ckpt, map_location="cpu")
+        try:
+            ## deepspeed
+            new_pl_sd = OrderedDict()
+            for key in state_dict['module'].keys():
+                new_pl_sd[key[16:]]=state_dict['module'][key]
+            model.load_state_dict(new_pl_sd, strict=full_strict)
+        except:
+            if "state_dict" in list(state_dict.keys()):
+                state_dict = state_dict["state_dict"]
+            model.load_state_dict(state_dict, strict=full_strict)
+        return model
+    load_checkpoint(model, ckpt, full_strict=True)
+    print('>>> model checkpoint loaded.')
+    return model
+
+
+def load_prompts(prompt_file):
+    f = open(prompt_file, 'r')
+    prompt_list = []
+    for idx, line in enumerate(f.readlines()):
+        l = line.strip()
+        if len(l) != 0:
+            prompt_list.append(l)
+        f.close()
+    return prompt_list
+
+def load_idx(prompt_file):
+    f = open(prompt_file, 'r')
+    idx_list = []
+    for idx, line in enumerate(f.readlines()):
+        l = line.strip()
+        if len(l) != 0:
+            indices = l.split(',')
+            indices_list = []
+            for index in indices:
+                indices_list.append(int(index))
+            idx_list.append(indices_list)
+        f.close()
+    return idx_list
+
+def load_traj(prompt_file):
+    f = open(prompt_file, 'r')
+    traj_list = []
+    for idx, line in enumerate(f.readlines()):
+        l = line.strip()
+        if len(l) != 0:
+            numbers = l.split(',')
+            numbers_list = []
+            for number_index in range(len(numbers)):
+                if number_index == 0:
+                    numbers_list.append(int(numbers[number_index]))
+                else:
+                    numbers_list.append(float(numbers[number_index]))
+            traj_list.append(numbers_list)
+        f.close()
+    return traj_list
+
+def load_video_batch(filepath_list, frame_stride, video_size=(256,256), video_frames=16):
+    '''
+    Notice about some special cases:
+    1. video_frames=-1 means to take all the frames (with fs=1)
+    2. when the total video frames is less than required, padding strategy will be used (repreated last frame)
+    '''
+    fps_list = []
+    batch_tensor = []
+    assert frame_stride > 0, "valid frame stride should be a positive interge!"
+    for filepath in filepath_list:
+        padding_num = 0
+        vidreader = VideoReader(filepath, ctx=cpu(0), width=video_size[1], height=video_size[0])
+        fps = vidreader.get_avg_fps()
+        total_frames = len(vidreader)
+        max_valid_frames = (total_frames-1) // frame_stride + 1
+        if video_frames < 0:
+            ## all frames are collected: fs=1 is a must
+            required_frames = total_frames
+            frame_stride = 1
+        else:
+            required_frames = video_frames
+        query_frames = min(required_frames, max_valid_frames)
+        frame_indices = [frame_stride*i for i in range(query_frames)]
+
+        ## [t,h,w,c] -> [c,t,h,w]
+        frames = vidreader.get_batch(frame_indices)
+        frame_tensor = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float()
+        frame_tensor = (frame_tensor / 255. - 0.5) * 2
+        if max_valid_frames < required_frames:
+            padding_num = required_frames - max_valid_frames
+            frame_tensor = torch.cat([frame_tensor, *([frame_tensor[:,-1:,:,:]]*padding_num)], dim=1)
+            print(f'{os.path.split(filepath)[1]} is not long enough: {padding_num} frames padded.')
+        batch_tensor.append(frame_tensor)
+        sample_fps = int(fps/frame_stride)
+        fps_list.append(sample_fps)
+    
+    return torch.stack(batch_tensor, dim=0)
+
+from PIL import Image
+def load_image_batch(filepath_list, image_size=(256,256)):
+    batch_tensor = []
+    for filepath in filepath_list:
+        _, filename = os.path.split(filepath)
+        _, ext = os.path.splitext(filename)
+        if ext == '.mp4':
+            vidreader = VideoReader(filepath, ctx=cpu(0), width=image_size[1], height=image_size[0])
+            frame = vidreader.get_batch([0])
+            img_tensor = torch.tensor(frame.asnumpy()).squeeze(0).permute(2, 0, 1).float()
+        elif ext == '.png' or ext == '.jpg':
+            img = Image.open(filepath).convert("RGB")
+            rgb_img = np.array(img, np.float32)
+            #bgr_img = cv2.imread(filepath, cv2.IMREAD_COLOR)
+            #bgr_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2RGB)
+            rgb_img = cv2.resize(rgb_img, (image_size[1],image_size[0]), interpolation=cv2.INTER_LINEAR)
+            img_tensor = torch.from_numpy(rgb_img).permute(2, 0, 1).float()
+        else:
+            print(f'ERROR: <{ext}> image loading only support format: [mp4], [png], [jpg]')
+            raise NotImplementedError
+        img_tensor = (img_tensor / 255. - 0.5) * 2
+        batch_tensor.append(img_tensor)
+    return torch.stack(batch_tensor, dim=0)
+
+
+def save_videos(batch_tensors, savedir, filenames, fps=10):
+    # b,samples,c,t,h,w
+    n_samples = batch_tensors.shape[1]
+    for idx, vid_tensor in enumerate(batch_tensors):
+        video = vid_tensor.detach().cpu()
+        video = torch.clamp(video.float(), -1., 1.)
+        video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
+        frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n_samples)) for framesheet in video] #[3, 1*h, n*w]
+        grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w]
+        grid = (grid + 1.0) / 2.0
+        grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+        savepath = os.path.join(savedir, f"{filenames[idx]}.mp4")
+        torchvision.io.write_video(savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'})
+
+def save_videos_with_bbox(batch_tensors, savedir, conddir, filenames, fps=10, input_traj=[]):
+    # b,samples,c,t,h,w
+    BOX_SIZE_H = input_traj[0][2] - input_traj[0][1]
+    BOX_SIZE_W = input_traj[0][4] - input_traj[0][3]
+    PATHS = plan_path(input_traj)
+    n_samples = batch_tensors.shape[1]
+    for idx, vid_tensor in enumerate(batch_tensors):
+        video = vid_tensor.detach().cpu()
+        video = torch.clamp(video.float(), -1., 1.)
