akhaliq HF staff pcuenq HF staff commited on
Commit
b35f791
β€’
1 Parent(s): a3ad55f

Keep demo files only (#6)

Browse files

- Keep only the demo files. (2ed30fcbc11a7862c942c3ad78bcde83a08f049e)


Co-authored-by: Pedro Cuenca <pcuenq@users.noreply.huggingface.co>

This view is limited to 50 files because it contains too many changes. Β  See raw diff
Files changed (50) hide show
  1. configs/generate.yaml +0 -99
  2. configs/train.yaml +0 -113
  3. dataset_creation/generate_img_dataset.py +0 -315
  4. dataset_creation/generate_txt_dataset.py +0 -113
  5. dataset_creation/prepare_dataset.py +0 -29
  6. dataset_creation/prepare_for_gpt.py +0 -25
  7. edit_cli.py +0 -128
  8. edit_dataset.py +0 -121
  9. environment.yaml +0 -38
  10. imgs/dataset.jpg +0 -0
  11. imgs/edit_app.jpg +0 -0
  12. imgs/prompt_app.jpg +0 -0
  13. main.py +0 -799
  14. metrics/clip_similarity.py +0 -47
  15. metrics/compute_metrics.py +0 -235
  16. prompt_app.py +0 -55
  17. requirements.txt +3 -24
  18. scripts/download_checkpoints.sh +0 -7
  19. scripts/download_data.sh +0 -27
  20. scripts/download_pretrained_sd.sh +0 -7
  21. stable_diffusion/LICENSE +0 -82
  22. stable_diffusion/README.md +0 -215
  23. stable_diffusion/Stable_Diffusion_v1_Model_Card.md +0 -144
  24. stable_diffusion/assets/a-painting-of-a-fire.png +0 -0
  25. stable_diffusion/assets/a-photograph-of-a-fire.png +0 -0
  26. stable_diffusion/assets/a-shirt-with-a-fire-printed-on-it.png +0 -0
  27. stable_diffusion/assets/a-shirt-with-the-inscription-'fire'.png +0 -0
  28. stable_diffusion/assets/a-watercolor-painting-of-a-fire.png +0 -0
  29. stable_diffusion/assets/birdhouse.png +0 -0
  30. stable_diffusion/assets/fire.png +0 -0
  31. stable_diffusion/assets/inpainting.png +0 -0
  32. stable_diffusion/assets/modelfigure.png +0 -0
  33. stable_diffusion/assets/rdm-preview.jpg +0 -0
  34. stable_diffusion/assets/reconstruction1.png +0 -0
  35. stable_diffusion/assets/reconstruction2.png +0 -0
  36. stable_diffusion/assets/results.gif.REMOVED.git-id +0 -1
  37. stable_diffusion/assets/rick.jpeg +0 -0
  38. stable_diffusion/assets/stable-samples/img2img/mountains-1.png +0 -0
  39. stable_diffusion/assets/stable-samples/img2img/mountains-2.png +0 -0
  40. stable_diffusion/assets/stable-samples/img2img/mountains-3.png +0 -0
  41. stable_diffusion/assets/stable-samples/img2img/sketch-mountains-input.jpg +0 -0
  42. stable_diffusion/assets/stable-samples/img2img/upscaling-in.png.REMOVED.git-id +0 -1
  43. stable_diffusion/assets/stable-samples/img2img/upscaling-out.png.REMOVED.git-id +0 -1
  44. stable_diffusion/assets/stable-samples/txt2img/000002025.png +0 -0
  45. stable_diffusion/assets/stable-samples/txt2img/000002035.png +0 -0
  46. stable_diffusion/assets/stable-samples/txt2img/merged-0005.png.REMOVED.git-id +0 -1
  47. stable_diffusion/assets/stable-samples/txt2img/merged-0006.png.REMOVED.git-id +0 -1
  48. stable_diffusion/assets/stable-samples/txt2img/merged-0007.png.REMOVED.git-id +0 -1
  49. stable_diffusion/assets/the-earth-is-on-fire,-oil-on-canvas.png +0 -0
  50. stable_diffusion/assets/txt2img-convsample.png +0 -0
configs/generate.yaml DELETED
@@ -1,99 +0,0 @@
1
- # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
- # See more details in LICENSE.
3
-
4
- model:
5
- base_learning_rate: 1.0e-04
6
- target: ldm.models.diffusion.ddpm_edit.LatentDiffusion
7
- params:
8
- linear_start: 0.00085
9
- linear_end: 0.0120
10
- num_timesteps_cond: 1
11
- log_every_t: 200
12
- timesteps: 1000
13
- first_stage_key: edited
14
- cond_stage_key: edit
15
- # image_size: 64
16
- # image_size: 32
17
- image_size: 16
18
- channels: 4
19
- cond_stage_trainable: false # Note: different from the one we trained before
20
- conditioning_key: hybrid
21
- monitor: val/loss_simple_ema
22
- scale_factor: 0.18215
23
- use_ema: true
24
- load_ema: true
25
-
26
- scheduler_config: # 10000 warmup steps
27
- target: ldm.lr_scheduler.LambdaLinearScheduler
28
- params:
29
- warm_up_steps: [ 0 ]
30
- cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
31
- f_start: [ 1.e-6 ]
32
- f_max: [ 1. ]
33
- f_min: [ 1. ]
34
-
35
- unet_config:
36
- target: ldm.modules.diffusionmodules.openaimodel.UNetModel
37
- params:
38
- image_size: 32 # unused
39
- in_channels: 8
40
- out_channels: 4
41
- model_channels: 320
42
- attention_resolutions: [ 4, 2, 1 ]
43
- num_res_blocks: 2
44
- channel_mult: [ 1, 2, 4, 4 ]
45
- num_heads: 8
46
- use_spatial_transformer: True
47
- transformer_depth: 1
48
- context_dim: 768
49
- use_checkpoint: True
50
- legacy: False
51
-
52
- first_stage_config:
53
- target: ldm.models.autoencoder.AutoencoderKL
54
- params:
55
- embed_dim: 4
56
- monitor: val/rec_loss
57
- ddconfig:
58
- double_z: true
59
- z_channels: 4
60
- resolution: 256
61
- in_channels: 3
62
- out_ch: 3
63
- ch: 128
64
- ch_mult:
65
- - 1
66
- - 2
67
- - 4
68
- - 4
69
- num_res_blocks: 2
70
- attn_resolutions: []
71
- dropout: 0.0
72
- lossconfig:
73
- target: torch.nn.Identity
74
-
75
- cond_stage_config:
76
- target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
77
-
78
- data:
79
- target: main.DataModuleFromConfig
80
- params:
81
- batch_size: 128
82
- num_workers: 1
83
- wrap: false
84
- validation:
85
- target: edit_dataset.EditDataset
86
- params:
87
- path: data/clip-filtered-dataset
88
- cache_dir: data/
89
- cache_name: data_10k
90
- split: val
91
- min_text_sim: 0.2
92
- min_image_sim: 0.75
93
- min_direction_sim: 0.2
94
- max_samples_per_prompt: 1
95
- min_resize_res: 512
96
- max_resize_res: 512
97
- crop_res: 512
98
- output_as_edit: False
99
- real_input: True
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
configs/train.yaml DELETED
@@ -1,113 +0,0 @@
1
- # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
- # See more details in LICENSE.
3
-
4
- model:
5
- base_learning_rate: 1.0e-04
6
- target: ldm.models.diffusion.ddpm_edit.LatentDiffusion
7
- params:
8
- ckpt_path: stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly.ckpt
9
- linear_start: 0.00085
10
- linear_end: 0.0120
11
- num_timesteps_cond: 1
12
- log_every_t: 200
13
- timesteps: 1000
14
- first_stage_key: edited
15
- cond_stage_key: edit
16
- image_size: 32
17
- channels: 4
18
- cond_stage_trainable: false # Note: different from the one we trained before
19
- conditioning_key: hybrid
20
- monitor: val/loss_simple_ema
21
- scale_factor: 0.18215
22
- use_ema: true
23
- load_ema: false
24
-
25
- scheduler_config: # 10000 warmup steps
26
- target: ldm.lr_scheduler.LambdaLinearScheduler
27
- params:
28
- warm_up_steps: [ 0 ]
29
- cycle_lengths: [ 10000000000000 ] # incredibly large number to prevent corner cases
30
- f_start: [ 1.e-6 ]
31
- f_max: [ 1. ]
32
- f_min: [ 1. ]
33
-
34
- unet_config:
35
- target: ldm.modules.diffusionmodules.openaimodel.UNetModel
36
- params:
37
- image_size: 32 # unused
38
- in_channels: 8
39
- out_channels: 4
40
- model_channels: 320
41
- attention_resolutions: [ 4, 2, 1 ]
42
- num_res_blocks: 2
43
- channel_mult: [ 1, 2, 4, 4 ]
44
- num_heads: 8
45
- use_spatial_transformer: True
46
- transformer_depth: 1
47
- context_dim: 768
48
- use_checkpoint: True
49
- legacy: False
50
-
51
- first_stage_config:
52
- target: ldm.models.autoencoder.AutoencoderKL
53
- params:
54
- embed_dim: 4
55
- monitor: val/rec_loss
56
- ddconfig:
57
- double_z: true
58
- z_channels: 4
59
- resolution: 256
60
- in_channels: 3
61
- out_ch: 3
62
- ch: 128
63
- ch_mult:
64
- - 1
65
- - 2
66
- - 4
67
- - 4
68
- num_res_blocks: 2
69
- attn_resolutions: []
70
- dropout: 0.0
71
- lossconfig:
72
- target: torch.nn.Identity
73
-
74
- cond_stage_config:
75
- target: ldm.modules.encoders.modules.FrozenCLIPEmbedder
76
-
77
- data:
78
- target: main.DataModuleFromConfig
79
- params:
80
- batch_size: 32
81
- num_workers: 2
82
- train:
83
- target: edit_dataset.EditDataset
84
- params:
85
- path: data/clip-filtered-dataset
86
- split: train
87
- min_resize_res: 256
88
- max_resize_res: 256
89
- crop_res: 256
90
- flip_prob: 0.5
91
- validation:
92
- target: edit_dataset.EditDataset
93
- params:
94
- path: data/clip-filtered-dataset
95
- split: val
96
- min_resize_res: 256
97
- max_resize_res: 256
98
- crop_res: 256
99
-
100
- lightning:
101
- callbacks:
102
- image_logger:
103
- target: main.ImageLogger
104
- params:
105
- batch_frequency: 2000
106
- max_images: 2
107
- increase_log_steps: False
108
-
109
- trainer:
110
- max_epochs: 2000
111
- benchmark: True
112
- accumulate_grad_batches: 4
113
- check_val_every_n_epoch: 4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dataset_creation/generate_img_dataset.py DELETED
@@ -1,315 +0,0 @@
1
- import argparse
2
- import json
3
- import sys
4
- from pathlib import Path
5
-
6
- import k_diffusion
7
- import numpy as np
8
- import torch
9
- import torch.nn as nn
10
- from einops import rearrange, repeat
11
- from omegaconf import OmegaConf
12
- from PIL import Image
13
- from pytorch_lightning import seed_everything
14
- from tqdm import tqdm
15
-
16
- sys.path.append("./")
17
- sys.path.append("./stable_diffusion")
18
-
19
- from ldm.modules.attention import CrossAttention
20
- from ldm.util import instantiate_from_config
21
- from metrics.clip_similarity import ClipSimilarity
22
-
23
-
24
- ################################################################################
25
- # Modified K-diffusion Euler ancestral sampler with prompt-to-prompt.
26
- # https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py
27
-
28
-
29
- def append_dims(x, target_dims):
30
- """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
31
- dims_to_append = target_dims - x.ndim
32
- if dims_to_append < 0:
33
- raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
34
- return x[(...,) + (None,) * dims_to_append]
35
-
36
-
37
- def to_d(x, sigma, denoised):
38
- """Converts a denoiser output to a Karras ODE derivative."""
39
- return (x - denoised) / append_dims(sigma, x.ndim)
40
-
41
-
42
- def get_ancestral_step(sigma_from, sigma_to):
43
- """Calculates the noise level (sigma_down) to step down to and the amount
44
- of noise to add (sigma_up) when doing an ancestral sampling step."""
45
- sigma_up = min(sigma_to, (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5)
46
- sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5
47
- return sigma_down, sigma_up
48
-
49
-
50
- def sample_euler_ancestral(model, x, sigmas, prompt2prompt_threshold=0.0, **extra_args):
51
- """Ancestral sampling with Euler method steps."""
52
- s_in = x.new_ones([x.shape[0]])
53
- for i in range(len(sigmas) - 1):
54
- prompt_to_prompt = prompt2prompt_threshold > i / (len(sigmas) - 2)
55
- for m in model.modules():
56
- if isinstance(m, CrossAttention):
57
- m.prompt_to_prompt = prompt_to_prompt
58
- denoised = model(x, sigmas[i] * s_in, **extra_args)
59
- sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1])
60
- d = to_d(x, sigmas[i], denoised)
61
- # Euler method
62
- dt = sigma_down - sigmas[i]
63
- x = x + d * dt
64
- if sigmas[i + 1] > 0:
65
- # Make noise the same across all samples in batch.
66
- x = x + torch.randn_like(x[:1]) * sigma_up
67
- return x
68
-
69
-
70
- ################################################################################
71
-
72
-
73
- def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False):
74
- print(f"Loading model from {ckpt}")
75
- pl_sd = torch.load(ckpt, map_location="cpu")
76
- if "global_step" in pl_sd:
77
- print(f"Global Step: {pl_sd['global_step']}")
78
- sd = pl_sd["state_dict"]
79
- if vae_ckpt is not None:
80
- print(f"Loading VAE from {vae_ckpt}")
81
- vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
82
- sd = {
83
- k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
84
- for k, v in sd.items()
85
- }
86
- model = instantiate_from_config(config.model)
87
- m, u = model.load_state_dict(sd, strict=False)
88
- if len(m) > 0 and verbose:
89
- print("missing keys:")
90
- print(m)
91
- if len(u) > 0 and verbose:
92
- print("unexpected keys:")
93
- print(u)
94
- return model
95
-
96
-
97
- class CFGDenoiser(nn.Module):
98
- def __init__(self, model):
99
- super().__init__()
100
- self.inner_model = model
101
-
102
- def forward(self, x, sigma, uncond, cond, cfg_scale):
103
- x_in = torch.cat([x] * 2)
104
- sigma_in = torch.cat([sigma] * 2)
105
- cond_in = torch.cat([uncond, cond])
106
- uncond, cond = self.inner_model(x_in, sigma_in, cond=cond_in).chunk(2)
107
- return uncond + (cond - uncond) * cfg_scale
108
-
109
-
110
- def to_pil(image: torch.Tensor) -> Image.Image:
111
- image = 255.0 * rearrange(image.cpu().numpy(), "c h w -> h w c")
112
- image = Image.fromarray(image.astype(np.uint8))
113
- return image
114
-
115
-
116
- def main():
117
- parser = argparse.ArgumentParser()
118
- parser.add_argument(
119
- "--out_dir",
120
- type=str,
121
- required=True,
122
- help="Path to output dataset directory.",
123
- )
124
- parser.add_argument(
125
- "--prompts_file",
126
- type=str,
127
- required=True,
128
- help="Path to prompts .jsonl file.",
129
- )
130
- parser.add_argument(
131
- "--ckpt",
132
- type=str,
133
- default="stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly.ckpt",
134
- help="Path to stable diffusion checkpoint.",
135
- )
136
- parser.add_argument(
137
- "--vae-ckpt",
138
- type=str,
139
- default="stable_diffusion/models/ldm/stable-diffusion-v1/vae-ft-mse-840000-ema-pruned.ckpt",
140
- help="Path to vae checkpoint.",
141
- )
142
- parser.add_argument(
143
- "--steps",
144
- type=int,
145
- default=100,
146
- help="Number of sampling steps.",
147
- )
148
- parser.add_argument(
149
- "--n-samples",
150
- type=int,
151
- default=100,
152
- help="Number of samples to generate per prompt (before CLIP filtering).",
153
- )
154
- parser.add_argument(
155
- "--max-out-samples",
156
- type=int,
157
- default=4,
158
- help="Max number of output samples to save per prompt (after CLIP filtering).",
159
- )
160
- parser.add_argument(
161
- "--n-partitions",
162
- type=int,
163
- default=1,
164
- help="Number of total partitions.",
165
- )
166
- parser.add_argument(
167
- "--partition",
168
- type=int,
169
- default=0,
170
- help="Partition index.",
171
- )
172
- parser.add_argument(
173
- "--min-p2p",
174
- type=float,
175
- default=0.1,
176
- help="Min prompt2prompt threshold (portion of denoising for which to fix self attention maps).",
177
- )
178
- parser.add_argument(
179
- "--max-p2p",
180
- type=float,
181
- default=0.9,
182
- help="Max prompt2prompt threshold (portion of denoising for which to fix self attention maps).",
183
- )
184
- parser.add_argument(
185
- "--min-cfg",
186
- type=float,
187
- default=7.5,
188
- help="Min classifier free guidance scale.",
189
- )
190
- parser.add_argument(
191
- "--max-cfg",
192
- type=float,
193
- default=15,
194
- help="Max classifier free guidance scale.",
195
- )
196
- parser.add_argument(
197
- "--clip-threshold",
198
- type=float,
199
- default=0.2,
200
- help="CLIP threshold for text-image similarity of each image.",
201
- )
202
- parser.add_argument(
203
- "--clip-dir-threshold",
204
- type=float,
205
- default=0.2,
206
- help="Directional CLIP threshold for similarity of change between pairs of text and pairs of images.",
207
- )
208
- parser.add_argument(
209
- "--clip-img-threshold",
210
- type=float,
211
- default=0.7,
212
- help="CLIP threshold for image-image similarity.",
213
- )
214
- opt = parser.parse_args()
215
-
216
- global_seed = torch.randint(1 << 32, ()).item()
217
- print(f"Global seed: {global_seed}")
218
- seed_everything(global_seed)
219
-
220
- model = load_model_from_config(
221
- OmegaConf.load("stable_diffusion/configs/stable-diffusion/v1-inference.yaml"),
222
- ckpt=opt.ckpt,
223
- vae_ckpt=opt.vae_ckpt,
224
- )
225
- model.cuda().eval()
226
- model_wrap = k_diffusion.external.CompVisDenoiser(model)
227
-
228
- clip_similarity = ClipSimilarity().cuda()
229
-
230
- out_dir = Path(opt.out_dir)
231
- out_dir.mkdir(exist_ok=True, parents=True)
232
-
233
- with open(opt.prompts_file) as fp:
234
- prompts = [json.loads(line) for line in fp]
235
-
236
- print(f"Partition index {opt.partition} ({opt.partition + 1} / {opt.n_partitions})")
237
- prompts = np.array_split(list(enumerate(prompts)), opt.n_partitions)[opt.partition]
238
-
239
- with torch.no_grad(), torch.autocast("cuda"), model.ema_scope():
240
- uncond = model.get_learned_conditioning(2 * [""])
241
- sigmas = model_wrap.get_sigmas(opt.steps)
242
-
243
- for i, prompt in tqdm(prompts, desc="Prompts"):
244
- prompt_dir = out_dir.joinpath(f"{i:07d}")
245
- prompt_dir.mkdir(exist_ok=True)
246
-
247
- with open(prompt_dir.joinpath("prompt.json"), "w") as fp:
248
- json.dump(prompt, fp)
249
-
250
- cond = model.get_learned_conditioning([prompt["caption"], prompt["output"]])
251
- results = {}
252
-
253
- with tqdm(total=opt.n_samples, desc="Samples") as progress_bar:
254
-
255
- while len(results) < opt.n_samples:
256
- seed = torch.randint(1 << 32, ()).item()
257
- if seed in results:
258
- continue
259
- torch.manual_seed(seed)
260
-
261
- x = torch.randn(1, 4, 512 // 8, 512 // 8, device="cuda") * sigmas[0]
262
- x = repeat(x, "1 ... -> n ...", n=2)
263
-
264
- model_wrap_cfg = CFGDenoiser(model_wrap)
265
- p2p_threshold = opt.min_p2p + torch.rand(()).item() * (opt.max_p2p - opt.min_p2p)
266
- cfg_scale = opt.min_cfg + torch.rand(()).item() * (opt.max_cfg - opt.min_cfg)
267
- extra_args = {"cond": cond, "uncond": uncond, "cfg_scale": cfg_scale}
268
- samples_ddim = sample_euler_ancestral(model_wrap_cfg, x, sigmas, p2p_threshold, **extra_args)
269
- x_samples_ddim = model.decode_first_stage(samples_ddim)
270
- x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
271
-
272
- x0 = x_samples_ddim[0]
273
- x1 = x_samples_ddim[1]
274
-
275
- clip_sim_0, clip_sim_1, clip_sim_dir, clip_sim_image = clip_similarity(
276
- x0[None], x1[None], [prompt["caption"]], [prompt["output"]]
277
- )
278
-
279
- results[seed] = dict(
