jadechoghari
commited on
Commit
•
f5bb4af
1
Parent(s):
ecbaf30
Create generate.py
Browse files- generate.py +375 -0
generate.py
ADDED
@@ -0,0 +1,375 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import torch
|
3 |
+
import numpy as np
|
4 |
+
from tqdm import tqdm
|
5 |
+
import os
|
6 |
+
from transformers import logging
|
7 |
+
|
8 |
+
from utils import CONTROLNET_DICT
|
9 |
+
from utils import load_config, save_config
|
10 |
+
from utils import get_controlnet_kwargs, get_frame_ids, get_latents_dir, init_model, seed_everything
|
11 |
+
from utils import prepare_control, load_latent, load_video, prepare_depth, save_video
|
12 |
+
from utils import register_time, register_attention_control, register_conv_control
|
13 |
+
|
14 |
+
import vidtome
|
15 |
+
|
16 |
+
# suppress partial model loading warning
|
17 |
+
logging.set_verbosity_error()
|
18 |
+
|
19 |
+
|
20 |
+
class Generator(nn.Module):
|
21 |
+
def __init__(self, pipe, scheduler, config):
|
22 |
+
super().__init__()
|
23 |
+
|
24 |
+
self.device = config.device
|
25 |
+
self.seed = config.seed
|
26 |
+
|
27 |
+
|
28 |
+
|
29 |
+
|
30 |
+
self.model_key = config.model_key
|
31 |
+
|
32 |
+
self.config = config
|
33 |
+
gene_config = config.generation
|
34 |
+
float_precision = gene_config.float_precision if "float_precision" in gene_config else config.float_precision
|
35 |
+
if float_precision == "fp16":
|
36 |
+
self.dtype = torch.float16
|
37 |
+
print("[INFO] float precision fp16. Use torch.float16.")
|
38 |
+
else:
|
39 |
+
self.dtype = torch.float32
|
40 |
+
print("[INFO] float precision fp32. Use torch.float32.")
|
41 |
+
|
42 |
+
self.pipe = pipe
|
43 |
+
self.vae = pipe.vae
|
44 |
+
self.tokenizer = pipe.tokenizer
|
45 |
+
self.unet = pipe.unet
|
46 |
+
self.text_encoder = pipe.text_encoder
|
47 |
+
if config.enable_xformers_memory_efficient_attention:
|
48 |
+
try:
|
49 |
+
pipe.enable_xformers_memory_efficient_attention()
|
50 |
+
except ModuleNotFoundError:
|
51 |
+
print("[WARNING] xformers not found. Disable xformers attention.")
|
52 |
+
self.n_timesteps = gene_config.n_timesteps
|
53 |
+
scheduler.set_timesteps(gene_config.n_timesteps, device=self.device)
|
54 |
+
self.scheduler = scheduler
|
55 |
+
|
56 |
+
self.batch_size = 2
|
57 |
+
self.control = gene_config.control
|
58 |
+
self.use_depth = config.sd_version == "depth"
|
59 |
+
self.use_controlnet = self.control in CONTROLNET_DICT.keys()
|
60 |
+
self.use_pnp = self.control == "pnp"
|
61 |
+
if self.use_controlnet:
|
62 |
+
self.controlnet = pipe.controlnet
|
63 |
+
self.controlnet_scale = gene_config.control_scale
|
64 |
+
elif self.use_pnp:
|
65 |
+
pnp_f_t = int(gene_config.n_timesteps * gene_config.pnp_f_t)
|
66 |
+
pnp_attn_t = int(gene_config.n_timesteps * gene_config.pnp_attn_t)
|
67 |
+
self.batch_size += 1
|
68 |
+
self.init_pnp(conv_injection_t=pnp_f_t, qk_injection_t=pnp_attn_t)
|
69 |
+
|
70 |
+
self.chunk_size = gene_config.chunk_size
|
71 |
+
self.chunk_ord = gene_config.chunk_ord
|
72 |
+
self.merge_global = gene_config.merge_global
|
73 |
+
self.local_merge_ratio = gene_config.local_merge_ratio
|
74 |
+
self.global_merge_ratio = gene_config.global_merge_ratio
|
75 |
+
self.global_rand = gene_config.global_rand
|
76 |
+
self.align_batch = gene_config.align_batch
|
77 |
+
|
78 |
+
self.prompt = gene_config.prompt
|
79 |
+
self.negative_prompt = gene_config.negative_prompt
|
80 |
+
self.guidance_scale = gene_config.guidance_scale
|
81 |
+
self.save_frame = gene_config.save_frame
|
82 |
+
|
83 |
+
self.frame_height, self.frame_width = config.height, config.width
|
84 |
+
self.work_dir = config.work_dir
|
85 |
+
|
86 |
+
self.chunk_ord = gene_config.chunk_ord
|
87 |
+
if "mix" in self.chunk_ord:
|
88 |
+
self.perm_div = float(self.chunk_ord.split("-")[-1]) if "-" in self.chunk_ord else 3.