+        video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
+        h_len = video.shape[3]
+        w_len = video.shape[4]
+        sub_h = int(BOX_SIZE_H * h_len) 
+        sub_w = int(BOX_SIZE_W * w_len)
+        for i in range(video.shape[1]):
+            single_video = video[:, i]
+            frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n_samples)) for framesheet in single_video] #[3, 1*h, n*w]
+            grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w]
+            grid = (grid + 1.0) / 2.0
+            grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
+            savepath = os.path.join(savedir, f"{filenames[idx]}_{str(i)}.mp4")
+            torchvision.io.write_video(savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'})
+            for j in range(video.shape[0]): 
+                h_start = int(PATHS[j][0] * h_len)
+                h_end = h_start + sub_h
+                w_start = int(PATHS[j][2] * w_len)
+                w_end = w_start + sub_w
+
+                h_start = max(1, h_start)
+                h_end = min(h_len-1, h_end)
+                w_start = max(1, w_start)
+                w_end = min(w_len-1, w_end)
+
+                grid[j, h_start-1:h_end+1, w_start-1:w_start+2, :] = torch.ones_like(grid[j, h_start-1:h_end+1, w_start-1:w_start+2, :]) * torch.Tensor([127, 255, 127]).view(1, 1, 3)
+                grid[j, h_start-1:h_end+1, w_end-2:w_end+1, :] = torch.ones_like(grid[j, h_start-1:h_end+1, w_end-2:w_end+1, :]) * torch.Tensor([127, 255, 127]).view(1, 1, 3)
+                grid[j, h_start-1:h_start+2, w_start-1:w_end+1, :] = torch.ones_like(grid[j, h_start-1:h_start+2, w_start-1:w_end+1, :]) * torch.Tensor([127, 255, 127]).view(1, 1, 3)
+                grid[j, h_end-2:h_end+1, w_start-1:w_end+1, :] = torch.ones_like(grid[j, h_end-2:h_end+1, w_start-1:w_end+1, :]) * torch.Tensor([127, 255, 127]).view(1, 1, 3)
+            bbox_savepath = os.path.join(conddir, f"{filenames[idx]}_{str(i)}.mp4")
+            torchvision.io.write_video(bbox_savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'})
+

scripts/evaluation/inference.py

@@ -0,0 +1,137 @@
+import argparse, os, sys, glob, yaml, math, random
+import datetime, time
+import numpy as np
+from omegaconf import OmegaConf
+from collections import OrderedDict
+from tqdm import trange, tqdm
+from einops import repeat
+from einops import rearrange, repeat
+from functools import partial
+import torch
+from pytorch_lightning import seed_everything
+
+from funcs import load_model_checkpoint, load_prompts, load_image_batch, get_filelist, save_videos
+from funcs import batch_ddim_sampling
+from utils.utils import instantiate_from_config
+
+
+def get_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--seed", type=int, default=20230211, help="seed for seed_everything")
+    parser.add_argument("--mode", default="base", type=str, help="which kind of inference mode: {'base', 'i2v'}")
+    parser.add_argument("--ckpt_path", type=str, default=None, help="checkpoint path")
+    parser.add_argument("--config", type=str, help="config (yaml) path")
+    parser.add_argument("--prompt_file", type=str, default=None, help="a text file containing many prompts")
+    parser.add_argument("--savedir", type=str, default=None, help="results saving path")
+    parser.add_argument("--savefps", type=str, default=10, help="video fps to generate")
+    parser.add_argument("--n_samples", type=int, default=1, help="num of samples per prompt",)
+    parser.add_argument("--ddim_steps", type=int, default=50, help="steps of ddim if positive, otherwise use DDPM",)
+    parser.add_argument("--ddim_eta", type=float, default=1.0, help="eta for ddim sampling (0.0 yields deterministic sampling)",)
+    parser.add_argument("--bs", type=int, default=1, help="batch size for inference")
+    parser.add_argument("--height", type=int, default=512, help="image height, in pixel space")
+    parser.add_argument("--width", type=int, default=512, help="image width, in pixel space")
+    parser.add_argument("--frames", type=int, default=-1, help="frames num to inference")
+    parser.add_argument("--fps", type=int, default=24)
+    parser.add_argument("--unconditional_guidance_scale", type=float, default=1.0, help="prompt classifier-free guidance")
+    parser.add_argument("--unconditional_guidance_scale_temporal", type=float, default=None, help="temporal consistency guidance")
+    ## for conditional i2v only
+    parser.add_argument("--cond_input", type=str, default=None, help="data dir of conditional input")
+    return parser
+
+
+def run_inference(args, gpu_num, gpu_no, **kwargs):
+    ## step 1: model config
+    ## -----------------------------------------------------------------
+    config = OmegaConf.load(args.config)
+    #data_config = config.pop("data", OmegaConf.create())
+    model_config = config.pop("model", OmegaConf.create())
+    model = instantiate_from_config(model_config)
+    model = model.cuda(gpu_no)
+    assert os.path.exists(args.ckpt_path), f"Error: checkpoint [{args.ckpt_path}] Not Found!"
+    model = load_model_checkpoint(model, args.ckpt_path)
+    model.eval()
+
+    ## sample shape
+    assert (args.height % 16 == 0) and (args.width % 16 == 0), "Error: image size [h,w] should be multiples of 16!"