280
- image_0=to_pil(x0),
281
- image_1=to_pil(x1),
282
- p2p_threshold=p2p_threshold,
283
- cfg_scale=cfg_scale,
284
- clip_sim_0=clip_sim_0[0].item(),
285
- clip_sim_1=clip_sim_1[0].item(),
286
- clip_sim_dir=clip_sim_dir[0].item(),
287
- clip_sim_image=clip_sim_image[0].item(),
288
- )
289
-
290
- progress_bar.update()
291
-
292
- # CLIP filter to get best samples for each prompt.
293
- metadata = [
294
- (result["clip_sim_dir"], seed)
295
- for seed, result in results.items()
296
- if result["clip_sim_image"] >= opt.clip_img_threshold
297
- and result["clip_sim_dir"] >= opt.clip_dir_threshold
298
- and result["clip_sim_0"] >= opt.clip_threshold
299
- and result["clip_sim_1"] >= opt.clip_threshold
300
- ]
301
- metadata.sort(reverse=True)
302
- for _, seed in metadata[: opt.max_out_samples]:
303
- result = results[seed]
304
- image_0 = result.pop("image_0")
305
- image_1 = result.pop("image_1")
306
- image_0.save(prompt_dir.joinpath(f"{seed}_0.jpg"), quality=100)
307
- image_1.save(prompt_dir.joinpath(f"{seed}_1.jpg"), quality=100)
308
- with open(prompt_dir.joinpath(f"metadata.jsonl"), "a") as fp:
309
- fp.write(f"{json.dumps(dict(seed=seed, **result))}\n")
310
-
311
- print("Done.")
312
-
313
-
314
- if __name__ == "__main__":
315
- main()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dataset_creation/generate_txt_dataset.py DELETED
@@ -1,113 +0,0 @@
1
- from __future__ import annotations
2
-
3
- import json
4
- import time
5
- from argparse import ArgumentParser
6
- from pathlib import Path
7
- from typing import Optional
8
-
9
- import datasets
10
- import numpy as np
11
- import openai
12
- from tqdm.auto import tqdm
13
-
14
-
15
- DELIMITER_0 = "\n##\n"
16
- DELIMITER_1 = "\n%%\n"
17
- STOP = "\nEND"
18
-
19
-
20
- def generate(
21
- openai_model: str,
22
- caption: str,
23
- num_retries: int = 3,
24
- max_tokens: int = 256,
25
- temperature: float = 0.7,
26
- top_p: float = 1.0,
27
- frequency_penalty: float = 0.1,
28
- presence_penalty: float = 0.0,
29
- sleep_on_error: float = 1.0,
30
- ) -> Optional[tuple[str, str]]:
31
- for _ in range(1 + num_retries):
32
- try:
33
- response = openai.Completion.create(
34
- model=openai_model,
35
- prompt=caption + DELIMITER_0,
36
- temperature=temperature,
37
- max_tokens=max_tokens,
38
- top_p=top_p,
39
- frequency_penalty=frequency_penalty,
40
- presence_penalty=presence_penalty,
41
- stop=[STOP],
42
- )
43
- except Exception as e:
44
- print(e)
45
- time.sleep(sleep_on_error)
46
- continue
47
- output = response["choices"][0]["text"].split(DELIMITER_1)
48
- if len(output) == 2:
49
- instruction, edited_caption = output
50
- results = openai.Moderation.create([instruction, edited_caption])["results"]
51
- if results[0]["flagged"] or results[1]["flagged"]:
52
- continue
53
- if caption.strip().strip(".!?").lower() != edited_caption.strip().strip(".!?").lower():
54
- return instruction, edited_caption
55
-
56
-
57
- def main(openai_model: str, num_samples: int, num_partitions: int, partition: int, seed: int):
58
- dataset = datasets.load_dataset("ChristophSchuhmann/improved_aesthetics_6.5plus", split="train")
59
- # Other datasets we considered that may be worth trying:
60
- # dataset = datasets.load_dataset("ChristophSchuhmann/MS_COCO_2017_URL_TEXT", split="train")
61
- # dataset = datasets.load_dataset("laion/laion-coco", split="train")
62
-
63
- np.random.seed(seed)
64
- permutation = np.array_split(np.random.permutation(len(dataset)), num_partitions)[partition]
65
- dataset = dataset[permutation]
66
- captions = dataset["TEXT"]
67
- urls = dataset["URL"]
68
- output_path = f"data/dataset=laion-aesthetics-6.5_model={openai_model}_samples={num_samples}_partition={partition}.jsonl" # fmt: skip
69
- print(f"Prompt file path: {output_path}")
70
-
71
- count = 0
72
- caption_set = set()
73
- url_set = set()
74
-
75
- if Path(output_path).exists():
76
- with open(output_path, "r") as f:
77
- for line in tqdm(f, desc="Resuming from existing prompts"):
78
- prompt = json.loads(line)
79
- if prompt["caption"] not in caption_set and prompt["url"] not in url_set:
80
- caption_set.add(prompt["caption"])
81
- url_set.add(prompt["url"])
82
- count += 1
83
-
84
- with open(output_path, "a") as fp:
85
- with tqdm(total=num_samples - count, desc="Generating instructions and edited captions") as progress_bar:
86
- for caption, url in zip(captions, urls):
87
- if caption in caption_set or url in url_set:
88
- continue
89
- if openai.Moderation.create(caption)["results"][0]["flagged"]:
90
- continue
91
- edit_output = generate(openai_model, caption)
92
- if edit_output is not None:
93
- edit, output = edit_output
94
- fp.write(f"{json.dumps(dict(caption=caption, edit=edit, output=output, url=url))}\n")
95
- count += 1
96
- progress_bar.update()
97
- caption_set.add(caption)
98
- url_set.add(url)
99
- if count == num_samples:
100
- break
101
-
102
-
103
- if __name__ == "__main__":
104
- parser = ArgumentParser()
105
- parser.add_argument("--openai-api-key", required=True, type=str)
106
- parser.add_argument("--openai-model", required=True, type=str)
107
- parser.add_argument("--num-samples", default=10000, type=int)
108
- parser.add_argument("--num-partitions", default=1, type=int)
109
- parser.add_argument("--partition", default=0, type=int)
110
- parser.add_argument("--seed", default=0, type=int)
111
- args = parser.parse_args()
112
- openai.api_key = args.openai_api_key
113
- main(args.openai_model, args.num_samples, args.num_partitions, args.partition, args.seed)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dataset_creation/prepare_dataset.py DELETED
@@ -1,29 +0,0 @@
1
- import json
2
- from argparse import ArgumentParser
3
- from pathlib import Path
4
-
5
- from tqdm.auto import tqdm
6
-
7
-
8
- def main():
9
- parser = ArgumentParser()
10
- parser.add_argument("dataset_dir")
11
- args = parser.parse_args()
12
- dataset_dir = Path(args.dataset_dir)
13
-
14
- seeds = []
15
- with tqdm(desc="Listing dataset image seeds") as progress_bar:
16
- for prompt_dir in dataset_dir.iterdir():
17
- if prompt_dir.is_dir():
18
- prompt_seeds = [image_path.name.split("_")[0] for image_path in sorted(prompt_dir.glob("*_0.jpg"))]
19
- if len(prompt_seeds) > 0:
20
- seeds.append((prompt_dir.name, prompt_seeds))
21
- progress_bar.update()
22
- seeds.sort()
23
-
24
- with open(dataset_dir.joinpath("seeds.json"), "w") as f:
25
- json.dump(seeds, f)
26
-
27
-
28
- if __name__ == "__main__":
29
- main()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dataset_creation/prepare_for_gpt.py DELETED
@@ -1,25 +0,0 @@
1
- import json
2
- from argparse import ArgumentParser
3
-
4
- from generate_txt_dataset import DELIMITER_0, DELIMITER_1, STOP
5
-
6
-
7
- def main(input_path: str, output_path: str):
8
- with open(input_path) as f:
9
- prompts = [json.loads(l) for l in f]
10
-
11
- with open(output_path, "w") as f:
12
- for prompt in prompts:
13
- prompt_for_gpt = {
14
- "prompt": f"{prompt['input']}{DELIMITER_0}",
15
- "completion": f"{prompt['edit']}{DELIMITER_1}{prompt['output']}{STOP}",
16
- }
17
- f.write(f"{json.dumps(prompt_for_gpt)}\n")
18
-
19
-
20
- if __name__ == "__main__":
21
- parser = ArgumentParser()
22
- parser.add_argument("--input-path", required=True, type=str)
23
- parser.add_argument("--output-path", required=True, type=str)
24
- args = parser.parse_args()
25
- main(args.input_path, args.output_path)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
edit_cli.py DELETED
@@ -1,128 +0,0 @@
1
- from __future__ import annotations
2
-
3
- import math
4
- import random
5
- import sys
6
- from argparse import ArgumentParser
7
-
8
- import einops
9
- import k_diffusion as K
10
- import numpy as np
11
- import torch
12
- import torch.nn as nn
13
- from einops import rearrange
14
- from omegaconf import OmegaConf
15
- from PIL import Image, ImageOps
16
- from torch import autocast
17
-
18
- sys.path.append("./stable_diffusion")
19
-
20
- from stable_diffusion.ldm.util import instantiate_from_config
21
-
22
-
23
- class CFGDenoiser(nn.Module):
24
- def __init__(self, model):
25
- super().__init__()
26
- self.inner_model = model
27
-
28
- def forward(self, z, sigma, cond, uncond, text_cfg_scale, image_cfg_scale):
29
- cfg_z = einops.repeat(z, "1 ... -> n ...", n=3)
30
- cfg_sigma = einops.repeat(sigma, "1 ... -> n ...", n=3)
31
- cfg_cond = {
32
- "c_crossattn": [torch.cat([cond["c_crossattn"][0], uncond["c_crossattn"][0], uncond["c_crossattn"][0]])],
33
- "c_concat": [torch.cat([cond["c_concat"][0], cond["c_concat"][0], uncond["c_concat"][0]])],
34
- }
35
- out_cond, out_img_cond, out_uncond = self.inner_model(cfg_z, cfg_sigma, cond=cfg_cond).chunk(3)
36
- return out_uncond + text_cfg_scale * (out_cond - out_img_cond) + image_cfg_scale * (out_img_cond - out_uncond)
37
-
38
-
39
- def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False):
40
- print(f"Loading model from {ckpt}")
41
- pl_sd = torch.load(ckpt, map_location="cpu")
42
- if "global_step" in pl_sd:
43
- print(f"Global Step: {pl_sd['global_step']}")
44
- sd = pl_sd["state_dict"]
45
- if vae_ckpt is not None:
46
- print(f"Loading VAE from {vae_ckpt}")
47
- vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
48
- sd = {
49
- k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
50
- for k, v in sd.items()
51
- }
52
- model = instantiate_from_config(config.model)
53
- m, u = model.load_state_dict(sd, strict=False)
54
- if len(m) > 0 and verbose:
55
- print("missing keys:")
56
- print(m)
57
- if len(u) > 0 and verbose:
58
- print("unexpected keys:")
59
- print(u)
60
- return model
61
-
62
-
63
- def main():
64
- parser = ArgumentParser()
65
- parser.add_argument("--resolution", default=512, type=int)
66
- parser.add_argument("--steps", default=100, type=int)
67
- parser.add_argument("--config", default="configs/generate.yaml", type=str)
68
- parser.add_argument("--ckpt", default="checkpoints/instruct-pix2pix-00-22000.ckpt", type=str)
69
- parser.add_argument("--vae-ckpt", default=None, type=str)
70
- parser.add_argument("--input", required=True, type=str)
71
- parser.add_argument("--output", required=True, type=str)
72
- parser.add_argument("--edit", required=True, type=str)
73
- parser.add_argument("--cfg-text", default=7.5, type=float)
74
- parser.add_argument("--cfg-image", default=1.5, type=float)
75
- parser.add_argument("--seed", type=int)
76
- args = parser.parse_args()
77
-
78
- config = OmegaConf.load(args.config)
79
- model = load_model_from_config(config, args.ckpt, args.vae_ckpt)
80
- model.eval().cuda()
81
- model_wrap = K.external.CompVisDenoiser(model)
82
- model_wrap_cfg = CFGDenoiser(model_wrap)
83
- null_token = model.get_learned_conditioning([""])
84
-
85
- seed = random.randint(0, 100000) if args.seed is None else args.seed
86
- input_image = Image.open(args.input).convert("RGB")
87
- width, height = input_image.size
88
- factor = args.resolution / max(width, height)
89
- factor = math.ceil(min(width, height) * factor / 64) * 64 / min(width, height)
90
- width = int((width * factor) // 64) * 64
91
- height = int((height * factor) // 64) * 64
92
- input_image = ImageOps.fit(input_image, (width, height), method=Image.Resampling.LANCZOS)
93
-
94
- if args.edit == "":
95
- input_image.save(args.output)
96
- return
97
-
98
- with torch.no_grad(), autocast("cuda"), model.ema_scope():
99
- cond = {}
100
- cond["c_crossattn"] = [model.get_learned_conditioning([args.edit])]
101
- input_image = 2 * torch.tensor(np.array(input_image)).float() / 255 - 1
102
- input_image = rearrange(input_image, "h w c -> 1 c h w").to(model.device)
103
- cond["c_concat"] = [model.encode_first_stage(input_image).mode()]
104
-
105
- uncond = {}
106
- uncond["c_crossattn"] = [null_token]
107
- uncond["c_concat"] = [torch.zeros_like(cond["c_concat"][0])]
108
-
109
- sigmas = model_wrap.get_sigmas(args.steps)
110
-
111
- extra_args = {
112
- "cond": cond,
113
- "uncond": uncond,
114
- "text_cfg_scale": args.cfg_text,
115
- "image_cfg_scale": args.cfg_image,
116
- }
117
- torch.manual_seed(seed)
118
- z = torch.randn_like(cond["c_concat"][0]) * sigmas[0]
119
- z = K.sampling.sample_euler_ancestral(model_wrap_cfg, z, sigmas, extra_args=extra_args)
120
- x = model.decode_first_stage(z)
121
- x = torch.clamp((x + 1.0) / 2.0, min=0.0, max=1.0)
122
- x = 255.0 * rearrange(x, "1 c h w -> h w c")
123
- edited_image = Image.fromarray(x.type(torch.uint8).cpu().numpy())
124
- edited_image.save(args.output)
125
-
126
-
127
- if __name__ == "__main__":
128
- main()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
edit_dataset.py DELETED
@@ -1,121 +0,0 @@
1
- from __future__ import annotations
2
-
3
- import json
4
- import math
5
- from pathlib import Path
6
- from typing import Any
7
-
8
- import numpy as np
9
- import torch
10
- import torchvision
11
- from einops import rearrange
12
- from PIL import Image
13
- from torch.utils.data import Dataset
14
-
15
-
16
- class EditDataset(Dataset):
17
- def __init__(
18
- self,
19
- path: str,
20
- split: str = "train",
21
- splits: tuple[float, float, float] = (0.9, 0.05, 0.05),
22
- min_resize_res: int = 256,
23
- max_resize_res: int = 256,
24
- crop_res: int = 256,
25
- flip_prob: float = 0.0,
26
- ):
27
- assert split in ("train", "val", "test")
28
- assert sum(splits) == 1
29
- self.path = path
30
- self.min_resize_res = min_resize_res
31
- self.max_resize_res = max_resize_res
32
- self.crop_res = crop_res
33
- self.flip_prob = flip_prob
34
-
35
- with open(Path(self.path, "seeds.json")) as f:
36
- self.seeds = json.load(f)
37
-
38
- split_0, split_1 = {
39
- "train": (0.0, splits[0]),
40
- "val": (splits[0], splits[0] + splits[1]),
41
- "test": (splits[0] + splits[1], 1.0),
42
- }[split]
43
-
44
- idx_0 = math.floor(split_0 * len(self.seeds))
45
- idx_1 = math.floor(split_1 * len(self.seeds))
46
- self.seeds = self.seeds[idx_0:idx_1]
47
-
48
- def __len__(self) -> int:
49
- return len(self.seeds)
50
-
51
- def __getitem__(self, i: int) -> dict[str, Any]:
52
- name, seeds = self.seeds[i]
53
- propt_dir = Path(self.path, name)
54
- seed = seeds[torch.randint(0, len(seeds), ()).item()]
55
- with open(propt_dir.joinpath("prompt.json")) as fp:
56
- prompt = json.load(fp)["edit"]
57
-
58
- image_0 = Image.open(propt_dir.joinpath(f"{seed}_0.jpg"))
59
- image_1 = Image.open(propt_dir.joinpath(f"{seed}_1.jpg"))
60
-
61
- reize_res = torch.randint(self.min_resize_res, self.max_resize_res + 1, ()).item()
62
- image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS)
63
- image_1 = image_1.resize((reize_res, reize_res), Image.Resampling.LANCZOS)
64
-
65
- image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w")
66
- image_1 = rearrange(2 * torch.tensor(np.array(image_1)).float() / 255 - 1, "h w c -> c h w")
67
-
68
- crop = torchvision.transforms.RandomCrop(self.crop_res)
69
- flip = torchvision.transforms.RandomHorizontalFlip(float(self.flip_prob))
70
- image_0, image_1 = flip(crop(torch.cat((image_0, image_1)))).chunk(2)
71
-
72
- return dict(edited=image_1, edit=dict(c_concat=image_0, c_crossattn=prompt))
73
-
74
-
75
- class EditDatasetEval(Dataset):
76
- def __init__(
77
- self,
78
- path: str,
79
- split: str = "train",
80
- splits: tuple[float, float, float] = (0.9, 0.05, 0.05),
81
- res: int = 256,
82
- ):
83
- assert split in ("train", "val", "test")
84
- assert sum(splits) == 1
85
- self.path = path
86
- self.res = res
87
-
88
- with open(Path(self.path, "seeds.json")) as f:
89
- self.seeds = json.load(f)
90
-
91
- split_0, split_1 = {
92
- "train": (0.0, splits[0]),
93
- "val": (splits[0], splits[0] + splits[1]),
94
- "test": (splits[0] + splits[1], 1.0),
95
- }[split]
96
-
97
- idx_0 = math.floor(split_0 * len(self.seeds))
98
- idx_1 = math.floor(split_1 * len(self.seeds))
99
- self.seeds = self.seeds[idx_0:idx_1]
100
-
101
- def __len__(self) -> int:
102
- return len(self.seeds)
103
-
104
- def __getitem__(self, i: int) -> dict[str, Any]:
105
- name, seeds = self.seeds[i]
106
- propt_dir = Path(self.path, name)
107
- seed = seeds[torch.randint(0, len(seeds), ()).item()]
108
- with open(propt_dir.joinpath("prompt.json")) as fp:
109
- prompt = json.load(fp)
110
- edit = prompt["edit"]
111
- input_prompt = prompt["input"]
112
- output_prompt = prompt["output"]
113
-
114
- image_0 = Image.open(propt_dir.joinpath(f"{seed}_0.jpg"))
115
-
116
- reize_res = torch.randint(self.res, self.res + 1, ()).item()
117
- image_0 = image_0.resize((reize_res, reize_res), Image.Resampling.LANCZOS)
118
-
119
- image_0 = rearrange(2 * torch.tensor(np.array(image_0)).float() / 255 - 1, "h w c -> c h w")
120
-
121
- return dict(image_0=image_0, input_prompt=input_prompt, edit=edit, output_prompt=output_prompt)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
environment.yaml DELETED
@@ -1,38 +0,0 @@
1
- # File modified by authors of InstructPix2Pix from original (https://github.com/CompVis/stable-diffusion).
2
- # See more details in LICENSE.
3
-
4
- name: ip2p
5
- channels:
6
- - pytorch
7
- - defaults
8
- dependencies:
9
- - python=3.8.5
10
- - pip=20.3
11
- - cudatoolkit=11.3
12
- - pytorch=1.11.0
13
- - torchvision=0.12.0
14
- - numpy=1.19.2
15
- - pip:
16
- - albumentations==0.4.3
17
- - datasets==2.8.0
18
- - diffusers
19
- - opencv-python==4.1.2.30
20
- - pudb==2019.2
21
- - invisible-watermark
22
- - imageio==2.9.0
23
- - imageio-ffmpeg==0.4.2
24
- - pytorch-lightning==1.4.2
25
- - omegaconf==2.1.1
26
- - test-tube>=0.7.5
27
- - streamlit>=0.73.1
28
- - einops==0.3.0
29
- - torch-fidelity==0.3.0
30
- - transformers==4.19.2
31
- - torchmetrics==0.6.0
32
- - kornia==0.6
33
- - -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
34
- - -e git+https://github.com/openai/CLIP.git@main#egg=clip
35
- - openai
36
- - gradio
37
- - seaborn
38
- - git+https://github.com/crowsonkb/k-diffusion.git
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
imgs/dataset.jpg DELETED
Binary file (100 kB)
 
