Create app.py

app.py

@@ -0,0 +1,170 @@
+import os
+
+os.system("git clone --recursive https://github.com/JD-P/cloob-latent-diffusion")
+os.system("cd cloob-latent-diffusion;pip install omegaconf pillow pytorch-lightning einops wandb ftfy regex ./CLIP")
+
+import argparse
+from functools import partial
+from pathlib import Path
+import sys
+sys.path.append('./cloob-latent-diffusion')
+sys.path.append('./cloob-latent-diffusion/cloob-training')
+sys.path.append('./cloob-latent-diffusion/latent-diffusion')
+sys.path.append('./cloob-latent-diffusion/taming-transformers')
+sys.path.append('./cloob-latent-diffusion/v-diffusion-pytorch')
+from omegaconf import OmegaConf
+from PIL import Image
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torchvision import transforms
+from torchvision.transforms import functional as TF
+from tqdm import trange
+from CLIP import clip
+from cloob_training import model_pt, pretrained
+import ldm.models.autoencoder
+from diffusion import sampling, utils
+import train_latent_diffusion as train
+from huggingface_hub import hf_hub_url, cached_download
+import random
+
+# Download the model files
+checkpoint = cached_download(hf_hub_url("huggan/distill-ccld-wa", filename="model_student.ckpt"))
+ae_model_path = cached_download(hf_hub_url("huggan/ccld_wa", filename="ae_model.ckpt"))
+ae_config_path = cached_download(hf_hub_url("huggan/ccld_wa", filename="ae_model.yaml"))
+
+# Define a few utility functions
+
+def parse_prompt(prompt, default_weight=3.):
+    if prompt.startswith('http://') or prompt.startswith('https://'):
+        vals = prompt.rsplit(':', 2)
+        vals = [vals[0] + ':' + vals[1], *vals[2:]]
+    else:
+        vals = prompt.rsplit(':', 1)
+    vals = vals + ['', default_weight][len(vals):]
+    return vals[0], float(vals[1])
+
+
+def resize_and_center_crop(image, size):
+    fac = max(size[0] / image.size[0], size[1] / image.size[1])
+    image = image.resize((int(fac * image.size[0]), int(fac * image.size[1])), Image.LANCZOS)
+    return TF.center_crop(image, size[::-1])
+
+
+# Load the models
+device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
+print('Using device:', device)
+print('loading models')
+# autoencoder
+ae_config = OmegaConf.load(ae_config_path)
+ae_model = ldm.models.autoencoder.AutoencoderKL(**ae_config.model.params)
+ae_model.eval().requires_grad_(False).to(device)
+ae_model.load_state_dict(torch.load(ae_model_path))
+n_ch, side_y, side_x = 4, 32, 32
+
+# diffusion model
+model = train.DiffusionModel(192, [1,1,2,2], autoencoder_scale=torch.tensor(4.3084))
+model.load_state_dict(torch.load(checkpoint, map_location='cpu'))
+model = model.to(device).eval().requires_grad_(False)
+
+# CLOOB
+cloob_config = pretrained.get_config('cloob_laion_400m_vit_b_16_16_epochs')
+cloob = model_pt.get_pt_model(cloob_config)
+checkpoint = pretrained.download_checkpoint(cloob_config)
+cloob.load_state_dict(model_pt.get_pt_params(cloob_config, checkpoint))
+cloob.eval().requires_grad_(False).to(device)
+
+
+# The key function: returns a list of n PIL images
+def generate(n=1, prompts=['a red circle'], images=[], seed=42, steps=15,
+             method='plms', eta=None):
+  zero_embed = torch.zeros([1, cloob.config['d_embed']], device=device)
+  target_embeds, weights = [zero_embed], []
+
+  for prompt in prompts:
+      txt, weight = parse_prompt(prompt)
+      target_embeds.append(cloob.text_encoder(cloob.tokenize(txt).to(device)).float())
+      weights.append(weight)
+
+  for prompt in images:
+      path, weight = parse_prompt(prompt)
+      img = Image.open(utils.fetch(path)).convert('RGB')
+      clip_size = cloob.config['image_encoder']['image_size']
+      img = resize_and_center_crop(img, (clip_size, clip_size))
+      batch = TF.to_tensor(img)[None].to(device)
