File size: 10,869 Bytes
21c4e64
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
from transformers import CLIPTextModel, CLIPTokenizer, logging
from diffusers import (
    AutoencoderKL,
    UNet2DConditionModel,
    PNDMScheduler,
    DDIMScheduler,
    StableDiffusionPipeline,
)
from diffusers.utils.import_utils import is_xformers_available

# suppress partial model loading warning
logging.set_verbosity_error()

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F


def seed_everything(seed):
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    # torch.backends.cudnn.deterministic = True
    # torch.backends.cudnn.benchmark = True


class StableDiffusion(nn.Module):
    def __init__(
        self,
        device,
        fp16=True,
        vram_O=False,
        sd_version="2.1",
        hf_key=None,
        t_range=[0.02, 0.98],
    ):
        super().__init__()

        self.device = device
        self.sd_version = sd_version

        if hf_key is not None:
            print(f"[INFO] using hugging face custom model key: {hf_key}")
            model_key = hf_key
        elif self.sd_version == "2.1":
            model_key = "stabilityai/stable-diffusion-2-1-base"
        elif self.sd_version == "2.0":
            model_key = "stabilityai/stable-diffusion-2-base"
        elif self.sd_version == "1.5":
            model_key = "runwayml/stable-diffusion-v1-5"
        else:
            raise ValueError(
                f"Stable-diffusion version {self.sd_version} not supported."
            )

        self.dtype = torch.float16 if fp16 else torch.float32

        # Create model
        pipe = StableDiffusionPipeline.from_pretrained(
            model_key, torch_dtype=self.dtype
        )

        if vram_O:
            pipe.enable_sequential_cpu_offload()
            pipe.enable_vae_slicing()
            pipe.unet.to(memory_format=torch.channels_last)
            pipe.enable_attention_slicing(1)
            # pipe.enable_model_cpu_offload()
        else:
            pipe.to(device)

        self.vae = pipe.vae
        self.tokenizer = pipe.tokenizer
        self.text_encoder = pipe.text_encoder
        self.unet = pipe.unet

        self.scheduler = DDIMScheduler.from_pretrained(
            model_key, subfolder="scheduler", torch_dtype=self.dtype
        )

        del pipe

        self.num_train_timesteps = self.scheduler.config.num_train_timesteps
        self.min_step = int(self.num_train_timesteps * t_range[0])
        self.max_step = int(self.num_train_timesteps * t_range[1])
        self.alphas = self.scheduler.alphas_cumprod.to(self.device)  # for convenience

        self.embeddings = None

    @torch.no_grad()
    def get_text_embeds(self, prompts, negative_prompts):
        pos_embeds = self.encode_text(prompts)  # [1, 77, 768]
        neg_embeds = self.encode_text(negative_prompts)
        self.embeddings = torch.cat([neg_embeds, pos_embeds], dim=0)  # [2, 77, 768]
    
    def encode_text(self, prompt):
        # prompt: [str]
        inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            return_tensors="pt",
        )
        embeddings = self.text_encoder(inputs.input_ids.to(self.device))[0]
        return embeddings

    @torch.no_grad()
    def refine(self, pred_rgb,
               guidance_scale=100, steps=50, strength=0.8,
        ):

        batch_size = pred_rgb.shape[0]
        pred_rgb_512 = F.interpolate(pred_rgb, (512, 512), mode='bilinear', align_corners=False)
        latents = self.encode_imgs(pred_rgb_512.to(self.dtype))
        # latents = torch.randn((1, 4, 64, 64), device=self.device, dtype=self.dtype)

        self.scheduler.set_timesteps(steps)
        init_step = int(steps * strength)
        latents = self.scheduler.add_noise(latents, torch.randn_like(latents), self.scheduler.timesteps[init_step])

        for i, t in enumerate(self.scheduler.timesteps[init_step:]):
    
            latent_model_input = torch.cat([latents] * 2)

            noise_pred = self.unet(
                latent_model_input, t, encoder_hidden_states=self.embeddings,
            ).sample

            noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
            noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond)
            
            latents = self.scheduler.step(noise_pred, t, latents).prev_sample

        imgs = self.decode_latents(latents) # [1, 3, 512, 512]
        return imgs

    def train_step(
        self,
        pred_rgb,
        step_ratio=None,
        guidance_scale=100,
        as_latent=False,
    ):
        
        batch_size = pred_rgb.shape[0]
        pred_rgb = pred_rgb.to(self.dtype)

        if as_latent:
            latents = F.interpolate(pred_rgb, (64, 64), mode="bilinear", align_corners=False) * 2 - 1
        else:
            # interp to 512x512 to be fed into vae.
            pred_rgb_512 = F.interpolate(pred_rgb, (512, 512), mode="bilinear", align_corners=False)
            # encode image into latents with vae, requires grad!
            latents = self.encode_imgs(pred_rgb_512)

        if step_ratio is not None:
            # dreamtime-like
            # t = self.max_step - (self.max_step - self.min_step) * np.sqrt(step_ratio)
            t = np.round((1 - step_ratio) * self.num_train_timesteps).clip(self.min_step, self.max_step)
            t = torch.full((batch_size,), t, dtype=torch.long, device=self.device)
        else:
            t = torch.randint(self.min_step, self.max_step + 1, (batch_size,), dtype=torch.long, device=self.device)