imgs/edit_app.jpg DELETED
Binary file (402 kB)
 
imgs/prompt_app.jpg DELETED
Binary file (70.4 kB)
 
main.py DELETED
@@ -1,799 +0,0 @@
1
- import argparse, os, sys, datetime, glob
2
- import numpy as np
3
- import time
4
- import torch
5
- import torchvision
6
- import pytorch_lightning as pl
7
- import json
8
- import pickle
9
-
10
- from packaging import version
11
- from omegaconf import OmegaConf
12
- from torch.utils.data import DataLoader, Dataset
13
- from functools import partial
14
- from PIL import Image
15
-
16
- import torch.distributed as dist
17
- from pytorch_lightning import seed_everything
18
- from pytorch_lightning.trainer import Trainer
19
- from pytorch_lightning.callbacks import ModelCheckpoint, Callback, LearningRateMonitor
20
- from pytorch_lightning.utilities.distributed import rank_zero_only
21
- from pytorch_lightning.utilities import rank_zero_info
22
- from pytorch_lightning.plugins import DDPPlugin
23
-
24
- sys.path.append("./stable_diffusion")
25
-
26
- from ldm.data.base import Txt2ImgIterableBaseDataset
27
- from ldm.util import instantiate_from_config
28
-
29
-
30
- def get_parser(**parser_kwargs):
31
- def str2bool(v):
32
- if isinstance(v, bool):
33
- return v
34
- if v.lower() in ("yes", "true", "t", "y", "1"):
35
- return True
36
- elif v.lower() in ("no", "false", "f", "n", "0"):
37
- return False
38
- else:
39
- raise argparse.ArgumentTypeError("Boolean value expected.")
40
-
41
- parser = argparse.ArgumentParser(**parser_kwargs)
42
- parser.add_argument(
43
- "-n",
44
- "--name",
45
- type=str,
46
- const=True,
47
- default="",
48
- nargs="?",
49
- help="postfix for logdir",
50
- )
51
- parser.add_argument(
52
- "-r",
53
- "--resume",
54
- type=str,
55
- const=True,
56
- default="",
57
- nargs="?",
58
- help="resume from logdir or checkpoint in logdir",
59
- )
60
- parser.add_argument(
61
- "-b",
62
- "--base",
63
- nargs="*",
64
- metavar="base_config.yaml",
65
- help="paths to base configs. Loaded from left-to-right. "
66
- "Parameters can be overwritten or added with command-line options of the form `--key value`.",
67
- default=list(),
68
- )
69
- parser.add_argument(
70
- "-t",
71
- "--train",
72
- type=str2bool,
73
- const=True,
74
- default=False,
75
- nargs="?",
76
- help="train",
77
- )
78
- parser.add_argument(
79
- "--no-test",
80
- type=str2bool,
81
- const=True,
82
- default=False,
83
- nargs="?",
84
- help="disable test",
85
- )
86
- parser.add_argument(
87
- "-p",
88
- "--project",
89
- help="name of new or path to existing project"
90
- )
91
- parser.add_argument(
92
- "-d",
93
- "--debug",
94
- type=str2bool,
95
- nargs="?",
96
- const=True,
97
- default=False,
98
- help="enable post-mortem debugging",
99
- )
100
- parser.add_argument(
101
- "-s",
102
- "--seed",
103
- type=int,
104
- default=23,
105
- help="seed for seed_everything",
106
- )
107
- parser.add_argument(
108
- "-f",
109
- "--postfix",
110
- type=str,
111
- default="",
112
- help="post-postfix for default name",
113
- )
114
- parser.add_argument(
115
- "-l",
116
- "--logdir",
117
- type=str,
118
- default="logs",
119
- help="directory for logging dat shit",
120
- )
121
- parser.add_argument(
122
- "--scale_lr",
123
- action="store_true",
124
- default=False,
125
- help="scale base-lr by ngpu * batch_size * n_accumulate",
126
- )
127
- return parser
128
-
129
-
130
- def nondefault_trainer_args(opt):
131
- parser = argparse.ArgumentParser()
132
- parser = Trainer.add_argparse_args(parser)
133
- args = parser.parse_args([])
134
- return sorted(k for k in vars(args) if getattr(opt, k) != getattr(args, k))
135
-
136
-
137
- class WrappedDataset(Dataset):
138
- """Wraps an arbitrary object with __len__ and __getitem__ into a pytorch dataset"""
139
-
140
- def __init__(self, dataset):
141
- self.data = dataset
142
-
143
- def __len__(self):
144
- return len(self.data)
145
-
146
- def __getitem__(self, idx):
147
- return self.data[idx]
148
-
149
-
150
- def worker_init_fn(_):
151
- worker_info = torch.utils.data.get_worker_info()
152
-
153
- dataset = worker_info.dataset
154
- worker_id = worker_info.id
155
-
156
- if isinstance(dataset, Txt2ImgIterableBaseDataset):
157
- split_size = dataset.num_records // worker_info.num_workers
158
- # reset num_records to the true number to retain reliable length information
159
- dataset.sample_ids = dataset.valid_ids[worker_id * split_size:(worker_id + 1) * split_size]
160
- current_id = np.random.choice(len(np.random.get_state()[1]), 1)
161
- return np.random.seed(np.random.get_state()[1][current_id] + worker_id)
162
- else:
163
- return np.random.seed(np.random.get_state()[1][0] + worker_id)
164
-
165
-
166
- class DataModuleFromConfig(pl.LightningDataModule):
167
- def __init__(self, batch_size, train=None, validation=None, test=None, predict=None,
168
- wrap=False, num_workers=None, shuffle_test_loader=False, use_worker_init_fn=False,
169
- shuffle_val_dataloader=False):
170
- super().__init__()
171
- self.batch_size = batch_size
172
- self.dataset_configs = dict()
173
- self.num_workers = num_workers if num_workers is not None else batch_size * 2
174
- self.use_worker_init_fn = use_worker_init_fn
175
- if train is not None:
176
- self.dataset_configs["train"] = train
177
- self.train_dataloader = self._train_dataloader
178
- if validation is not None:
179
- self.dataset_configs["validation"] = validation
180
- self.val_dataloader = partial(self._val_dataloader, shuffle=shuffle_val_dataloader)
181
- if test is not None:
182
- self.dataset_configs["test"] = test
183
- self.test_dataloader = partial(self._test_dataloader, shuffle=shuffle_test_loader)
184
- if predict is not None:
185
- self.dataset_configs["predict"] = predict
186
- self.predict_dataloader = self._predict_dataloader
187
- self.wrap = wrap
188
-
189
- def prepare_data(self):
190
- for data_cfg in self.dataset_configs.values():
191
- instantiate_from_config(data_cfg)
192
-
193
- def setup(self, stage=None):
194
- self.datasets = dict(
195
- (k, instantiate_from_config(self.dataset_configs[k]))
196
- for k in self.dataset_configs)
197
- if self.wrap:
198
- for k in self.datasets:
199
- self.datasets[k] = WrappedDataset(self.datasets[k])
200
-
201
- def _train_dataloader(self):
202
- is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset)
203
- if is_iterable_dataset or self.use_worker_init_fn:
204
- init_fn = worker_init_fn
205
- else:
206
- init_fn = None
207
- return DataLoader(self.datasets["train"], batch_size=self.batch_size,
208
- num_workers=self.num_workers, shuffle=False if is_iterable_dataset else True,
209
- worker_init_fn=init_fn, persistent_workers=True)
210
-
211
- def _val_dataloader(self, shuffle=False):
212
- if isinstance(self.datasets['validation'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn:
213
- init_fn = worker_init_fn
214
- else:
215
- init_fn = None
216
- return DataLoader(self.datasets["validation"],
217
- batch_size=self.batch_size,
218
- num_workers=self.num_workers,
219
- worker_init_fn=init_fn,
220
- shuffle=shuffle, persistent_workers=True)
221
-
222
- def _test_dataloader(self, shuffle=False):
223
- is_iterable_dataset = isinstance(self.datasets['train'], Txt2ImgIterableBaseDataset)
224
- if is_iterable_dataset or self.use_worker_init_fn:
225
- init_fn = worker_init_fn
226
- else:
227
- init_fn = None
228
-
229
- # do not shuffle dataloader for iterable dataset
230
- shuffle = shuffle and (not is_iterable_dataset)
231
-
232
- return DataLoader(self.datasets["test"], batch_size=self.batch_size,
233
- num_workers=self.num_workers, worker_init_fn=init_fn, shuffle=shuffle, persistent_workers=True)
234
-
235
- def _predict_dataloader(self, shuffle=False):
236
- if isinstance(self.datasets['predict'], Txt2ImgIterableBaseDataset) or self.use_worker_init_fn:
237
- init_fn = worker_init_fn
238
- else:
239
- init_fn = None
240
- return DataLoader(self.datasets["predict"], batch_size=self.batch_size,
241
- num_workers=self.num_workers, worker_init_fn=init_fn, persistent_workers=True)
242
-
243
-
244
- class SetupCallback(Callback):
245
- def __init__(self, resume, now, logdir, ckptdir, cfgdir, config, lightning_config):
246
- super().__init__()
247
- self.resume = resume
248
- self.now = now
249
- self.logdir = logdir
250
- self.ckptdir = ckptdir
251
- self.cfgdir = cfgdir
252
- self.config = config
253
- self.lightning_config = lightning_config
254
-
255
- def on_keyboard_interrupt(self, trainer, pl_module):
256
- if trainer.global_rank == 0:
257
- print("Summoning checkpoint.")
258
- ckpt_path = os.path.join(self.ckptdir, "last.ckpt")
259
- trainer.save_checkpoint(ckpt_path)
260
-
261
- def on_pretrain_routine_start(self, trainer, pl_module):
262
- if trainer.global_rank == 0:
263
- # Create logdirs and save configs
264
- # os.makedirs(self.logdir, exist_ok=True)
265
- # os.makedirs(self.ckptdir, exist_ok=True)
266
- # os.makedirs(self.cfgdir, exist_ok=True)
267
-
268
- if "callbacks" in self.lightning_config:
269
- if 'metrics_over_trainsteps_checkpoint' in self.lightning_config['callbacks']:
270
- os.makedirs(os.path.join(self.ckptdir, 'trainstep_checkpoints'), exist_ok=True)
271
- print("Project config")
272
- print(OmegaConf.to_yaml(self.config))
273
- OmegaConf.save(self.config,
274
- os.path.join(self.cfgdir, "{}-project.yaml".format(self.now)))
275
-
276
- print("Lightning config")
277
- print(OmegaConf.to_yaml(self.lightning_config))
278
- OmegaConf.save(OmegaConf.create({"lightning": self.lightning_config}),
279
- os.path.join(self.cfgdir, "{}-lightning.yaml".format(self.now)))
280
-
281
- def get_world_size():
282
- if not dist.is_available():
283
- return 1
284
- if not dist.is_initialized():
285
- return 1
286
- return dist.get_world_size()
287
-
288
- def all_gather(data):
289
- """
290
- Run all_gather on arbitrary picklable data (not necessarily tensors)
291
- Args:
292
- data: any picklable object
293
- Returns:
294
- list[data]: list of data gathered from each rank
295
- """
296
- world_size = get_world_size()
297
- if world_size == 1:
298
- return [data]
299
-
300
- # serialized to a Tensor
301
- origin_size = None
302
- if not isinstance(data, torch.Tensor):
303
- buffer = pickle.dumps(data)
304
- storage = torch.ByteStorage.from_buffer(buffer)
305
- tensor = torch.ByteTensor(storage).to("cuda")
306
- else:
307
- origin_size = data.size()
308
- tensor = data.reshape(-1)
309
-
310
- tensor_type = tensor.dtype
311
-
312
- # obtain Tensor size of each rank
313
- local_size = torch.LongTensor([tensor.numel()]).to("cuda")
314
- size_list = [torch.LongTensor([0]).to("cuda") for _ in range(world_size)]
315
- dist.all_gather(size_list, local_size)
316
- size_list = [int(size.item()) for size in size_list]
317
- max_size = max(size_list)
318
-
319
- # receiving Tensor from all ranks
320
- # we pad the tensor because torch all_gather does not support
321
- # gathering tensors of different shapes
322
- tensor_list = []
323
- for _ in size_list:
324
- tensor_list.append(torch.FloatTensor(size=(max_size,)).cuda().to(tensor_type))
325
- if local_size != max_size:
326
- padding = torch.FloatTensor(size=(max_size - local_size,)).cuda().to(tensor_type)
327
- tensor = torch.cat((tensor, padding), dim=0)
328
- dist.all_gather(tensor_list, tensor)
329
-
330
- data_list = []
331
- for size, tensor in zip(size_list, tensor_list):
332
- if origin_size is None:
333
- buffer = tensor.cpu().numpy().tobytes()[:size]
334
- data_list.append(pickle.loads(buffer))
335
- else:
336
- buffer = tensor[:size]
337
- data_list.append(buffer)
338
-
339
- if origin_size is not None:
340
- new_shape = [-1] + list(origin_size[1:])
341
- resized_list = []
342
- for data in data_list:
343
- # suppose the difference of tensor size exist in first dimension
344
- data = data.reshape(new_shape)
345
- resized_list.append(data)
346
-
347
- return resized_list
348
- else:
349
- return data_list
350
-
351
- class ImageLogger(Callback):
352
- def __init__(self, batch_frequency, max_images, clamp=True, increase_log_steps=True,
353
- rescale=True, disabled=False, log_on_batch_idx=False, log_first_step=False,
354
- log_images_kwargs=None):
355
- super().__init__()
356
- self.rescale = rescale
357
- self.batch_freq = batch_frequency
358
- self.max_images = max_images
359
- self.logger_log_images = {
360
- pl.loggers.TestTubeLogger: self._testtube,
361
- }
362
- self.log_steps = [2 ** n for n in range(6, int(np.log2(self.batch_freq)) + 1)]
363
- if not increase_log_steps:
364
- self.log_steps = [self.batch_freq]
365
- self.clamp = clamp
366
- self.disabled = disabled
367
- self.log_on_batch_idx = log_on_batch_idx
368
- self.log_images_kwargs = log_images_kwargs if log_images_kwargs else {}
369
- self.log_first_step = log_first_step
370
-
371
- @rank_zero_only
372
- def _testtube(self, pl_module, images, batch_idx, split):
373
- for k in images:
374
- grid = torchvision.utils.make_grid(images[k])
375
- grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w
376
-
377
- tag = f"{split}/{k}"
378
- pl_module.logger.experiment.add_image(
379
- tag, grid,
380
- global_step=pl_module.global_step)
381
-
382
- @rank_zero_only
383
- def log_local(self, save_dir, split, images, prompts,
384
- global_step, current_epoch, batch_idx):
385
- root = os.path.join(save_dir, "images", split)
386
- names = {"reals": "before", "inputs": "after", "reconstruction": "before-vq", "samples": "after-gen"}
387
- # print(root)
388
- for k in images:
389
- grid = torchvision.utils.make_grid(images[k], nrow=8)
390
- if self.rescale:
391
- grid = (grid + 1.0) / 2.0 # -1,1 -> 0,1; c,h,w
392