|
89 |
+
self.chunk_ord = "mix"
|
90 |
+
# Patch VidToMe to model
|
91 |
+
self.activate_vidtome()
|
92 |
+
|
93 |
+
if gene_config.use_lora:
|
94 |
+
self.pipe.load_lora_weights(**gene_config.lora)
|
95 |
+
|
96 |
+
def activate_vidtome(self):
|
97 |
+
vidtome.apply_patch(self.pipe, self.local_merge_ratio, self.merge_global, self.global_merge_ratio,
|
98 |
+
seed = self.seed, batch_size = self.batch_size, align_batch = self.use_pnp or self.align_batch, global_rand = self.global_rand)
|
99 |
+
|
100 |
+
@torch.no_grad()
|
101 |
+
def get_text_embeds_input(self, prompt, negative_prompt):
|
102 |
+
text_embeds = self.get_text_embeds(
|
103 |
+
prompt, negative_prompt, self.device)
|
104 |
+
if self.use_pnp:
|
105 |
+
pnp_guidance_embeds = self.get_text_embeds("", device=self.device)
|
106 |
+
text_embeds = torch.cat(
|
107 |
+
[pnp_guidance_embeds, text_embeds], dim=0)
|
108 |
+
return text_embeds
|
109 |
+
|
110 |
+
@torch.no_grad()
|
111 |
+
def get_text_embeds(self, prompt, negative_prompt=None, device="cuda"):
|
112 |
+
text_input = self.tokenizer(prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
|
113 |
+
truncation=True, return_tensors='pt')
|
114 |
+
text_embeddings = self.text_encoder(text_input.input_ids.to(device))[0]
|
115 |
+
if negative_prompt is not None:
|
116 |
+
uncond_input = self.tokenizer(negative_prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
|
117 |
+
return_tensors='pt')
|
118 |
+
uncond_embeddings = self.text_encoder(
|
119 |
+
uncond_input.input_ids.to(device))[0]
|
120 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
121 |
+
return text_embeddings
|
122 |
+
|
123 |
+
@torch.no_grad()
|
124 |
+
def prepare_data(self, data_path, latent_path, frame_ids):
|
125 |
+
self.frames = load_video(data_path, self.frame_height,
|
126 |
+
self.frame_width, frame_ids=frame_ids, device=self.device)
|
127 |
+
self.init_noise = load_latent(
|
128 |
+
latent_path, t=self.scheduler.timesteps[0], frame_ids=frame_ids).to(self.dtype).to(self.device)
|
129 |
+
|
130 |
+
if self.use_depth:
|
131 |
+
self.depths = prepare_depth(
|
132 |
+
self.pipe, self.frames, frame_ids, self.work_dir).to(self.init_noise)
|
133 |
+
|
134 |
+
if self.use_controlnet:
|
135 |
+
self.controlnet_images = prepare_control(
|
136 |
+
self.control, self.frames, frame_ids, self.work_dir).to(self.init_noise)
|
137 |
+
|
138 |
+
@torch.no_grad()
|
139 |
+
def decode_latents(self, latents):
|
140 |
+
with torch.autocast(device_type=self.device, dtype=self.dtype):
|
141 |
+