+    ## latent noise shape
+    h, w = args.height // 8, args.width // 8
+    frames = model.temporal_length if args.frames < 0 else args.frames
+    channels = model.channels
+    
+    ## saving folders
+    os.makedirs(args.savedir, exist_ok=True)
+
+    ## step 2: load data
+    ## -----------------------------------------------------------------
+    assert os.path.exists(args.prompt_file), "Error: prompt file NOT Found!"
+    prompt_list = load_prompts(args.prompt_file)
+    num_samples = len(prompt_list)
+    filename_list = [f"{id+1:04d}" for id in range(num_samples)]
+
+    samples_split = num_samples // gpu_num
+    residual_tail = num_samples % gpu_num
+    print(f'[rank:{gpu_no}] {samples_split}/{num_samples} samples loaded.')
+    indices = list(range(samples_split*gpu_no, samples_split*(gpu_no+1)))
+    if gpu_no == 0 and residual_tail != 0:
+        indices = indices + list(range(num_samples-residual_tail, num_samples))
+    prompt_list_rank = [prompt_list[i] for i in indices]
+
+    ## conditional input
+    if args.mode == "i2v":
+        ## each video or frames dir per prompt
+        cond_inputs = get_filelist(args.cond_input, ext='[mpj][pn][4gj]')   # '[mpj][pn][4gj]'
+        assert len(cond_inputs) == num_samples, f"Error: conditional input ({len(cond_inputs)}) NOT match prompt ({num_samples})!"
+        filename_list = [f"{os.path.split(cond_inputs[id])[-1][:-4]}" for id in range(num_samples)]
+        cond_inputs_rank = [cond_inputs[i] for i in indices]
+
+    filename_list_rank = [filename_list[i] for i in indices]
+
+    ## step 3: run over samples
+    ## -----------------------------------------------------------------
+    start = time.time()
+    n_rounds = len(prompt_list_rank) // args.bs
+    n_rounds = n_rounds+1 if len(prompt_list_rank) % args.bs != 0 else n_rounds
+    for idx in range(0, n_rounds):
+        print(f'[rank:{gpu_no}] batch-{idx+1} ({args.bs})x{args.n_samples} ...')
+        idx_s = idx*args.bs
+        idx_e = min(idx_s+args.bs, len(prompt_list_rank))
+        batch_size = idx_e - idx_s
+        filenames = filename_list_rank[idx_s:idx_e]
+        noise_shape = [batch_size, channels, frames, h, w]
+        fps = torch.tensor([args.fps]*batch_size).to(model.device).long()
+
+        prompts = prompt_list_rank[idx_s:idx_e]
+        if isinstance(prompts, str):
+            prompts = [prompts]
+        #prompts = batch_size * [""]
+        text_emb = model.get_learned_conditioning(prompts)
+
+        if args.mode == 'base':
+            cond = {"c_crossattn": [text_emb], "fps": fps}
+        elif args.mode == 'i2v':
+            #cond_images = torch.zeros(noise_shape[0],3,224,224).to(model.device)
+            cond_images = load_image_batch(cond_inputs_rank[idx_s:idx_e], (args.height, args.width))
+            cond_images = cond_images.to(model.device)
+            img_emb = model.get_image_embeds(cond_images)
+            imtext_cond = torch.cat([text_emb, img_emb], dim=1)
+            cond = {"c_crossattn": [imtext_cond], "fps": fps}
+        else:
+            raise NotImplementedError
+
+        ## inference
+        batch_samples = batch_ddim_sampling(model, cond, noise_shape, args.n_samples, \
+                                                args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, **kwargs)
+        ## b,samples,c,t,h,w
+        save_videos(batch_samples, args.savedir, filenames, fps=args.savefps)
+
+    print(f"Saved in {args.savedir}. Time used: {(time.time() - start):.2f} seconds")
+
+
+if __name__ == '__main__':
+    now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
+    print("@CoLVDM Inference: %s"%now)
+    parser = get_parser()
+    args = parser.parse_args()
+    seed_everything(args.seed)
+    rank, gpu_num = 0, 1
+    run_inference(args, gpu_num, rank)

scripts/evaluation/inference_freetraj.py

@@ -0,0 +1,157 @@
+import argparse, os, sys, glob, yaml, math, random
+import datetime, time
+import numpy as np
+from omegaconf import OmegaConf
+from collections import OrderedDict
+from tqdm import trange, tqdm
+from einops import repeat
+from einops import rearrange, repeat
+from functools import partial
+import torch
+from pytorch_lightning import seed_everything
+
+from funcs import load_model_checkpoint, load_prompts, load_idx, load_traj, load_image_batch, get_filelist, save_videos, save_videos_with_bbox
+from funcs import batch_ddim_sampling_freetraj
+from utils.utils import instantiate_from_config
+
+
+def get_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--seed", type=int, default=20230211, help="seed for seed_everything")
+    parser.add_argument("--mode", default="base", type=str, help="which kind of inference mode: {'base', 'i2v'}")
+    parser.add_argument("--ckpt_path", type=str, default=None, help="checkpoint path")
+    parser.add_argument("--config", type=str, help="config (yaml) path")
+    parser.add_argument("--prompt_file", type=str, default=None, help="a text file containing many prompts")
+    parser.add_argument("--savedir", type=str, default=None, help="results saving path")
+    parser.add_argument("--savefps", type=str, default=10, help="video fps to generate")
+    parser.add_argument("--n_samples", type=int, default=1, help="num of samples per prompt")
+    parser.add_argument("--ddim_steps", type=int, default=50, help="steps of ddim if positive, otherwise use DDPM")
+    parser.add_argument("--ddim_eta", type=float, default=1.0, help="eta for ddim sampling (0.0 yields deterministic sampling)")
+    parser.add_argument("--bs", type=int, default=1, help="batch size for inference")
+    parser.add_argument("--height", type=int, default=512, help="image height, in pixel space")
+    parser.add_argument("--width", type=int, default=512, help="image width, in pixel space")
+    parser.add_argument("--frames", type=int, default=-1, help="frames num to inference")
+    parser.add_argument("--fps", type=int, default=24)
+    parser.add_argument("--unconditional_guidance_scale", type=float, default=1.0, help="prompt classifier-free guidance")
+    parser.add_argument("--unconditional_guidance_scale_temporal", type=float, default=None, help="temporal consistency guidance")
+    ## for conditional i2v only
+    parser.add_argument("--cond_input", type=str, default=None, help="data dir of conditional input")
+
+    # FreeTraj
+    parser.add_argument("--ddim_edit", type=int, default=6, help="steps of ddim for edited attention")
+    parser.add_argument("--idx_file", type=str, default=None, help="a index file containing many prompts")
+    parser.add_argument("--traj_file", type=str, default=None, help="a path file containing many prompts")
+    return parser
+
+
+def run_inference(args, gpu_num, gpu_no, **kwargs):
+    ## step 1: model config
+    ## -----------------------------------------------------------------
+    config = OmegaConf.load(args.config)
+    #data_config = config.pop("data", OmegaConf.create())
+    model_config = config.pop("model", OmegaConf.create())
+    model = instantiate_from_config(model_config)
+    model = model.cuda(gpu_no)
+    assert os.path.exists(args.ckpt_path), f"Error: checkpoint [{args.ckpt_path}] Not Found!"