- grid = grid.transpose(0, 1).transpose(1, 2).squeeze(-1)
393
- grid = grid.numpy()
394
- grid = (grid * 255).astype(np.uint8)
395
- filename = "gs-{:06}_e-{:06}_b-{:06}_{}.png".format(
396
- global_step,
397
- current_epoch,
398
- batch_idx,
399
- names[k])
400
- path = os.path.join(root, filename)
401
- os.makedirs(os.path.split(path)[0], exist_ok=True)
402
- # print(path)
403
- Image.fromarray(grid).save(path)
404
-
405
- filename = "gs-{:06}_e-{:06}_b-{:06}_prompt.json".format(
406
- global_step,
407
- current_epoch,
408
- batch_idx)
409
- path = os.path.join(root, filename)
410
- with open(path, "w") as f:
411
- for p in prompts:
412
- f.write(f"{json.dumps(p)}\n")
413
-
414
- def log_img(self, pl_module, batch, batch_idx, split="train"):
415
- check_idx = batch_idx if self.log_on_batch_idx else pl_module.global_step
416
- if (self.check_frequency(check_idx) and # batch_idx % self.batch_freq == 0
417
- hasattr(pl_module, "log_images") and
418
- callable(pl_module.log_images) and
419
- self.max_images > 0) or (split == "val" and batch_idx == 0):
420
- logger = type(pl_module.logger)
421
-
422
- is_train = pl_module.training
423
- if is_train:
424
- pl_module.eval()
425
-
426
- with torch.no_grad():
427
- images = pl_module.log_images(batch, split=split, **self.log_images_kwargs)
428
-
429
- prompts = batch["edit"]["c_crossattn"][:self.max_images]
430
- prompts = [p for ps in all_gather(prompts) for p in ps]
431
-
432
- for k in images:
433
- N = min(images[k].shape[0], self.max_images)
434
- images[k] = images[k][:N]
435
- images[k] = torch.cat(all_gather(images[k][:N]))
436
- if isinstance(images[k], torch.Tensor):
437
- images[k] = images[k].detach().cpu()
438
- if self.clamp:
439
- images[k] = torch.clamp(images[k], -1., 1.)
440
-
441
- self.log_local(pl_module.logger.save_dir, split, images, prompts,
442
- pl_module.global_step, pl_module.current_epoch, batch_idx)
443
-
444
- logger_log_images = self.logger_log_images.get(logger, lambda *args, **kwargs: None)
445
- logger_log_images(pl_module, images, pl_module.global_step, split)
446
-
447
- if is_train:
448
- pl_module.train()
449
-
450
- def check_frequency(self, check_idx):
451
- if ((check_idx % self.batch_freq) == 0 or (check_idx in self.log_steps)) and (
452
- check_idx > 0 or self.log_first_step):
453
- if len(self.log_steps) > 0:
454
- self.log_steps.pop(0)
455
- return True
456
- return False
457
-
458
- def on_train_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):
459
- if not self.disabled and (pl_module.global_step > 0 or self.log_first_step):
460
- self.log_img(pl_module, batch, batch_idx, split="train")
461
-
462
- def on_validation_batch_end(self, trainer, pl_module, outputs, batch, batch_idx, dataloader_idx):
463
- if not self.disabled and pl_module.global_step > 0:
464
- self.log_img(pl_module, batch, batch_idx, split="val")
465
- if hasattr(pl_module, 'calibrate_grad_norm'):
466
- if (pl_module.calibrate_grad_norm and batch_idx % 25 == 0) and batch_idx > 0:
467
- self.log_gradients(trainer, pl_module, batch_idx=batch_idx)
468
-
469
-
470
- class CUDACallback(Callback):
471
- # see https://github.com/SeanNaren/minGPT/blob/master/mingpt/callback.py
472
- def on_train_epoch_start(self, trainer, pl_module):
473
- # Reset the memory use counter
474
- torch.cuda.reset_peak_memory_stats(trainer.root_gpu)
475
- torch.cuda.synchronize(trainer.root_gpu)
476
- self.start_time = time.time()
477
-
478
- def on_train_epoch_end(self, trainer, pl_module, outputs):
479
- torch.cuda.synchronize(trainer.root_gpu)
480
- max_memory = torch.cuda.max_memory_allocated(trainer.root_gpu) / 2 ** 20
481
- epoch_time = time.time() - self.start_time
482
-
483
- try:
484
- max_memory = trainer.training_type_plugin.reduce(max_memory)
485
- epoch_time = trainer.training_type_plugin.reduce(epoch_time)
486
-
487
- rank_zero_info(f"Average Epoch time: {epoch_time:.2f} seconds")
488
- rank_zero_info(f"Average Peak memory {max_memory:.2f}MiB")
489
- except AttributeError:
490
- pass
491
-
492
-
493
- if __name__ == "__main__":
494
- # custom parser to specify config files, train, test and debug mode,
495
- # postfix, resume.
496
- # `--key value` arguments are interpreted as arguments to the trainer.
497
- # `nested.key=value` arguments are interpreted as config parameters.
498
- # configs are merged from left-to-right followed by command line parameters.
499
-
500
- # model:
501
- # base_learning_rate: float
502
- # target: path to lightning module
503
- # params:
504
- # key: value
505
- # data:
506
- # target: main.DataModuleFromConfig
507
- # params:
508
- # batch_size: int
509
- # wrap: bool
510
- # train:
511
- # target: path to train dataset
512
- # params:
513
- # key: value
514
- # validation:
515
- # target: path to validation dataset
516
- # params:
517
- # key: value
518
- # test:
519
- # target: path to test dataset
520
- # params:
521
- # key: value
522
- # lightning: (optional, has sane defaults and can be specified on cmdline)
523
- # trainer:
524
- # additional arguments to trainer
525
- # logger:
526
- # logger to instantiate
527
- # modelcheckpoint:
528
- # modelcheckpoint to instantiate
529
- # callbacks:
530
- # callback1:
531
- # target: importpath
532
- # params:
533
- # key: value
534
-
535
- now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
536
-
537
- # add cwd for convenience and to make classes in this file available when
538
- # running as `python main.py`
539
- # (in particular `main.DataModuleFromConfig`)
540
- sys.path.append(os.getcwd())
541
-
542
- parser = get_parser()
543
- parser = Trainer.add_argparse_args(parser)
544
-
545
- opt, unknown = parser.parse_known_args()
546
-
547
- assert opt.name
548
- cfg_fname = os.path.split(opt.base[0])[-1]
549
- cfg_name = os.path.splitext(cfg_fname)[0]
550
- nowname = f"{cfg_name}_{opt.name}"
551
- logdir = os.path.join(opt.logdir, nowname)
552
- ckpt = os.path.join(logdir, "checkpoints", "last.ckpt")
553
- resume = False
554
-
555
- if os.path.isfile(ckpt):
556
- opt.resume_from_checkpoint = ckpt
557
- base_configs = sorted(glob.glob(os.path.join(logdir, "configs/*.yaml")))
558
- opt.base = base_configs + opt.base
559
- _tmp = logdir.split("/")
560
- nowname = _tmp[-1]
561
- resume = True
562
-
563
- ckptdir = os.path.join(logdir, "checkpoints")
564
- cfgdir = os.path.join(logdir, "configs")
565
-
566
- os.makedirs(logdir, exist_ok=True)
567
- os.makedirs(ckptdir, exist_ok=True)
568
- os.makedirs(cfgdir, exist_ok=True)
569
-
570
- try:
571
- # init and save configs
572
- configs = [OmegaConf.load(cfg) for cfg in opt.base]
573
- cli = OmegaConf.from_dotlist(unknown)
574
- config = OmegaConf.merge(*configs, cli)
575
-
576
- if resume:
577
- # By default, when finetuning from Stable Diffusion, we load the EMA-only checkpoint to initialize all weights.
578
- # If resuming InstructPix2Pix from a finetuning checkpoint, instead load both EMA and non-EMA weights.
579
- config.model.params.load_ema = True
580
-
581
- lightning_config = config.pop("lightning", OmegaConf.create())
582
- # merge trainer cli with config
583
- trainer_config = lightning_config.get("trainer", OmegaConf.create())
584
- # default to ddp
585
- trainer_config["accelerator"] = "ddp"
586
- for k in nondefault_trainer_args(opt):
587
- trainer_config[k] = getattr(opt, k)
588
- if not "gpus" in trainer_config:
589
- del trainer_config["accelerator"]
590
- cpu = True
591
- else:
592
- gpuinfo = trainer_config["gpus"]
593
- print(f"Running on GPUs {gpuinfo}")
594
- cpu = False
595
- trainer_opt = argparse.Namespace(**trainer_config)
596
- lightning_config.trainer = trainer_config
597
-
598
- # model
599
- model = instantiate_from_config(config.model)
600
-
601
- # trainer and callbacks
602
- trainer_kwargs = dict()
603
-
604
- # default logger configs
605
- default_logger_cfgs = {
606
- "wandb": {
607
- "target": "pytorch_lightning.loggers.WandbLogger",
608
- "params": {
609
- "name": nowname,
610
- "save_dir": logdir,
611
- "id": nowname,
612
- }
613
- },
614
- "testtube": {
615
- "target": "pytorch_lightning.loggers.TestTubeLogger",
616
- "params": {
617
- "name": "testtube",
618
- "save_dir": logdir,
619
- }
620
- },
621
- }
622
- default_logger_cfg = default_logger_cfgs["wandb"]
623
- if "logger" in lightning_config:
624
- logger_cfg = lightning_config.logger
625
- else:
626
- logger_cfg = OmegaConf.create()
627
- logger_cfg = OmegaConf.merge(default_logger_cfg, logger_cfg)
628
- trainer_kwargs["logger"] = instantiate_from_config(logger_cfg)
629
-
630
- # modelcheckpoint - use TrainResult/EvalResult(checkpoint_on=metric) to
631
- # specify which metric is used to determine best models
632
- default_modelckpt_cfg = {
633
- "target": "pytorch_lightning.callbacks.ModelCheckpoint",
634
- "params": {
635
- "dirpath": ckptdir,
636
- "filename": "{epoch:06}",
637
- "verbose": True,
638
- "save_last": True,
639
- }
640
- }
641
-
642
- if "modelcheckpoint" in lightning_config:
643
- modelckpt_cfg = lightning_config.modelcheckpoint
644
- else:
645
- modelckpt_cfg = OmegaConf.create()
646
- modelckpt_cfg = OmegaConf.merge(default_modelckpt_cfg, modelckpt_cfg)
647
- print(f"Merged modelckpt-cfg: \n{modelckpt_cfg}")
648
- if version.parse(pl.__version__) < version.parse('1.4.0'):
649
- trainer_kwargs["checkpoint_callback"] = instantiate_from_config(modelckpt_cfg)
650
-
651
- # add callback which sets up log directory
652
- default_callbacks_cfg = {
653
- "setup_callback": {
654
- "target": "main.SetupCallback",
655
- "params": {
656
- "resume": opt.resume,
657
- "now": now,
658
- "logdir": logdir,
659
- "ckptdir": ckptdir,
660
- "cfgdir": cfgdir,
661
- "config": config,
662
- "lightning_config": lightning_config,
663
- }
664
- },
665
- "image_logger": {
666
- "target": "main.ImageLogger",
667
- "params": {
668
- "batch_frequency": 750,
669
- "max_images": 4,
670
- "clamp": True
671
- }
672
- },
673
- "learning_rate_logger": {
674
- "target": "main.LearningRateMonitor",
675
- "params": {
676
- "logging_interval": "step",
677
- # "log_momentum": True
678
- }
679
- },
680
- "cuda_callback": {
681
- "target": "main.CUDACallback"
682
- },
683
- }
684
- if version.parse(pl.__version__) >= version.parse('1.4.0'):
685
- default_callbacks_cfg.update({'checkpoint_callback': modelckpt_cfg})
686
-
687
- if "callbacks" in lightning_config:
688
- callbacks_cfg = lightning_config.callbacks
689
- else:
690
- callbacks_cfg = OmegaConf.create()
691
-
692
- print(
693
- 'Caution: Saving checkpoints every n train steps without deleting. This might require some free space.')
694
- default_metrics_over_trainsteps_ckpt_dict = {
695
- 'metrics_over_trainsteps_checkpoint': {
696
- "target": 'pytorch_lightning.callbacks.ModelCheckpoint',
697
- 'params': {
698
- "dirpath": os.path.join(ckptdir, 'trainstep_checkpoints'),
699
- "filename": "{epoch:06}-{step:09}",
700
- "verbose": True,
701
- 'save_top_k': -1,
702
- 'every_n_train_steps': 1000,
703
- 'save_weights_only': True
704
- }
705
- }
706
- }
707
- default_callbacks_cfg.update(default_metrics_over_trainsteps_ckpt_dict)
708
-
709
- callbacks_cfg = OmegaConf.merge(default_callbacks_cfg, callbacks_cfg)
710
- if 'ignore_keys_callback' in callbacks_cfg and hasattr(trainer_opt, 'resume_from_checkpoint'):
711
- callbacks_cfg.ignore_keys_callback.params['ckpt_path'] = trainer_opt.resume_from_checkpoint
712
- elif 'ignore_keys_callback' in callbacks_cfg:
713
- del callbacks_cfg['ignore_keys_callback']
714
-
715
- trainer_kwargs["callbacks"] = [instantiate_from_config(callbacks_cfg[k]) for k in callbacks_cfg]
716
-
717
- trainer = Trainer.from_argparse_args(trainer_opt, plugins=DDPPlugin(find_unused_parameters=False), **trainer_kwargs)
718
- trainer.logdir = logdir ###
719
-
720
- # data
721
- data = instantiate_from_config(config.data)
722
- # NOTE according to https://pytorch-lightning.readthedocs.io/en/latest/datamodules.html
723
- # calling these ourselves should not be necessary but it is.
724
- # lightning still takes care of proper multiprocessing though
725
- data.prepare_data()
726
- data.setup()
727
- print("#### Data #####")
728
- for k in data.datasets:
729
- print(f"{k}, {data.datasets[k].__class__.__name__}, {len(data.datasets[k])}")
730
-
731
- # configure learning rate
732
- bs, base_lr = config.data.params.batch_size, config.model.base_learning_rate
733
- if not cpu:
734
- ngpu = len(lightning_config.trainer.gpus.strip(",").split(','))
735
- else:
736
- ngpu = 1
737
- if 'accumulate_grad_batches' in lightning_config.trainer:
738
- accumulate_grad_batches = lightning_config.trainer.accumulate_grad_batches
739
- else:
740
- accumulate_grad_batches = 1
741
- print(f"accumulate_grad_batches = {accumulate_grad_batches}")
742
- lightning_config.trainer.accumulate_grad_batches = accumulate_grad_batches
743
- if opt.scale_lr:
744
- model.learning_rate = accumulate_grad_batches * ngpu * bs * base_lr
745
- print(
746
- "Setting learning rate to {:.2e} = {} (accumulate_grad_batches) * {} (num_gpus) * {} (batchsize) * {:.2e} (base_lr)".format(
747
- model.learning_rate, accumulate_grad_batches, ngpu, bs, base_lr))
748
- else:
749
- model.learning_rate = base_lr
750
- print("++++ NOT USING LR SCALING ++++")
751
- print(f"Setting learning rate to {model.learning_rate:.2e}")
752
-
753
-
754
- # allow checkpointing via USR1
755
- def melk(*args, **kwargs):
756
- # run all checkpoint hooks
757
- if trainer.global_rank == 0:
758
- print("Summoning checkpoint.")
759
- ckpt_path = os.path.join(ckptdir, "last.ckpt")
760
- trainer.save_checkpoint(ckpt_path)
761
-
762
-
763