        # w(t), sigma_t^2
        w = (1 - self.alphas[t]).view(batch_size, 1, 1, 1)

        # predict the noise residual with unet, NO grad!
        with torch.no_grad():
            # add noise
            noise = torch.randn_like(latents)
            latents_noisy = self.scheduler.add_noise(latents, noise, t)
            # pred noise
            latent_model_input = torch.cat([latents_noisy] * 2)
            tt = torch.cat([t] * 2)

            noise_pred = self.unet(
                latent_model_input, tt, encoder_hidden_states=self.embeddings.repeat(batch_size, 1, 1)
            ).sample

            # perform guidance (high scale from paper!)
            noise_pred_uncond, noise_pred_pos = noise_pred.chunk(2)
            noise_pred = noise_pred_uncond + guidance_scale * (
                noise_pred_pos - noise_pred_uncond
            )

        grad = w * (noise_pred - noise)
        grad = torch.nan_to_num(grad)

        # seems important to avoid NaN...
        # grad = grad.clamp(-1, 1)

        target = (latents - grad).detach()
        loss = 0.5 * F.mse_loss(latents.float(), target, reduction='sum') / latents.shape[0]

        return loss

    @torch.no_grad()
    def produce_latents(
        self,
        height=512,
        width=512,
        num_inference_steps=50,
        guidance_scale=7.5,
        latents=None,
    ):
        if latents is None:
            latents = torch.randn(
                (
                    self.embeddings.shape[0] // 2,
                    self.unet.in_channels,
                    height // 8,
                    width // 8,
                ),
                device=self.device,
            )

        self.scheduler.set_timesteps(num_inference_steps)

        for i, t in enumerate(self.scheduler.timesteps):
            # expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
            latent_model_input = torch.cat([latents] * 2)
            # predict the noise residual
            noise_pred = self.unet(
                latent_model_input, t, encoder_hidden_states=self.embeddings
            ).sample

            # perform guidance
            noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
            noise_pred = noise_pred_uncond + guidance_scale * (
                noise_pred_cond - noise_pred_uncond
            )

            # compute the previous noisy sample x_t -> x_t-1
            latents = self.scheduler.step(noise_pred, t, latents).prev_sample

        return latents

    def decode_latents(self, latents):
        latents = 1 / self.vae.config.scaling_factor * latents

        imgs = self.vae.decode(latents).sample
        imgs = (imgs / 2 + 0.5).clamp(0, 1)

        return imgs

    def encode_imgs(self, imgs):
        # imgs: [B, 3, H, W]

        imgs = 2 * imgs - 1

        posterior = self.vae.encode(imgs).latent_dist
        latents = posterior.sample() * self.vae.config.scaling_factor

        return latents

    def prompt_to_img(
        self,
        prompts,
        negative_prompts="",
        height=512,
        width=512,
        num_inference_steps=50,
        guidance_scale=7.5,
        latents=None,
    ):
        if isinstance(prompts, str):
            prompts = [prompts]

        if isinstance(negative_prompts, str):
            negative_prompts = [negative_prompts]

        # Prompts -> text embeds
        self.get_text_embeds(prompts, negative_prompts)
        
        # Text embeds -> img latents
        latents = self.produce_latents(
            height=height,
            width=width,
            latents=latents,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
        )  # [1, 4, 64, 64]

        # Img latents -> imgs
        imgs = self.decode_latents(latents)  # [1, 3, 512, 512]

        # Img to Numpy
        imgs = imgs.detach().cpu().permute(0, 2, 3, 1).numpy()
        imgs = (imgs * 255).round().astype("uint8")

        return imgs


if __name__ == "__main__":
    import argparse
    import matplotlib.pyplot as plt

    parser = argparse.ArgumentParser()
    parser.add_argument("prompt", type=str)
    parser.add_argument("--negative", default="", type=str)
    parser.add_argument(
        "--sd_version",
        type=str,
        default="2.1",
        choices=["1.5", "2.0", "2.1"],
        help="stable diffusion version",
    )
    parser.add_argument(
        "--hf_key",
        type=str,
        default=None,
        help="hugging face Stable diffusion model key",
    )
    parser.add_argument("--fp16", action="store_true", help="use float16 for training")
    parser.add_argument(
        "--vram_O", action="store_true", help="optimization for low VRAM usage"
    )
    parser.add_argument("-H", type=int, default=512)
    parser.add_argument("-W", type=int, default=512)
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument("--steps", type=int, default=50)
    opt = parser.parse_args()

    seed_everything(opt.seed)

    device = torch.device("cuda")

    sd = StableDiffusion(device, opt.fp16, opt.vram_O, opt.sd_version, opt.hf_key)

    imgs = sd.prompt_to_img(opt.prompt, opt.negative, opt.H, opt.W, opt.steps)

    # visualize image
    plt.imshow(imgs[0])
    plt.show()