+    model = load_model_checkpoint(model, args.ckpt_path)
+    model.eval()
+
+    ## sample shape
+    assert (args.height % 16 == 0) and (args.width % 16 == 0), "Error: image size [h,w] should be multiples of 16!"
+    ## latent noise shape
+    h, w = args.height // 8, args.width // 8
+    frames = model.temporal_length if args.frames < 0 else args.frames
+    channels = model.channels
+    
+    ## saving folders
+    os.makedirs(args.savedir, exist_ok=True)
+    bboxdir = os.path.join(args.savedir, "bbox")
+    os.makedirs(bboxdir, exist_ok=True)
+
+    ## step 2: load data
+    ## -----------------------------------------------------------------
+    assert os.path.exists(args.prompt_file), "Error: prompt file NOT Found!"
+    prompt_list = load_prompts(args.prompt_file)
+    idx_list_rank = load_idx(args.idx_file)
+    input_traj = load_traj(args.traj_file)
+    print(prompt_list)
+    print(idx_list_rank)
+    print(input_traj)
+
+    num_samples = len(prompt_list)
+    filename_list = [f"{id+1:04d}" for id in range(num_samples)]
+
+    samples_split = num_samples // gpu_num
+    residual_tail = num_samples % gpu_num
+    print(f'[rank:{gpu_no}] {samples_split}/{num_samples} samples loaded.')
+    indices = list(range(samples_split*gpu_no, samples_split*(gpu_no+1)))
+    if gpu_no == 0 and residual_tail != 0:
+        indices = indices + list(range(num_samples-residual_tail, num_samples))
+    prompt_list_rank = [prompt_list[i] for i in indices]
+
+    ## conditional input
+    if args.mode == "i2v":
+        ## each video or frames dir per prompt
+        cond_inputs = get_filelist(args.cond_input, ext='[mpj][pn][4gj]')   # '[mpj][pn][4gj]'
+        assert len(cond_inputs) == num_samples, f"Error: conditional input ({len(cond_inputs)}) NOT match prompt ({num_samples})!"
+        filename_list = [f"{os.path.split(cond_inputs[id])[-1][:-4]}" for id in range(num_samples)]
+        cond_inputs_rank = [cond_inputs[i] for i in indices]
+
+    filename_list_rank = [filename_list[i] for i in indices]
+    
+    assert len(idx_list_rank) == len(filename_list_rank), "Error: metas are not paired!"
+
+
+    ## step 3: run over samples
+    ## -----------------------------------------------------------------
+    start = time.time()
+    n_rounds = len(prompt_list_rank) // args.bs
+    n_rounds = n_rounds+1 if len(prompt_list_rank) % args.bs != 0 else n_rounds
+    for idx in range(0, n_rounds):
+        print(f'[rank:{gpu_no}] batch-{idx+1} ({args.bs})x{args.n_samples} ...', flush=True)
+        idx_s = idx*args.bs
+        idx_e = min(idx_s+args.bs, len(prompt_list_rank))
+        batch_size = idx_e - idx_s
+        filenames = filename_list_rank[idx_s:idx_e]
+        noise_shape = [batch_size, channels, frames, h, w]
+        fps = torch.tensor([args.fps]*batch_size).to(model.device).long()
+
+        idx_list = idx_list_rank[idx_s:idx_e][0]
+        # print(idx_list)
+
+        prompts = prompt_list_rank[idx_s:idx_e]
+        if isinstance(prompts, str):
+            prompts = [prompts]
+        #prompts = batch_size * [""]
+        text_emb = model.get_learned_conditioning(prompts)
+
+        if args.mode == 'base':
+            cond = {"c_crossattn": [text_emb], "fps": fps}
+        elif args.mode == 'i2v':
+            #cond_images = torch.zeros(noise_shape[0],3,224,224).to(model.device)
+            cond_images = load_image_batch(cond_inputs_rank[idx_s:idx_e], (args.height, args.width))
+            cond_images = cond_images.to(model.device)
+            img_emb = model.get_image_embeds(cond_images)
+            imtext_cond = torch.cat([text_emb, img_emb], dim=1)
+            cond = {"c_crossattn": [imtext_cond], "fps": fps}
+        else:
+            raise NotImplementedError
+
+        ## inference
+        batch_samples = batch_ddim_sampling_freetraj(model, cond, noise_shape, args.n_samples, \
+                                                args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, idx_list=idx_list, input_traj=input_traj, args=args, **kwargs)
+        ## b,samples,c,t,h,w
+        # save_videos(batch_samples, args.savedir, filenames, fps=args.savefps)