- def divein(*args, **kwargs):
764
- if trainer.global_rank == 0:
765
- import pudb;
766
- pudb.set_trace()
767
-
768
-
769
- import signal
770
-
771
- signal.signal(signal.SIGUSR1, melk)
772
- signal.signal(signal.SIGUSR2, divein)
773
-
774
- # run
775
- if opt.train:
776
- try:
777
- trainer.fit(model, data)
778
- except Exception:
779
- melk()
780
- raise
781
- if not opt.no_test and not trainer.interrupted:
782
- trainer.test(model, data)
783
- except Exception:
784
- if opt.debug and trainer.global_rank == 0:
785
- try:
786
- import pudb as debugger
787
- except ImportError:
788
- import pdb as debugger
789
- debugger.post_mortem()
790
- raise
791
- finally:
792
- # move newly created debug project to debug_runs
793
- if opt.debug and not opt.resume and trainer.global_rank == 0:
794
- dst, name = os.path.split(logdir)
795
- dst = os.path.join(dst, "debug_runs", name)
796
- os.makedirs(os.path.split(dst)[0], exist_ok=True)
797
- os.rename(logdir, dst)
798
- if trainer.global_rank == 0:
799
- print(trainer.profiler.summary())
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
metrics/clip_similarity.py DELETED
@@ -1,47 +0,0 @@
1
- from __future__ import annotations
2
-
3
- import clip
4
- import torch
5
- import torch.nn as nn
6
- import torch.nn.functional as F
7
- from einops import rearrange
8
-
9
-
10
- class ClipSimilarity(nn.Module):
11
- def __init__(self, name: str = "ViT-L/14"):
12
- super().__init__()
13
- assert name in ("RN50", "RN101", "RN50x4", "RN50x16", "RN50x64", "ViT-B/32", "ViT-B/16", "ViT-L/14", "ViT-L/14@336px") # fmt: skip
14
- self.size = {"RN50x4": 288, "RN50x16": 384, "RN50x64": 448, "ViT-L/14@336px": 336}.get(name, 224)
15
-
16
- self.model, _ = clip.load(name, device="cpu", download_root="./")
17
- self.model.eval().requires_grad_(False)
18
-
19
- self.register_buffer("mean", torch.tensor((0.48145466, 0.4578275, 0.40821073)))
20
- self.register_buffer("std", torch.tensor((0.26862954, 0.26130258, 0.27577711)))
21
-
22
- def encode_text(self, text: list[str]) -> torch.Tensor:
23
- text = clip.tokenize(text, truncate=True).to(next(self.parameters()).device)
24
- text_features = self.model.encode_text(text)
25
- text_features = text_features / text_features.norm(dim=1, keepdim=True)
26
- return text_features
27
-
28
- def encode_image(self, image: torch.Tensor) -> torch.Tensor: # Input images in range [0, 1].
29
- image = F.interpolate(image.float(), size=self.size, mode="bicubic", align_corners=False)
30
- image = image - rearrange(self.mean, "c -> 1 c 1 1")
31
- image = image / rearrange(self.std, "c -> 1 c 1 1")
32
- image_features = self.model.encode_image(image)
33
- image_features = image_features / image_features.norm(dim=1, keepdim=True)
34
- return image_features
35
-
36
- def forward(
37
- self, image_0: torch.Tensor, image_1: torch.Tensor, text_0: list[str], text_1: list[str]
38
- ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
39
- image_features_0 = self.encode_image(image_0)
40
- image_features_1 = self.encode_image(image_1)
41
- text_features_0 = self.encode_text(text_0)
42
- text_features_1 = self.encode_text(text_1)
43
- sim_0 = F.cosine_similarity(image_features_0, text_features_0)
44
- sim_1 = F.cosine_similarity(image_features_1, text_features_1)
45
- sim_direction = F.cosine_similarity(image_features_1 - image_features_0, text_features_1 - text_features_0)
46
- sim_image = F.cosine_similarity(image_features_0, image_features_1)
47
- return sim_0, sim_1, sim_direction, sim_image
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
metrics/compute_metrics.py DELETED
@@ -1,235 +0,0 @@
1
- from __future__ import annotations
2
-
3
- import math
4
- import random
5
- import sys
6
- from argparse import ArgumentParser
7
-
8
- import einops
9
- import k_diffusion as K
10
- import numpy as np
11
- import torch
12
- import torch.nn as nn
13
- from tqdm.auto import tqdm
14
- from einops import rearrange
15
- from omegaconf import OmegaConf
16
- from PIL import Image, ImageOps
17
- from torch import autocast
18
-
19
- import json
20
- import matplotlib.pyplot as plt
21
- import seaborn
22
- from pathlib import Path
23
-
24
- sys.path.append("./")
25
-
26
- from clip_similarity import ClipSimilarity
27
- from edit_dataset import EditDatasetEval
28
-
29
- sys.path.append("./stable_diffusion")
30
-
31
- from ldm.util import instantiate_from_config
32
-
33
-
34
- class CFGDenoiser(nn.Module):
35
- def __init__(self, model):
36
- super().__init__()
37
- self.inner_model = model
38
-
39
- def forward(self, z, sigma, cond, uncond, text_cfg_scale, image_cfg_scale):
40
- cfg_z = einops.repeat(z, "1 ... -> n ...", n=3)
41
- cfg_sigma = einops.repeat(sigma, "1 ... -> n ...", n=3)
42
- cfg_cond = {
43
- "c_crossattn": [torch.cat([cond["c_crossattn"][0], uncond["c_crossattn"][0], uncond["c_crossattn"][0]])],
44
- "c_concat": [torch.cat([cond["c_concat"][0], cond["c_concat"][0], uncond["c_concat"][0]])],
45
- }
46
- out_cond, out_img_cond, out_uncond = self.inner_model(cfg_z, cfg_sigma, cond=cfg_cond).chunk(3)
47
- return out_uncond + text_cfg_scale * (out_cond - out_img_cond) + image_cfg_scale * (out_img_cond - out_uncond)
48
-
49
-
50
- def load_model_from_config(config, ckpt, vae_ckpt=None, verbose=False):
51
- print(f"Loading model from {ckpt}")
52
- pl_sd = torch.load(ckpt, map_location="cpu")
53
- if "global_step" in pl_sd:
54
- print(f"Global Step: {pl_sd['global_step']}")
55
- sd = pl_sd["state_dict"]
56
- if vae_ckpt is not None:
57
- print(f"Loading VAE from {vae_ckpt}")
58
- vae_sd = torch.load(vae_ckpt, map_location="cpu")["state_dict"]
59
- sd = {
60
- k: vae_sd[k[len("first_stage_model.") :]] if k.startswith("first_stage_model.") else v
61
- for k, v in sd.items()
62
- }
63
- model = instantiate_from_config(config.model)
64
- m, u = model.load_state_dict(sd, strict=False)
65
- if len(m) > 0 and verbose:
66
- print("missing keys:")
67
- print(m)
68
- if len(u) > 0 and verbose:
69
- print("unexpected keys:")
70
- print(u)
71
- return model
72
-
73
- class ImageEditor(nn.Module):
74
- def __init__(self, config, ckpt, vae_ckpt=None):
75
- super().__init__()
76
-
77
- config = OmegaConf.load(config)
78
- self.model = load_model_from_config(config, ckpt, vae_ckpt)
79
- self.model.eval().cuda()
80
- self.model_wrap = K.external.CompVisDenoiser(self.model)
81
- self.model_wrap_cfg = CFGDenoiser(self.model_wrap)
82
- self.null_token = self.model.get_learned_conditioning([""])
83
-
84
- def forward(
85
- self,
86
- image: torch.Tensor,
87
- edit: str,
88
- scale_txt: float = 7.5,
89
- scale_img: float = 1.0,
90
- steps: int = 100,
91
- ) -> torch.Tensor:
92
- assert image.dim() == 3
93
- assert image.size(1) % 64 == 0
94
- assert image.size(2) % 64 == 0
95
- with torch.no_grad(), autocast("cuda"), self.model.ema_scope():
96
- cond = {
97
- "c_crossattn": [self.model.get_learned_conditioning([edit])],
98
- "c_concat": [self.model.encode_first_stage(image[None]).mode()],
99
- }
100
- uncond = {
101
- "c_crossattn": [self.model.get_learned_conditioning([""])],
102
- "c_concat": [torch.zeros_like(cond["c_concat"][0])],
103
- }
104
- extra_args = {
105
- "uncond": uncond,
106
- "cond": cond,
107
- "image_cfg_scale": scale_img,
108
- "text_cfg_scale": scale_txt,
109
- }
110
- sigmas = self.model_wrap.get_sigmas(steps)
111
- x = torch.randn_like(cond["c_concat"][0]) * sigmas[0]
112
- x = K.sampling.sample_euler_ancestral(self.model_wrap_cfg, x, sigmas, extra_args=extra_args)
113
- x = self.model.decode_first_stage(x)[0]
114
- return x
115
-
116
-
117
- def compute_metrics(config,
118
- model_path,
119
- vae_ckpt,
120
- data_path,
121
- output_path,
122
- scales_img,
123
- scales_txt,
124
- num_samples = 5000,
125
- split = "test",
126
- steps = 50,
127
- res = 512,
128
- seed = 0):
129
- editor = ImageEditor(config, model_path, vae_ckpt).cuda()
130
- clip_similarity = ClipSimilarity().cuda()
131
-
132
-
133
-
134
- outpath = Path(output_path, f"n={num_samples}_p={split}_s={steps}_r={res}_e={seed}.jsonl")
135
- Path(output_path).mkdir(parents=True, exist_ok=True)
136
-
137
- for scale_txt in scales_txt:
138
- for scale_img in scales_img:
139
- dataset = EditDatasetEval(
140
- path=data_path,
141
- split=split,
142
- res=res
143
- )
144
- assert num_samples <= len(dataset)
145
- print(f'Processing t={scale_txt}, i={scale_img}')
146
- torch.manual_seed(seed)
147
- perm = torch.randperm(len(dataset))
148
- count = 0
149
- i = 0
150
-
151
- sim_0_avg = 0
152
- sim_1_avg = 0
153
- sim_direction_avg = 0
154
- sim_image_avg = 0
155
- count = 0
156
-
157
- pbar = tqdm(total=num_samples)
158
- while count < num_samples:
159
-
160
- idx = perm[i].item()
161
- sample = dataset[idx]
162
- i += 1
163
-
164
- gen = editor(sample["image_0"].cuda(), sample["edit"], scale_txt=scale_txt, scale_img=scale_img, steps=steps)
165
-
166
- sim_0, sim_1, sim_direction, sim_image = clip_similarity(
167
- sample["image_0"][None].cuda(), gen[None].cuda(), [sample["input_prompt"]], [sample["output_prompt"]]
168
- )
169
- sim_0_avg += sim_0.item()
170
- sim_1_avg += sim_1.item()
171
- sim_direction_avg += sim_direction.item()
172
- sim_image_avg += sim_image.item()
173
- count += 1
174
- pbar.update(count)
175
- pbar.close()
176
-
177
- sim_0_avg /= count
178
- sim_1_avg /= count
179
- sim_direction_avg /= count
180
- sim_image_avg /= count
181
-
182
- with open(outpath, "a") as f:
183
- f.write(f"{json.dumps(dict(sim_0=sim_0_avg, sim_1=sim_1_avg, sim_direction=sim_direction_avg, sim_image=sim_image_avg, num_samples=num_samples, split=split, scale_txt=scale_txt, scale_img=scale_img, steps=steps, res=res, seed=seed))}\n")
184
- return outpath
185
-
186
- def plot_metrics(metrics_file, output_path):
187
-
188
- with open(metrics_file, 'r') as f:
189
- data = [json.loads(line) for line in f]
190
-
191
- plt.rcParams.update({'font.size': 11.5})
192
- seaborn.set_style("darkgrid")
193
- plt.figure(figsize=(20.5* 0.7, 10.8* 0.7), dpi=200)
194
-
195
- x = [d["sim_direction"] for d in data]
196
- y = [d["sim_image"] for d in data]
197
-
198
- plt.plot(x, y, marker='o', linewidth=2, markersize=4)
199
-
200
- plt.xlabel("CLIP Text-Image Direction Similarity", labelpad=10)
201
- plt.ylabel("CLIP Image Similarity", labelpad=10)
202
-
203
- plt.savefig(Path(output_path) / Path("plot.pdf"), bbox_inches="tight")
204
-
205
- def main():
206
- parser = ArgumentParser()
207
- parser.add_argument("--resolution", default=512, type=int)
208
- parser.add_argument("--steps", default=100, type=int)
209
- parser.add_argument("--config", default="configs/generate.yaml", type=str)
210
- parser.add_argument("--output_path", default="analysis/", type=str)
211
- parser.add_argument("--ckpt", default="checkpoints/instruct-pix2pix-00-22000.ckpt", type=str)
212
- parser.add_argument("--dataset", default="data/clip-filtered-dataset/", type=str)
213
- parser.add_argument("--vae-ckpt", default=None, type=str)
214
- args = parser.parse_args()
215
-
216
- scales_img = [1.0, 1.2, 1.4, 1.6, 1.8, 2.0, 2.2]
217
- scales_txt = [7.5]
218
-
219
- metrics_file = compute_metrics(
220
- args.config,
221
- args.ckpt,
222
- args.vae_ckpt,
223
- args.dataset,
224
- args.output_path,
225
- scales_img,
226
- scales_txt
227
- steps = args.steps
228
- )
229
-
230
- plot_metrics(metrics_file, args.output_path)
231
-
232
-
233
-
234
- if __name__ == "__main__":
235
- main()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
prompt_app.py DELETED
@@ -1,55 +0,0 @@
1
- from __future__ import annotations
2
-
3
- from argparse import ArgumentParser
4
-
5
- import datasets
6
- import gradio as gr
7
- import numpy as np
8
- import openai
9
-
10
- from dataset_creation.generate_txt_dataset import generate
11
-
12
-
13
- def main(openai_model: str):
14
- dataset = datasets.load_dataset("ChristophSchuhmann/improved_aesthetics_6.5plus", split="train")
15
- captions = dataset[np.random.permutation(len(dataset))]["TEXT"]
16
- index = 0
17
-
18
- def click_random():
19
- nonlocal index
20
- output = captions[index]
21
- index = (index + 1) % len(captions)
22
- return output
23
-
24
- def click_generate(input: str):
25
- if input == "":
26
- raise gr.Error("Input caption is missing!")
27
- edit_output = generate(openai_model, input)
28
- if edit_output is None:
29
- return "Failed :(", "Failed :("
30
- return edit_output
31
-
32
- with gr.Blocks(css="footer {visibility: hidden}") as demo:
33
- txt_input = gr.Textbox(lines=3, label="Input Caption", interactive=True, placeholder="Type image caption here...") # fmt: skip
34
- txt_edit = gr.Textbox(lines=1, label="GPT-3 Instruction", interactive=False)
35
- txt_output = gr.Textbox(lines=3, label="GPT3 Edited Caption", interactive=False)
36
-
37
- with gr.Row():
38
- clear_btn = gr.Button("Clear")
39
- random_btn = gr.Button("Random Input")
40
- generate_btn = gr.Button("Generate Instruction + Edited Caption")
41
-
42
- clear_btn.click(fn=lambda: ("", "", ""), inputs=[], outputs=[txt_input, txt_edit, txt_output])
43
- random_btn.click(fn=click_random, inputs=[], outputs=[txt_input])
44
- generate_btn.click(fn=click_generate, inputs=[txt_input], outputs=[txt_edit, txt_output])
45
-
46
- demo.launch(share=True)
47
-
48
-
49
- if __name__ == "__main__":
50
- parser = ArgumentParser()
51
- parser.add_argument("--openai-api-key", required=True, type=str)
52
- parser.add_argument("--openai-model", required=True, type=str)
53
- args = parser.parse_args()
54
- openai.api_key = args.openai_api_key
55
- main(args.openai_model)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
requirements.txt CHANGED
@@ -1,27 +1,6 @@
 