+      embed = F.normalize(cloob.image_encoder(cloob.normalize(batch)).float(), dim=-1)
+      target_embeds.append(embed)
+      weights.append(weight)
+
+  weights = torch.tensor([1 - sum(weights), *weights], device=device)
+
+  torch.manual_seed(seed)
+
+  def cfg_model_fn(x, t):
+      n = x.shape[0]
+      n_conds = len(target_embeds)
+      x_in = x.repeat([n_conds, 1, 1, 1])
+      t_in = t.repeat([n_conds])
+      clip_embed_in = torch.cat([*target_embeds]).repeat_interleave(n, 0)
+      vs = model(x_in, t_in, clip_embed_in).view([n_conds, n, *x.shape[1:]])
+      v = vs.mul(weights[:, None, None, None, None]).sum(0)
+      return v
+
+  def run(x, steps):
+      if method == 'ddpm':
+          return sampling.sample(cfg_model_fn, x, steps, 1., {})
+      if method == 'ddim':
+          return sampling.sample(cfg_model_fn, x, steps, eta, {})
+      if method == 'prk':
+          return sampling.prk_sample(cfg_model_fn, x, steps, {})
+      if method == 'plms':
+          return sampling.plms_sample(cfg_model_fn, x, steps, {})
+      if method == 'pie':
+          return sampling.pie_sample(cfg_model_fn, x, steps, {})
+      if method == 'plms2':
+          return sampling.plms2_sample(cfg_model_fn, x, steps, {})
+      assert False
+
+  batch_size = n
+  x = torch.randn([n, n_ch, side_y, side_x], device=device)
+  t = torch.linspace(1, 0, steps + 1, device=device)[:-1]
+  steps = utils.get_spliced_ddpm_cosine_schedule(t)
+  pil_ims = []
+  for i in trange(0, n, batch_size):
+      cur_batch_size = min(n - i, batch_size)
+      out_latents = run(x[i:i+cur_batch_size], steps)
+      outs = ae_model.decode(out_latents * torch.tensor(2.55).to(device))
+      for j, out in enumerate(outs):
+          pil_ims.append(utils.to_pil_image(out))
+
+  return pil_ims
+  
+  
+import gradio as gr
+
+def gen_ims(prompt, im_prompt=None, seed=None, n_steps=10, method='plms'):
+  if seed == None :
+    seed = random.randint(0, 10000)
+  print( prompt, im_prompt, seed, n_steps)
+  prompts = [prompt]
+  im_prompts = []
+  if im_prompt != None:
+    im_prompts = [im_prompt]
+  pil_ims = generate(n=1, prompts=prompts, images=im_prompts, seed=seed, steps=n_steps, method=method)
+  return pil_ims[0]
+
+iface = gr.Interface(fn=gen_ims, 
+  inputs=[#gr.inputs.Slider(minimum=1, maximum=1, step=1, default=1,label="Number of images"),
+          #gr.inputs.Slider(minimum=0, maximum=200, step=1, label='Random seed', default=0),
+          gr.inputs.Textbox(label="Text prompt"),
+          gr.inputs.Image(optional=True, label="Image prompt", type='filepath'),
+          #gr.inputs.Slider(minimum=10, maximum=35, step=1, default=15,label="Number of steps")
+          ], 
+  outputs=[gr.outputs.Image(type="pil", label="Generated Image")],
+  examples=[["An iceberg, oil on canvas"],["A martian landscape, in the style of Monet"], ['A peaceful meadow, pastel crayons'], ["A painting of a vase of flowers"], ["A ship leaving the port in the summer, oil on canvas"]],
+  title='Generate art from text prompts :',
+  description="By typing a text prompt or providing an image prompt, and pressing submit you can generate images based on this prompt. The model was trained on images from the [WikiArt](https://huggingface.co/datasets/huggan/wikiart) dataset, comprised mostly of paintings.",
+  article = 'The model is a distilled version of a cloob-conditioned latent diffusion model fine-tuned on the WikiArt dataset. You can find more information on this model on the [model card](https://huggingface.co/huggan/distill-ccld-wa). According to the [Latent Diffusion paper](https://arxiv.org/abs/2112.10752): \"Deep learning modules tend to reproduce or exacerbate biases that are already present in the data\".'
+
+)
+iface.launch(enable_queue=True) # , debug=True for colab debugging