+        save_videos_with_bbox(batch_samples, args.savedir, bboxdir, filenames, fps=args.savefps, input_traj=input_traj)
+
+    print(f"Saved in {args.savedir}. Time used: {(time.time() - start):.2f} seconds")
+
+
+if __name__ == '__main__':
+    now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
+    print("@CoLVDM Inference: %s"%now)
+    parser = get_parser()
+    args = parser.parse_args()
+    seed_everything(args.seed)
+    rank, gpu_num = 0, 1
+    run_inference(args, gpu_num, rank)

scripts/run_text2video.sh

@@ -0,0 +1,21 @@
+name="base_512_v2"
+
+ckpt='checkpoints/base_512_v2/model.ckpt'
+config='configs/inference_t2v_512_v2.0.yaml'
+
+prompt_file="prompts/freetraj/text.txt"
+res_dir="results"
+
+python3 scripts/evaluation/inference.py \
+--seed 123 \
+--mode 'base' \
+--ckpt_path $ckpt \
+--config $config \
+--savedir $res_dir/$name \
+--n_samples 1 \
+--bs 1 --height 320 --width 512 \
+--unconditional_guidance_scale 12.0 \
+--ddim_steps 50 \
+--ddim_eta 0.0 \
+--prompt_file $prompt_file \
+--fps 16

scripts/run_text2video_freetraj.sh

@@ -0,0 +1,27 @@
+name="base_512_v2_ddim6_l"
+
+ckpt='checkpoints/base_512_v2/model.ckpt'
+config='configs/inference_t2v_freetraj_512_v2.0.yaml'
+
+res_dir="results_freetraj"
+prompt_file="prompts/freetraj/text.txt"
+idx_file="prompts/freetraj/idx.txt"
+traj_file="prompts/freetraj/traj_l.txt"
+
+
+python3 scripts/evaluation/inference_freetraj.py \
+--seed 123 \
+--mode 'base' \
+--ckpt_path $ckpt \
+--config $config \
+--savedir $res_dir/$name \
+--n_samples 1 \
+--bs 1 --height 320 --width 512 \
+--unconditional_guidance_scale 12.0 \
+--ddim_steps 50 \
+--ddim_eta 0.0 \
+--prompt_file $prompt_file \
+--idx_file $idx_file \
+--traj_file $traj_file \
+--ddim_edit 6 \
+--fps 16

utils/utils.py

@@ -0,0 +1,77 @@
+import importlib
+import numpy as np
+import cv2
+import torch
+import torch.distributed as dist
+
+
+def count_params(model, verbose=False):
+    total_params = sum(p.numel() for p in model.parameters())
+    if verbose:
+        print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
+    return total_params
+
+
+def check_istarget(name, para_list):
+    """ 
+    name: full name of source para
+    para_list: partial name of target para 
+    """
+    istarget=False
+    for para in para_list:
+        if para in name:
+            return True
+    return istarget
+
+
+def instantiate_from_config(config):
+    if not "target" in config:
+        if config == '__is_first_stage__':
+            return None
+        elif config == "__is_unconditional__":
+            return None
+        raise KeyError("Expected key `target` to instantiate.")
+    return get_obj_from_str(config["target"])(**config.get("params", dict()))
+
+
+def get_obj_from_str(string, reload=False):
+    module, cls = string.rsplit(".", 1)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)
+
+
+def load_npz_from_dir(data_dir):
+    data = [np.load(os.path.join(data_dir, data_name))['arr_0'] for data_name in os.listdir(data_dir)]
+    data = np.concatenate(data, axis=0)
+    return data
+
+
+def load_npz_from_paths(data_paths):
+    data = [np.load(data_path)['arr_0'] for data_path in data_paths]
+    data = np.concatenate(data, axis=0)
+    return data   
+
+
+def resize_numpy_image(image, max_resolution=512 * 512, resize_short_edge=None):
+    h, w = image.shape[:2]
+    if resize_short_edge is not None:
+        k = resize_short_edge / min(h, w)
+    else:
+        k = max_resolution / (h * w)
+        k = k**0.5
+    h = int(np.round(h * k / 64)) * 64
+    w = int(np.round(w * k / 64)) * 64
+    image = cv2.resize(image, (w, h), interpolation=cv2.INTER_LANCZOS4)
+    return image
+
+
+def setup_dist(args):
+    if dist.is_initialized():
+        return
+    torch.cuda.set_device(args.local_rank)
+    torch.distributed.init_process_group(
+        'nccl',
+        init_method='env://'
+    )

utils/utils_freetraj.py

@@ -0,0 +1,291 @@
+import torch
+import torch.fft as fft
+import math
+
+
+def get_longpath(BOX_SIZE_H=0.3, BOX_SIZE_W=0.3, input_mode=4):
+
+    if input_mode == 1:
+        # mode 1
+        inputs = [[0, 0, 0 + BOX_SIZE_H, 0, 0 + BOX_SIZE_W], 