1
  torch
2
- torchvision==0.12.0
3
- numpy==1.19.2
4
- albumentations==0.4.3
5
- datasets==2.8.0
6
- diffusers
7
- opencv-python==4.1.2.30
8
- pudb==2019.2
9
- invisible-watermark
10
- imageio==2.9.0
11
- imageio-ffmpeg==0.4.2
12
- pytorch-lightning==1.4.2
13
- omegaconf==2.1.1
14
- test-tube>=0.7.5
15
- streamlit>=0.73.1
16
- einops==0.3.0
17
- torch-fidelity==0.3.0
18
  transformers
19
- torchmetrics==0.6.0
20
- kornia==0.6
21
- -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers
22
- -e git+https://github.com/openai/CLIP.git@main#egg=clip
23
- huggingface-hub
24
- openai
25
- seaborn
26
- git+https://github.com/crowsonkb/k-diffusion.git
27
  git+https://github.com/huggingface/diffusers
 
1
+ -f --extra-index-url https://download.pytorch.org/whl/cu116
2
  torch
3
+ torchvision
4
+ numpy
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
  transformers
 
 
 
 
 
 
 
 
6
  git+https://github.com/huggingface/diffusers
scripts/download_checkpoints.sh DELETED
@@ -1,7 +0,0 @@
1
- #!/bin/bash
2
-
3
- SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
4
-
5
- mkdir -p $SCRIPT_DIR/../checkpoints
6
-
7
- curl http://instruct-pix2pix.eecs.berkeley.edu/instruct-pix2pix-00-22000.ckpt -o $SCRIPT_DIR/../checkpoints/instruct-pix2pix-00-22000.ckpt
 
 
 
 
 