+                [7, 1-BOX_SIZE_H, 1, (1-BOX_SIZE_W) / 15 * 7, (1-BOX_SIZE_W) / 15 * 7 + BOX_SIZE_W], 
+                [8, 1-BOX_SIZE_H, 1, (1-BOX_SIZE_W) / 15 * 8, (1-BOX_SIZE_W) / 15 * 8 + BOX_SIZE_W], 
+                [15, 0, 0 + BOX_SIZE_H, 1-BOX_SIZE_W, 1],
+                [16, 0.1, 0.1 + BOX_SIZE_H, 0.9-BOX_SIZE_W, 0.9],
+                [25, 0.1, 0.1 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W],
+                [31, 0.9-BOX_SIZE_H, 0.9, 0.1, 0.1 + BOX_SIZE_W],
+                [32, 1-BOX_SIZE_H, 1, 0, 0 + BOX_SIZE_W],
+                [39, 0, 0 + BOX_SIZE_H, (1-BOX_SIZE_W) / 15 * 7, (1-BOX_SIZE_W) / 15 * 7 + BOX_SIZE_W],
+                [40, 0, 0 + BOX_SIZE_H, (1-BOX_SIZE_W) / 15 * 8, (1-BOX_SIZE_W) / 15 * 8 + BOX_SIZE_W],
+                [47, 1-BOX_SIZE_H, 1, 1-BOX_SIZE_W, 1],
+                [48, 0.9-BOX_SIZE_H, 0.9, 0.9-BOX_SIZE_W, 0.9],
+                [57, 0.9-BOX_SIZE_H, 0.9, 0.1, 0.1 + BOX_SIZE_W],
+                [63, 0.1, 0.1 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W]]
+    elif input_mode == 2:
+        # mode 2
+        inputs = [[0, 0.1, 0.1 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W], 
+                  [6, 0.9-BOX_SIZE_H, 0.9, 0.1, 0.1 + BOX_SIZE_W], 
+                  [15, 0.9-BOX_SIZE_H, 0.9, 0.9-BOX_SIZE_W, 0.9],
+                  [16, 0.9-BOX_SIZE_H, 0.9, 0.9-BOX_SIZE_W, 0.9], 
+                  [22, 0.1, 0.1 + BOX_SIZE_H, 0.9-BOX_SIZE_W, 0.9], 
+                  [31, 0.1, 0.1 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W],
+                  [32, 0.1, 0.1 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W],
+                  [41, 0.1, 0.1 + BOX_SIZE_H, 0.9-BOX_SIZE_W, 0.9],
+                  [47, 0.9-BOX_SIZE_H, 0.9, 0.9-BOX_SIZE_W, 0.9],
+                  [48, 0.9-BOX_SIZE_H, 0.9, 0.9-BOX_SIZE_W, 0.9],
+                  [57, 0.9-BOX_SIZE_H, 0.9, 0.1, 0.1 + BOX_SIZE_W],
+                  [63, 0.1, 0.1 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W]]
+    elif input_mode == 3:
+        # mode 3 ||||
+        inputs = [[0, 0, 0 + BOX_SIZE_H, 0, 0 + BOX_SIZE_W],
+            [9, 1-BOX_SIZE_H, 1, (1-BOX_SIZE_W) / 7 * 1, (1-BOX_SIZE_W) / 7 * 1 + BOX_SIZE_W],
+            [18, 0, 0 + BOX_SIZE_H, (1-BOX_SIZE_W) / 7 * 2, (1-BOX_SIZE_W) / 7 * 2 + BOX_SIZE_W],
+            [27, 1-BOX_SIZE_H, 1, (1-BOX_SIZE_W) / 7 * 3, (1-BOX_SIZE_W) / 7 * 3 + BOX_SIZE_W],
+            [36, 0, 0 + BOX_SIZE_H, (1-BOX_SIZE_W) / 7 * 4, (1-BOX_SIZE_W) / 7 * 4 + BOX_SIZE_W],
+            [45, 1-BOX_SIZE_H, 1, (1-BOX_SIZE_W) / 7 * 5, (1-BOX_SIZE_W) / 7 * 5 + BOX_SIZE_W],
+            [54, 0, 0 + BOX_SIZE_H, (1-BOX_SIZE_W) / 7 * 6, (1-BOX_SIZE_W) / 7 * 6 + BOX_SIZE_W],
+            [63, 1-BOX_SIZE_H, 1, 1-BOX_SIZE_W, 1]]
+    elif input_mode == 4:
+        # mode 4 ----
+        inputs = [[0, 0, 0 + BOX_SIZE_H, 0, 0 + BOX_SIZE_W],
+            [9, (1-BOX_SIZE_H) / 7 * 1, (1-BOX_SIZE_H) / 7 * 1 + BOX_SIZE_H, 1-BOX_SIZE_W, 1],
+            [18, (1-BOX_SIZE_H) / 7 * 2, (1-BOX_SIZE_H) / 7 * 2 + BOX_SIZE_H, 0, 0 + BOX_SIZE_W],
+            [27, (1-BOX_SIZE_H) / 7 * 3, (1-BOX_SIZE_H) / 7 * 3 + BOX_SIZE_H, 1-BOX_SIZE_W, 1],
+            [36, (1-BOX_SIZE_H) / 7 * 4, (1-BOX_SIZE_H) / 7 * 4 + BOX_SIZE_H, 0, 0 + BOX_SIZE_W],
+            [45, (1-BOX_SIZE_H) / 7 * 5, (1-BOX_SIZE_H) / 7 * 5 + BOX_SIZE_H, 1-BOX_SIZE_W, 1],
+            [54, (1-BOX_SIZE_H) / 7 * 6, (1-BOX_SIZE_H) / 7 * 6 + BOX_SIZE_H, 0, 0 + BOX_SIZE_W],
+            [63, 1-BOX_SIZE_H, 1, 1-BOX_SIZE_W, 1]]
+    else:
+        print('error')
+        exit()
+
+    outputs = plan_path(inputs)
+    # print(outputs)
+    return outputs
+
+def get_path(BOX_SIZE_H=0.3, BOX_SIZE_W=0.3, input_mode=0):
+
+    if input_mode == 0:
+        # \ d
+        inputs = [[0, 0, 0 + BOX_SIZE_H, 0, 0 + BOX_SIZE_W], [15, 1-BOX_SIZE_H, 1, 1-BOX_SIZE_W, 1]] 
+    elif input_mode == 1:
+        # / re d
+        inputs = [[0, 0, 0 + BOX_SIZE_H, 1-BOX_SIZE_W, 1], [15, 1-BOX_SIZE_H, 1, 0, 0 + BOX_SIZE_W]] 
+    elif input_mode == 2:        
+        # L
+        inputs = [[0, 0.1, 0.1 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W], [6, 0.9-BOX_SIZE_H, 0.9, 0.1, 0.1 + BOX_SIZE_W], [15, 0.9-BOX_SIZE_H, 0.9, 0.9-BOX_SIZE_W, 0.9]] 