 
 
 
scripts/download_data.sh DELETED
@@ -1,27 +0,0 @@
1
- #!/bin/bash
2
-
3
- # Make data folder relative to script location
4
- SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
5
-
6
- mkdir -p $SCRIPT_DIR/../data
7
-
8
- # Copy text datasets
9
- wget -q --show-progress http://instruct-pix2pix.eecs.berkeley.edu/gpt-generated-prompts.jsonl -O $SCRIPT_DIR/../data/gpt-generated-prompts.jsonl
10
- wget -q --show-progress http://instruct-pix2pix.eecs.berkeley.edu/human-written-prompts.jsonl -O $SCRIPT_DIR/../data/human-written-prompts.jsonl
11
-
12
- # If dataset name isn't provided, exit.
13
- if [ -z $1 ]
14
- then
15
- exit 0
16
- fi
17
-
18
- # Copy dataset files
19
- mkdir $SCRIPT_DIR/../data/$1
20
- wget -A zip,json -R "index.html*" -q --show-progress -r --no-parent http://instruct-pix2pix.eecs.berkeley.edu/$1/ -nd -P $SCRIPT_DIR/../data/$1/
21
-
22
- # Unzip to folders
23
- unzip $SCRIPT_DIR/../data/$1/\*.zip -d $SCRIPT_DIR/../data/$1/
24
-
25
- # Cleanup
26
- rm -f $SCRIPT_DIR/../data/$1/*.zip
27
- rm -f $SCRIPT_DIR/../data/$1/*.html
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
scripts/download_pretrained_sd.sh DELETED
@@ -1,7 +0,0 @@
1
- #!/bin/bash
2
-
3
- SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
4
-
5
- mkdir -p $SCRIPT_DIR/../stable_diffusion/models/ldm/stable-diffusion-v1
6
- curl -L https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.ckpt -o $SCRIPT_DIR/../stable_diffusion/models/ldm/stable-diffusion-v1/v1-5-pruned-emaonly.ckpt
7
- curl -L https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.ckpt -o $SCRIPT_DIR/../stable_diffusion/models/ldm/stable-diffusion-v1/vae-ft-mse-840000-ema-pruned.ckpt
 
 
 
 
 
 
 
 
stable_diffusion/LICENSE DELETED
@@ -1,82 +0,0 @@
1
- Copyright (c) 2022 Robin Rombach and Patrick Esser and contributors
2
-
3
- CreativeML Open RAIL-M
4
- dated August 22, 2022
5
-
6
- Section I: PREAMBLE
7
-
8
- Multimodal generative models are being widely adopted and used, and have the potential to transform the way artists, among other individuals, conceive and benefit from AI or ML technologies as a tool for content creation.
9
-
10
- Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations.
11
-
12
- In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the Model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for art and content generation.
13
-
14
- Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this License aims to strike a balance between both in order to enable responsible open-science in the field of AI.
15
-
16
- This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model.
17
-
18
- NOW THEREFORE, You and Licensor agree as follows:
19
-
20
- 1. Definitions
21
-
22
- - "License" means the terms and conditions for use, reproduction, and Distribution as defined in this document.
23
- - "Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License.
24
- - "Output" means the results of operating a Model as embodied in informational content resulting therefrom.
25
- - "Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material.
26
- - "Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model.
27
- - "Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any.
28
- - "Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access.
29
- - "Licensor" means the copyright owner or entity authorized by the copyright owner that is granting the License, including the persons or entities that may have rights in the Model and/or distributing the Model.
30
- - "You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, image generator.
31
- - "Third Parties" means individuals or legal entities that are not under common control with Licensor or You.
32
- - "Contribution" means any work of authorship, including the original version of the Model and any modifications or additions to that Model or Derivatives of the Model thereof, that is intentionally submitted to Licensor for inclusion in the Model by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Model, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
33
- - "Contributor" means Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Model.
34
-
35
- Section II: INTELLECTUAL PROPERTY RIGHTS
36
-
37
- Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III.
38
-
39
- 2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model.
40
- 3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Model to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material or a Contribution incorporated within the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or Work shall terminate as of the date such litigation is asserted or filed.
41
-
42
- Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION
43
-
44
- 4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions:
45
- Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material.
46
- You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License;
47
- You must cause any modified files to carry prominent notices stating that You changed the files;
48
- You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model.
49
- You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. - for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License.
50
- 5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5).
51
- 6. The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
52
-
53
- Section IV: OTHER PROVISIONS
54
-
55
- 7. Updates and Runtime Restrictions. To the maximum extent permitted by law, Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License, update the Model through electronic means, or modify the Output of the Model based on updates. You shall undertake reasonable efforts to use the latest version of the Model.
56
- 8. Trademarks and related. Nothing in this License permits You to make use of Licensors’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by the Licensors.
57
- 9. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Model and the Complementary Material (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License.
58
- 10. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
59
- 11. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
60
- 12. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein.
61
-
62
- END OF TERMS AND CONDITIONS
63
-
64
-
65
-
66
-
67
- Attachment A
68
-
69
- Use Restrictions
70
-
71
- You agree not to use the Model or Derivatives of the Model:
72
- - In any way that violates any applicable national, federal, state, local or international law or regulation;
73
- - For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
74
- - To generate or disseminate verifiably false information and/or content with the purpose of harming others;
75
- - To generate or disseminate personal identifiable information that can be used to harm an individual;
76
- - To defame, disparage or otherwise harass others;
77
- - For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation;
78
- - For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics;
79
- - To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
80
- - For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories;
81
- - To provide medical advice and medical results interpretation;
82
- - To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
stable_diffusion/README.md DELETED
@@ -1,215 +0,0 @@
1
- # Stable Diffusion
2
- *Stable Diffusion was made possible thanks to a collaboration with [Stability AI](https://stability.ai/) and [Runway](https://runwayml.com/) and builds upon our previous work:*
3
-
4
- [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://ommer-lab.com/research/latent-diffusion-models/)<br/>
5
- [Robin Rombach](https://github.com/rromb)\*,
6
- [Andreas Blattmann](https://github.com/ablattmann)\*,
7
- [Dominik Lorenz](https://github.com/qp-qp)\,
8
- [Patrick Esser](https://github.com/pesser),
9
- [BjΓΆrn Ommer](https://hci.iwr.uni-heidelberg.de/Staff/bommer)<br/>
10
- _[CVPR '22 Oral](https://openaccess.thecvf.com/content/CVPR2022/html/Rombach_High-Resolution_Image_Synthesis_With_Latent_Diffusion_Models_CVPR_2022_paper.html) |
11
- [GitHub](https://github.com/CompVis/latent-diffusion) | [arXiv](https://arxiv.org/abs/2112.10752) | [Project page](https://ommer-lab.com/research/latent-diffusion-models/)_
12
-
13
- ![txt2img-stable2](assets/stable-samples/txt2img/merged-0006.png)
14
- [Stable Diffusion](#stable-diffusion-v1) is a latent text-to-image diffusion
15
- model.
16
- Thanks to a generous compute donation from [Stability AI](https://stability.ai/) and support from [LAION](https://laion.ai/), we were able to train a Latent Diffusion Model on 512x512 images from a subset of the [LAION-5B](https://laion.ai/blog/laion-5b/) database.
17
- Similar to Google's [Imagen](https://arxiv.org/abs/2205.11487),
18
- this model uses a frozen CLIP ViT-L/14 text encoder to condition the model on text prompts.
19
- With its 860M UNet and 123M text encoder, the model is relatively lightweight and runs on a GPU with at least 10GB VRAM.
20
- See [this section](#stable-diffusion-v1) below and the [model card](https://huggingface.co/CompVis/stable-diffusion).
21
-
22
-
23
- ## Requirements
24
- A suitable [conda](https://conda.io/) environment named `ldm` can be created
25
- and activated with:
26
-
27
- ```
28
- conda env create -f environment.yaml
29
- conda activate ldm
30
- ```
31
-
32
- You can also update an existing [latent diffusion](https://github.com/CompVis/latent-diffusion) environment by running
33
-
34
- ```
35
- conda install pytorch torchvision -c pytorch
36
- pip install transformers==4.19.2 diffusers invisible-watermark
37
- pip install -e .
38
- ```
39
-
40
-
41
- ## Stable Diffusion v1
42
-
43
- Stable Diffusion v1 refers to a specific configuration of the model
44
- architecture that uses a downsampling-factor 8 autoencoder with an 860M UNet
45
- and CLIP ViT-L/14 text encoder for the diffusion model. The model was pretrained on 256x256 images and
46
- then finetuned on 512x512 images.
47
-
48
- *Note: Stable Diffusion v1 is a general text-to-image diffusion model and therefore mirrors biases and (mis-)conceptions that are present
49
- in its training data.
50
- Details on the training procedure and data, as well as the intended use of the model can be found in the corresponding [model card](Stable_Diffusion_v1_Model_Card.md).*
51
-
52
- The weights are available via [the CompVis organization at Hugging Face](https://huggingface.co/CompVis) under [a license which contains specific use-based restrictions to prevent misuse and harm as informed by the model card, but otherwise remains permissive](LICENSE). While commercial use is permitted under the terms of the license, **we do not recommend using the provided weights for services or products without additional safety mechanisms and considerations**, since there are [known limitations and biases](Stable_Diffusion_v1_Model_Card.md#limitations-and-bias) of the weights, and research on safe and ethical deployment of general text-to-image models is an ongoing effort. **The weights are research artifacts and should be treated as such.**
53
-
54
- [The CreativeML OpenRAIL M license](LICENSE) is an [Open RAIL M license](https://www.licenses.ai/blog/2022/8/18/naming-convention-of-responsible-ai-licenses), adapted from the work that [BigScience](https://bigscience.huggingface.co/) and [the RAIL Initiative](https://www.licenses.ai/) are jointly carrying in the area of responsible AI licensing. See also [the article about the BLOOM Open RAIL license](https://bigscience.huggingface.co/blog/the-bigscience-rail-license) on which our license is based.
55
-
56
- ### Weights
57
-
58
- We currently provide the following checkpoints:
59
-
60
- - `sd-v1-1.ckpt`: 237k steps at resolution `256x256` on [laion2B-en](https://huggingface.co/datasets/laion/laion2B-en).
61
- 194k steps at resolution `512x512` on [laion-high-resolution](https://huggingface.co/datasets/laion/laion-high-resolution) (170M examples from LAION-5B with resolution `>= 1024x1024`).
62
- - `sd-v1-2.ckpt`: Resumed from `sd-v1-1.ckpt`.
63
- 515k steps at resolution `512x512` on [laion-aesthetics v2 5+](https://laion.ai/blog/laion-aesthetics/) (a subset of laion2B-en with estimated aesthetics score `> 5.0`, and additionally
64
- filtered to images with an original size `>= 512x512`, and an estimated watermark probability `< 0.5`. The watermark estimate is from the [LAION-5B](https://laion.ai/blog/laion-5b/) metadata, the aesthetics score is estimated using the [LAION-Aesthetics Predictor V2](https://github.com/christophschuhmann/improved-aesthetic-predictor)).
65
- - `sd-v1-3.ckpt`: Resumed from `sd-v1-2.ckpt`. 195k steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
66
- - `sd-v1-4.ckpt`: Resumed from `sd-v1-2.ckpt`. 225k steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
67
-
68
- Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
69
- 5.0, 6.0, 7.0, 8.0) and 50 PLMS sampling
70
- steps show the relative improvements of the checkpoints:
71
- ![sd evaluation results](assets/v1-variants-scores.jpg)
72
-
73
-
74
-
75
- ### Text-to-Image with Stable Diffusion
76
- ![txt2img-stable2](assets/stable-samples/txt2img/merged-0005.png)
77
- ![txt2img-stable2](assets/stable-samples/txt2img/merged-0007.png)
78
-
79
- Stable Diffusion is a latent diffusion model conditioned on the (non-pooled) text embeddings of a CLIP ViT-L/14 text encoder.
80
- We provide a [reference script for sampling](#reference-sampling-script), but
81
- there also exists a [diffusers integration](#diffusers-integration), which we
82
- expect to see more active community development.
83
-
84
- #### Reference Sampling Script
85
-
86
- We provide a reference sampling script, which incorporates
87
-
88
- - a [Safety Checker Module](https://github.com/CompVis/stable-diffusion/pull/36),
89
- to reduce the probability of explicit outputs,
90
- - an [invisible watermarking](https://github.com/ShieldMnt/invisible-watermark)
91
- of the outputs, to help viewers [identify the images as machine-generated](scripts/tests/test_watermark.py).
92
-
93
- After [obtaining the `stable-diffusion-v1-*-original` weights](#weights), link them
94
- ```
95
- mkdir -p models/ldm/stable-diffusion-v1/
96
- ln -s <path/to/model.ckpt> models/ldm/stable-diffusion-v1/model.ckpt
97
- ```
98
- and sample with
99
- ```
100
- python scripts/txt2img.py --prompt "a photograph of an astronaut riding a horse" --plms
101
- ```
102
-
103
- By default, this uses a guidance scale of `--scale 7.5`, [Katherine Crowson's implementation](https://github.com/CompVis/latent-diffusion/pull/51) of the [PLMS](https://arxiv.org/abs/2202.09778) sampler,
104
- and renders images of size 512x512 (which it was trained on) in 50 steps. All supported arguments are listed below (type `python scripts/txt2img.py --help`).
105
-
106
-
107
- ```commandline
108
- usage: txt2img.py [-h] [--prompt [PROMPT]] [--outdir [OUTDIR]] [--skip_grid] [--skip_save] [--ddim_steps DDIM_STEPS] [--plms] [--laion400m] [--fixed_code] [--ddim_eta DDIM_ETA]
109
- [--n_iter N_ITER] [--H H] [--W W] [--C C] [--f F] [--n_samples N_SAMPLES] [--n_rows N_ROWS] [--scale SCALE] [--from-file FROM_FILE] [--config CONFIG] [--ckpt CKPT]
110
- [--seed SEED] [--precision {full,autocast}]
111
-
112
- optional arguments:
113
- -h, --help show this help message and exit
114
- --prompt [PROMPT] the prompt to render
115
- --outdir [OUTDIR] dir to write results to
116
- --skip_grid do not save a grid, only individual samples. Helpful when evaluating lots of samples
117
- --skip_save do not save individual samples. For speed measurements.
118
- --ddim_steps DDIM_STEPS
119
- number of ddim sampling steps
120
- --plms use plms sampling
121
- --laion400m uses the LAION400M model
122
- --fixed_code if enabled, uses the same starting code across samples
123
- --ddim_eta DDIM_ETA ddim eta (eta=0.0 corresponds to deterministic sampling
124
- --n_iter N_ITER sample this often
125
- --H H image height, in pixel space
126
- --W W image width, in pixel space
127
- --C C latent channels
128
- --f F downsampling factor
129
- --n_samples N_SAMPLES
130
- how many samples to produce for each given prompt. A.k.a. batch size
131
- --n_rows N_ROWS rows in the grid (default: n_samples)
132
- --scale SCALE unconditional guidance scale: eps = eps(x, empty) + scale * (eps(x, cond) - eps(x, empty))
133
- --from-file FROM_FILE
134
- if specified, load prompts from this file
135
- --config CONFIG path to config which constructs model
136
- --ckpt CKPT path to checkpoint of model
137
- --seed SEED the seed (for reproducible sampling)
138
- --precision {full,autocast}
139
- evaluate at this precision
140
- ```
141
- Note: The inference config for all v1 versions is designed to be used with EMA-only checkpoints.
142
- For this reason `use_ema=False` is set in the configuration, otherwise the code will try to switch from
143
- non-EMA to EMA weights. If you want to examine the effect of EMA vs no EMA, we provide "full" checkpoints
144
- which contain both types of weights. For these, `use_ema=False` will load and use the non-EMA weights.
145
-
146
-
147
- #### Diffusers Integration
148
-
149
- A simple way to download and sample Stable Diffusion is by using the [diffusers library](https://github.com/huggingface/diffusers/tree/main#new--stable-diffusion-is-now-fully-compatible-with-diffusers):
150
- ```py
151
- # make sure you're logged in with `huggingface-cli login`
152
- from torch import autocast
153
- from diffusers import StableDiffusionPipeline
154
-
155
- pipe = StableDiffusionPipeline.from_pretrained(
156
- "CompVis/stable-diffusion-v1-4",
157
- use_auth_token=True
158
- ).to("cuda")
159
-
160
- prompt = "a photo of an astronaut riding a horse on mars"
161
- with autocast("cuda"):
162
- image = pipe(prompt)["sample"][0]
163
-
164
- image.save("astronaut_rides_horse.png")
165
- ```
166
-
167
-
168
- ### Image Modification with Stable Diffusion
169
-
170
- By using a diffusion-denoising mechanism as first proposed by [SDEdit](https://arxiv.org/abs/2108.01073), the model can be used for different
171
- tasks such as text-guided image-to-image translation and upscaling. Similar to the txt2img sampling script,
172
- we provide a script to perform image modification with Stable Diffusion.
173
-
174
- The following describes an example where a rough sketch made in [Pinta](https://www.pinta-project.com/) is converted into a detailed artwork.
175
- ```
176
- python scripts/img2img.py --prompt "A fantasy landscape, trending on artstation" --init-img <path-to-img.jpg> --strength 0.8
177
- ```
178
- Here, strength is a value between 0.0 and 1.0, that controls the amount of noise that is added to the input image.
179
- Values that approach 1.0 allow for lots of variations but will also produce images that are not semantically consistent with the input. See the following example.
180
-
181
- **Input**
182
-
183
- ![sketch-in](assets/stable-samples/img2img/sketch-mountains-input.jpg)
184
-
185
- **Outputs**
186
-
187
- ![out3](assets/stable-samples/img2img/mountains-3.png)
188
- ![out2](assets/stable-samples/img2img/mountains-2.png)
189
-
190
- This procedure can, for example, also be used to upscale samples from the base model.
191
-
192
-
193
- ## Comments
194
-
195
- - Our codebase for the diffusion models builds heavily on [OpenAI's ADM codebase](https://github.com/openai/guided-diffusion)
196
- and [https://github.com/lucidrains/denoising-diffusion-pytorch](https://github.com/lucidrains/denoising-diffusion-pytorch).
197
- Thanks for open-sourcing!
198
-
199
- - The implementation of the transformer encoder is from [x-transformers](https://github.com/lucidrains/x-transformers) by [lucidrains](https://github.com/lucidrains?tab=repositories).
200
-
201
-
202
- ## BibTeX
203
-
204
- ```
205
- @misc{rombach2021highresolution,
206
- title={High-Resolution Image Synthesis with Latent Diffusion Models},
207
- author={Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and BjΓΆrn Ommer},
208
- year={2021},
209
- eprint={2112.10752},
210
- archivePrefix={arXiv},
211
- primaryClass={cs.CV}
212
- }
213
- ```
214
-
215
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
stable_diffusion/Stable_Diffusion_v1_Model_Card.md DELETED
@@ -1,144 +0,0 @@
1
- # Stable Diffusion v1 Model Card
2
- This model card focuses on the model associated with the Stable Diffusion model, available [here](https://github.com/CompVis/stable-diffusion).
3
-
4
- ## Model Details
5
- - **Developed by:** Robin Rombach, Patrick Esser
6
- - **Model type:** Diffusion-based text-to-image generation model
7
- - **Language(s):** English
8
- - **License:** [Proprietary](LICENSE)
9
- - **Model Description:** This is a model that can be used to generate and modify images based on text prompts. It is a [Latent Diffusion Model](https://arxiv.org/abs/2112.10752) that uses a fixed, pretrained text encoder ([CLIP ViT-L/14](https://arxiv.org/abs/2103.00020)) as suggested in the [Imagen paper](https://arxiv.org/abs/2205.11487).
10
- - **Resources for more information:** [GitHub Repository](https://github.com/CompVis/stable-diffusion), [Paper](https://arxiv.org/abs/2112.10752).
11
- - **Cite as:**
12
-
13
- @InProceedings{Rombach_2022_CVPR,
14
- author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
15
- title = {High-Resolution Image Synthesis With Latent Diffusion Models},
16
- booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
17
- month = {June},
18
- year = {2022},
19
- pages = {10684-10695}
20
- }
21
-
22
- # Uses
23
-
24
- ## Direct Use
25
- The model is intended for research purposes only. Possible research areas and
26
- tasks include
27
-
28
- - Safe deployment of models which have the potential to generate harmful content.
29
- - Probing and understanding the limitations and biases of generative models.
30
- - Generation of artworks and use in design and other artistic processes.
31
- - Applications in educational or creative tools.
32
- - Research on generative models.
33
-
34
- Excluded uses are described below.
35
-
36
- ### Misuse, Malicious Use, and Out-of-Scope Use
37
- _Note: This section is taken from the [DALLE-MINI model card](https://huggingface.co/dalle-mini/dalle-mini), but applies in the same way to Stable Diffusion v1_.
38
-
39
- The model should not be used to intentionally create or disseminate images that create hostile or alienating environments for people. This includes generating images that people would foreseeably find disturbing, distressing, or offensive; or content that propagates historical or current stereotypes.
40
-
41
- #### Out-of-Scope Use
42
- The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.
43
-
44
- #### Misuse and Malicious Use
45
- Using the model to generate content that is cruel to individuals is a misuse of this model. This includes, but is not limited to:
46
-
47
- - Generating demeaning, dehumanizing, or otherwise harmful representations of people or their environments, cultures, religions, etc.
48
- - Intentionally promoting or propagating discriminatory content or harmful stereotypes.
49
- - Impersonating individuals without their consent.
50
- - Sexual content without consent of the people who might see it.
51
- - Mis- and disinformation
52
- - Representations of egregious violence and gore
53
- - Sharing of copyrighted or licensed material in violation of its terms of use.
54
- - Sharing content that is an alteration of copyrighted or licensed material in violation of its terms of use.
55
-
56
- ## Limitations and Bias
57
-
58
- ### Limitations
59
-
60
- - The model does not achieve perfect photorealism
61
- - The model cannot render legible text
62
- - The model does not perform well on more difficult tasks which involve compositionality, such as rendering an image corresponding to β€œA red cube on top of a blue sphere”
63
- - Faces and people in general may not be generated properly.
64
- - The model was trained mainly with English captions and will not work as well in other languages.
65
- - The autoencoding part of the model is lossy
66
- - The model was trained on a large-scale dataset
67
- [LAION-5B](https://laion.ai/blog/laion-5b/) which contains adult material
68
- and is not fit for product use without additional safety mechanisms and
69
- considerations.
70
- - No additional measures were used to deduplicate the dataset. As a result, we observe some degree of memorization for images that are duplicated in the training data.
71
- The training data can be searched at [https://rom1504.github.io/clip-retrieval/](https://rom1504.github.io/clip-retrieval/) to possibly assist in the detection of memorized images.
72
-
73
- ### Bias
74
- While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.
75
- Stable Diffusion v1 was primarily trained on subsets of [LAION-2B(en)](https://laion.ai/blog/laion-5b/),
76
- which consists of images that are limited to English descriptions.
77
- Texts and images from communities and cultures that use other languages are likely to be insufficiently accounted for.
78
- This affects the overall output of the model, as white and western cultures are often set as the default. Further, the
79
- ability of the model to generate content with non-English prompts is significantly worse than with English-language prompts.
80
- Stable Diffusion v1 mirrors and exacerbates biases to such a degree that viewer discretion must be advised irrespective of the input or its intent.
81
-
82
-
83
- ## Training
84
-
85
- **Training Data**
86
- The model developers used the following dataset for training the model:
87
-
88
- - LAION-5B and subsets thereof (see next section)
89
-
90
- **Training Procedure**
91
- Stable Diffusion v1 is a latent diffusion model which combines an autoencoder with a diffusion model that is trained in the latent space of the autoencoder. During training,
92
-
93
- - Images are encoded through an encoder, which turns images into latent representations. The autoencoder uses a relative downsampling factor of 8 and maps images of shape H x W x 3 to latents of shape H/f x W/f x 4
94
- - Text prompts are encoded through a ViT-L/14 text-encoder.
95
- - The non-pooled output of the text encoder is fed into the UNet backbone of the latent diffusion model via cross-attention.
96
- - The loss is a reconstruction objective between the noise that was added to the latent and the prediction made by the UNet.
97
-
98
- We currently provide the following checkpoints:
99
-
100
- - `sd-v1-1.ckpt`: 237k steps at resolution `256x256` on [laion2B-en](https://huggingface.co/datasets/laion/laion2B-en).
101
- 194k steps at resolution `512x512` on [laion-high-resolution](https://huggingface.co/datasets/laion/laion-high-resolution) (170M examples from LAION-5B with resolution `>= 1024x1024`).
102
- - `sd-v1-2.ckpt`: Resumed from `sd-v1-1.ckpt`.
103
- 515k steps at resolution `512x512` on [laion-aesthetics v2 5+](https://laion.ai/blog/laion-aesthetics/) (a subset of laion2B-en with estimated aesthetics score `> 5.0`, and additionally
104
- filtered to images with an original size `>= 512x512`, and an estimated watermark probability `< 0.5`. The watermark estimate is from the [LAION-5B](https://laion.ai/blog/laion-5b/) metadata, the aesthetics score is estimated using the [LAION-Aesthetics Predictor V2](https://github.com/christophschuhmann/improved-aesthetic-predictor)).
105
- - `sd-v1-3.ckpt`: Resumed from `sd-v1-2.ckpt`. 195k steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
106
- - `sd-v1-4.ckpt`: Resumed from `sd-v1-2.ckpt`. 225k steps at resolution `512x512` on "laion-aesthetics v2 5+" and 10\% dropping of the text-conditioning to improve [classifier-free guidance sampling](https://arxiv.org/abs/2207.12598).
107
-
108
- - **Hardware:** 32 x 8 x A100 GPUs
109
- - **Optimizer:** AdamW
110
- - **Gradient Accumulations**: 2
111
- - **Batch:** 32 x 8 x 2 x 4 = 2048
112
- - **Learning rate:** warmup to 0.0001 for 10,000 steps and then kept constant
113
-
114
- ## Evaluation Results
115
- Evaluations with different classifier-free guidance scales (1.5, 2.0, 3.0, 4.0,
116
- 5.0, 6.0, 7.0, 8.0) and 50 PLMS sampling
117
- steps show the relative improvements of the checkpoints:
118
-
119
- ![pareto](assets/v1-variants-scores.jpg)
120
-
121
- Evaluated using 50 PLMS steps and 10000 random prompts from the COCO2017 validation set, evaluated at 512x512 resolution. Not optimized for FID scores.
122
-
123
- ## Environmental Impact
124
-
125
- **Stable Diffusion v1** **Estimated Emissions**
126
- Based on that information, we estimate the following CO2 emissions using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). The hardware, runtime, cloud provider, and compute region were utilized to estimate the carbon impact.
127
-
128
- - **Hardware Type:** A100 PCIe 40GB
129
- - **Hours used:** 150000
130
- - **Cloud Provider:** AWS
131
- - **Compute Region:** US-east
132
- - **Carbon Emitted (Power consumption x Time x Carbon produced based on location of power grid):** 11250 kg CO2 eq.
133
-
134
- ## Citation
135
- @InProceedings{Rombach_2022_CVPR,
136
- author = {Rombach, Robin and Blattmann, Andreas and Lorenz, Dominik and Esser, Patrick and Ommer, Bj\"orn},
137
- title = {High-Resolution Image Synthesis With Latent Diffusion Models},
138
- booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
139
- month = {June},
140
- year = {2022},
141
- pages = {10684-10695}
142
- }
143
-
144
- *This model card was written by: Robin Rombach and Patrick Esser and is based on the [DALL-E Mini model card](https://huggingface.co/dalle-mini/dalle-mini).*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
stable_diffusion/assets/a-painting-of-a-fire.png DELETED
Binary file (667 kB)
 