+    elif input_mode == 3:     
+        # re L
+        inputs = [[0, 0.9-BOX_SIZE_H, 0.9, 0.9-BOX_SIZE_W, 0.9], [6, 0.1, 0.1 + BOX_SIZE_H, 0.9-BOX_SIZE_W, 0.9], [15, 0.1, 0.1 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W]] 
+    elif input_mode == 4:     
+        # V
+        inputs = [[0, 0, 0 + BOX_SIZE_H, 0, 0 + BOX_SIZE_W], [7, 1-BOX_SIZE_H, 1, (1-BOX_SIZE_W) / 15 * 7, (1-BOX_SIZE_W) / 15 * 7 + BOX_SIZE_W], [8, 1-BOX_SIZE_H, 1, (1-BOX_SIZE_W) / 15 * 8, (1-BOX_SIZE_W) / 15 * 8 + BOX_SIZE_W], [15, 0, 0 + BOX_SIZE_H, 1-BOX_SIZE_W, 1]]
+    elif input_mode == 5:     
+        # re V
+        inputs = [[0, 1-BOX_SIZE_H, 1, 1-BOX_SIZE_W, 1], [7, 0, 0 + BOX_SIZE_H, (1-BOX_SIZE_W) / 15 * 8, (1-BOX_SIZE_W) / 15 * 8 + BOX_SIZE_W], [8, 0, 0 + BOX_SIZE_H, (1-BOX_SIZE_W) / 15 * 7, (1-BOX_SIZE_W) / 15 * 7 + BOX_SIZE_W], [15, 1-BOX_SIZE_H, 1, 0, 0 + BOX_SIZE_W]]
+    elif input_mode == 6:    
+        # -- goback
+        inputs = [[0, 0.35, 0.35 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W], [7, 0.35, 0.35 + BOX_SIZE_H, 0.9-BOX_SIZE_W, 0.9], [8, 0.35, 0.35 + BOX_SIZE_H, 0.9-BOX_SIZE_W, 0.9], [15, 0.35, 0.35 + BOX_SIZE_H, 0.1, 0.1 + BOX_SIZE_W]] 
+    elif input_mode == 7:    
+        # tri
+        inputs = [[0, 0.1, 0.1 + BOX_SIZE_H, 0.35, 0.35 + BOX_SIZE_W], [5, 0.9-BOX_SIZE_H, 0.9, 0.9-BOX_SIZE_W, 0.9], [10, 0.9-BOX_SIZE_H, 0.9, 0.1, 0.1 + BOX_SIZE_W], [15, 0.1, 0.1 + BOX_SIZE_H, 0.35, 0.35 + BOX_SIZE_W]]
+
+    outputs = plan_path(inputs)
+    return outputs
+
+# input: List([frame, h_start, h_end, w_start, w_end], ...)
+# return: List([h_start, h_end, w_start, w_end], ...)
+def plan_path(input, video_length = 16):
+    len_input = len(input)
+    path = [input[0][1:]]
+    for i in range(1, len_input):
+        start = input[i-1]
+        end = input[i]
+        start_frame = start[0]
+        end_frame = end[0]
+        h_start_change = (end[1] - start[1]) / (end_frame - start_frame)
+        h_end_change = (end[2] - start[2]) / (end_frame - start_frame)
+        w_start_change = (end[3] - start[3]) / (end_frame - start_frame)
+        w_end_change = (end[4] - start[4]) / (end_frame - start_frame)
+        for j in range(start_frame+1, end_frame + 1):
+            increase_frame = j - start_frame
+            path += [[increase_frame * h_start_change + start[1], increase_frame * h_end_change + start[2], increase_frame * w_start_change + start[3], increase_frame * w_end_change + start[4]]]
+ 
+    if input[0][0] > 0:
+        h_change = path[1][0] - path[0][0]
+        w_change = path[1][2] - path[0][2]
+        for i in range(input[0][0]):
+            path = [path[0][0] - h_change, path[0][1] - h_change, path[0][2] - w_change, path[0][3] - w_change] + path
+
+    if input[-1][0] < video_length - 1:
+        h_change = path[-1][0] - path[-2][0]
+        w_change = path[-1][2] - path[-2][2]
+        for i in range(video_length - 1 - input[-1][0]):
+            path = path + [path[-1][0] + h_change, path[-1][1] + h_change, path[-1][2] + w_change, path[-1][3] + w_change]
+
+    return path
+
+
+def gaussian_2d(x=0, y=0, mx=0, my=0, sx=1, sy=1):
+    """ 2d Gaussian weight function
+    """
+    gaussian_map = (
+        1
+        / (2 * math.pi * sx * sy)
+        * torch.exp(-((x - mx) ** 2 / (2 * sx**2) + (y - my) ** 2 / (2 * sy**2)))
+    )
+    gaussian_map.div_(gaussian_map.max())
+    return gaussian_map
+
+def gaussian_weight(height=32, width=32, KERNEL_DIVISION=3.0):
+
+    x = torch.linspace(0, height, height)
+    y = torch.linspace(0, width, width)
+    x, y = torch.meshgrid(x, y, indexing="ij")
+    noise_patch = (
+                    gaussian_2d(
+                        x,
+                        y,
+                        mx=int(height / 2),
+                        my=int(width / 2),
+                        sx=float(height / KERNEL_DIVISION),
+                        sy=float(width / KERNEL_DIVISION),
+                    )
+                ).half()
+    return noise_patch
+
+def freq_mix_3d(x, noise, LPF):
+    """
+    Noise reinitialization.