stable_diffusion/assets/a-photograph-of-a-fire.png DELETED
Binary file (611 kB)
 
stable_diffusion/assets/a-shirt-with-a-fire-printed-on-it.png DELETED
Binary file (624 kB)
 
stable_diffusion/assets/a-shirt-with-the-inscription-'fire'.png DELETED
Binary file (561 kB)
 
stable_diffusion/assets/a-watercolor-painting-of-a-fire.png DELETED
Binary file (722 kB)
 
stable_diffusion/assets/birdhouse.png DELETED
Binary file (775 kB)
 
stable_diffusion/assets/fire.png DELETED
Binary file (626 kB)
 
stable_diffusion/assets/inpainting.png DELETED
Binary file (319 kB)
 
stable_diffusion/assets/modelfigure.png DELETED
Binary file (73.8 kB)
 
stable_diffusion/assets/rdm-preview.jpg DELETED
Binary file (327 kB)
 
stable_diffusion/assets/reconstruction1.png DELETED
Binary file (807 kB)
 
stable_diffusion/assets/reconstruction2.png DELETED
Binary file (981 kB)
 
stable_diffusion/assets/results.gif.REMOVED.git-id DELETED
@@ -1 +0,0 @@
1
- 82b6590e670a32196093cc6333ea19e6547d07de
 
 
stable_diffusion/assets/rick.jpeg DELETED
Binary file (232 kB)
 
stable_diffusion/assets/stable-samples/img2img/mountains-1.png DELETED
Binary file (625 kB)
 
stable_diffusion/assets/stable-samples/img2img/mountains-2.png DELETED
Binary file (658 kB)
 
stable_diffusion/assets/stable-samples/img2img/mountains-3.png DELETED
Binary file (656 kB)
 
stable_diffusion/assets/stable-samples/img2img/sketch-mountains-input.jpg DELETED
Binary file (179 kB)
 
stable_diffusion/assets/stable-samples/img2img/upscaling-in.png.REMOVED.git-id DELETED
@@ -1 +0,0 @@
1
- 501c31c21751664957e69ce52cad1818b6d2f4ce
 
 
stable_diffusion/assets/stable-samples/img2img/upscaling-out.png.REMOVED.git-id DELETED
@@ -1 +0,0 @@
1
- 1c4bb25a779f34d86b2d90e584ac67af91bb1303
 
 
stable_diffusion/assets/stable-samples/txt2img/000002025.png DELETED
Binary file (968 kB)
 
stable_diffusion/assets/stable-samples/txt2img/000002035.png DELETED
Binary file (996 kB)
 
stable_diffusion/assets/stable-samples/txt2img/merged-0005.png.REMOVED.git-id DELETED
@@ -1 +0,0 @@
1
- ca0a1af206555f0f208a1ab879e95efedc1b1c5b
 
 
stable_diffusion/assets/stable-samples/txt2img/merged-0006.png.REMOVED.git-id DELETED
@@ -1 +0,0 @@
1
- 999f3703230580e8c89e9081abd6a1f8f50896d4
 
 
stable_diffusion/assets/stable-samples/txt2img/merged-0007.png.REMOVED.git-id DELETED
@@ -1 +0,0 @@
1
- af390acaf601283782d6f479d4cade4d78e30b26
 
 
stable_diffusion/assets/the-earth-is-on-fire,-oil-on-canvas.png DELETED
Binary file (678 kB)
 
stable_diffusion/assets/txt2img-convsample.png DELETED
Binary file (310 kB)