+
+    Args:
+        x: diffused latent
+        noise: randomly sampled noise
+        LPF: low pass filter
+    """
+    # FFT
+    x_freq = fft.fftn(x, dim=(-3, -2, -1))
+    x_freq = fft.fftshift(x_freq, dim=(-3, -2, -1))
+    noise_freq = fft.fftn(noise, dim=(-3, -2, -1))
+    noise_freq = fft.fftshift(noise_freq, dim=(-3, -2, -1))
+
+    # frequency mix
+    HPF = 1 - LPF
+    x_freq_low = x_freq * LPF
+    noise_freq_high = noise_freq * HPF
+    x_freq_mixed = x_freq_low + noise_freq_high # mix in freq domain
+
+    # IFFT
+    x_freq_mixed = fft.ifftshift(x_freq_mixed, dim=(-3, -2, -1))
+    x_mixed = fft.ifftn(x_freq_mixed, dim=(-3, -2, -1)).real
+
+    return x_mixed
+
+
+def get_freq_filter(shape, device, filter_type, n, d_s, d_t):
+    """
+    Form the frequency filter for noise reinitialization.
+
+    Args:
+        shape: shape of latent (B, C, T, H, W)
+        filter_type: type of the freq filter
+        n: (only for butterworth) order of the filter, larger n ~ ideal, smaller n ~ gaussian
+        d_s: normalized stop frequency for spatial dimensions (0.0-1.0)
+        d_t: normalized stop frequency for temporal dimension (0.0-1.0)
+    """
+    if filter_type == "gaussian":
+        return gaussian_low_pass_filter(shape=shape, d_s=d_s, d_t=d_t).to(device)
+    elif filter_type == "ideal":
+        return ideal_low_pass_filter(shape=shape, d_s=d_s, d_t=d_t).to(device)
+    elif filter_type == "box":
+        return box_low_pass_filter(shape=shape, d_s=d_s, d_t=d_t).to(device)
+    elif filter_type == "butterworth":
+        return butterworth_low_pass_filter(shape=shape, n=n, d_s=d_s, d_t=d_t).to(device)
+    else:
+        raise NotImplementedError
+
+def gaussian_low_pass_filter(shape, d_s=0.25, d_t=0.25):
+    """
+    Compute the gaussian low pass filter mask.
+
+    Args:
+        shape: shape of the filter (volume)
+        d_s: normalized stop frequency for spatial dimensions (0.0-1.0)
+        d_t: normalized stop frequency for temporal dimension (0.0-1.0)
+    """
+    T, H, W = shape[-3], shape[-2], shape[-1]
+    mask = torch.zeros(shape)
+    if d_s==0 or d_t==0:
+        return mask
+    for t in range(T):
+        for h in range(H):
+            for w in range(W):
+                d_square = (((d_s/d_t)*(2*t/T-1))**2 + (2*h/H-1)**2 + (2*w/W-1)**2)
+                mask[..., t,h,w] = math.exp(-1/(2*d_s**2) * d_square)
+    return mask
+
+
+def butterworth_low_pass_filter(shape, n=4, d_s=0.25, d_t=0.25):
+    """
+    Compute the butterworth low pass filter mask.
+
+    Args:
+        shape: shape of the filter (volume)
+        n: order of the filter, larger n ~ ideal, smaller n ~ gaussian
+        d_s: normalized stop frequency for spatial dimensions (0.0-1.0)
+        d_t: normalized stop frequency for temporal dimension (0.0-1.0)
+    """
+    T, H, W = shape[-3], shape[-2], shape[-1]
+    mask = torch.zeros(shape)
+    if d_s==0 or d_t==0:
+        return mask
+    for t in range(T):
+        for h in range(H):
+            for w in range(W):
+                d_square = (((d_s/d_t)*(2*t/T-1))**2 + (2*h/H-1)**2 + (2*w/W-1)**2)
+                mask[..., t,h,w] = 1 / (1 + (d_square / d_s**2)**n)
+    return mask
+
+
+def ideal_low_pass_filter(shape, d_s=0.25, d_t=0.25):
+    """
+    Compute the ideal low pass filter mask.
+
+    Args:
+        shape: shape of the filter (volume)
+        d_s: normalized stop frequency for spatial dimensions (0.0-1.0)
+        d_t: normalized stop frequency for temporal dimension (0.0-1.0)
+    """
+    T, H, W = shape[-3], shape[-2], shape[-1]
+    mask = torch.zeros(shape)
+    if d_s==0 or d_t==0:
+        return mask
+    for t in range(T):
+        for h in range(H):
+            for w in range(W):
+                d_square = (((d_s/d_t)*(2*t/T-1))**2 + (2*h/H-1)**2 + (2*w/W-1)**2)
+                mask[..., t,h,w] =  1 if d_square <= d_s*2 else 0
+    return mask
+
+
+def box_low_pass_filter(shape, d_s=0.25, d_t=0.25):
+    """
+    Compute the ideal low pass filter mask (approximated version).
+
+    Args:
+        shape: shape of the filter (volume)
+        d_s: normalized stop frequency for spatial dimensions (0.0-1.0)
+        d_t: normalized stop frequency for temporal dimension (0.0-1.0)
+    """
+    T, H, W = shape[-3], shape[-2], shape[-1]
+    mask = torch.zeros(shape)
+    if d_s==0 or d_t==0:
+        return mask
+
+    threshold_s = round(int(H // 2) * d_s)
+    threshold_t = round(T // 2 * d_t)
+
+    cframe, crow, ccol = T // 2, H // 2, W //2
+    mask[..., cframe - threshold_t:cframe + threshold_t, crow - threshold_s:crow + threshold_s, ccol - threshold_s:ccol + threshold_s] = 1.0
+
+    return mask