latents = 1 / 0.18215 * latents
|
142 |
+
imgs = self.vae.decode(latents).sample
|
143 |
+
imgs = (imgs / 2 + 0.5).clamp(0, 1)
|
144 |
+
return imgs
|
145 |
+
|
146 |
+
@torch.no_grad()
|
147 |
+
def decode_latents_batch(self, latents):
|
148 |
+
imgs = []
|
149 |
+
batch_latents = latents.split(self.batch_size, dim=0)
|
150 |
+
for latent in batch_latents:
|
151 |
+
imgs += [self.decode_latents(latent)]
|
152 |
+
imgs = torch.cat(imgs)
|
153 |
+
return imgs
|
154 |
+
|
155 |
+
@torch.no_grad()
|
156 |
+
def encode_imgs(self, imgs):
|
157 |
+
with torch.autocast(device_type=self.device, dtype=self.dtype):
|
158 |
+
imgs = 2 * imgs - 1
|
159 |
+
posterior = self.vae.encode(imgs).latent_dist
|
160 |
+
latents = posterior.mean * 0.18215
|
161 |
+
return latents
|
162 |
+
|
163 |
+
@torch.no_grad()
|
164 |
+
def encode_imgs_batch(self, imgs):
|
165 |
+
latents = []
|
166 |
+
batch_imgs = imgs.split(self.batch_size, dim=0)
|
167 |
+
for img in batch_imgs:
|
168 |
+
latents += [self.encode_imgs(img)]
|
169 |
+
latents = torch.cat(latents)
|
170 |
+
return latents
|
171 |
+
|
172 |
+
def get_chunks(self, flen):
|
173 |
+
x_index = torch.arange(flen)
|
174 |
+
|
175 |
+
# The first chunk has a random length
|
176 |
+
rand_first = np.random.randint(0, self.chunk_size) + 1
|
177 |
+
chunks = x_index[rand_first:].split(self.chunk_size, dim=0)
|
178 |
+
chunks = [x_index[:rand_first]] + list(chunks) if len(chunks[0]) > 0 else [x_index[:rand_first]]
|
179 |
+
if np.random.rand() > 0.5:
|
180 |
+
chunks = chunks[::-1]
|
181 |
+
|
182 |
+
# Chunk order only matter when we do global token merging
|
183 |
+
if self.merge_global == False:
|
184 |
+
return chunks
|
185 |
+
|
186 |
+
# Chunk order. "seq": sequential order. "rand": full permutation. "mix": partial permutation.
|
187 |
+
if self.chunk_ord == "rand":
|
188 |
+
order = torch.randperm(len(chunks))
|
189 |
+
elif self.chunk_ord == "mix":
|
190 |
+
randord = torch.randperm(len(chunks)).tolist()
|
191 |
+
rand_len = int(len(randord) / self.perm_div)
|
192 |
+
seqord = sorted(randord[rand_len:])
|
193 |
+
if rand_len > 0:
|
194 |
+
randord = randord[:rand_len]
|
195 |
+
if abs(seqord[-1] - randord[-1]) < abs(seqord[0] - randord[-1]):
|
196 |
+
seqord = seqord[::-1]
|
197 |
+
order = randord + seqord
|
198 |
+
else:
|
199 |
+
order = seqord
|
200 |
+
else:
|
201 |
+
order = torch.arange(len(chunks))
|
202 |
+
chunks = [chunks[i] for i in order]
|
203 |
+
return chunks
|
204 |
+
|
205 |
+
@torch.no_grad()
|
206 |
+
def ddim_sample(self, x, conds):
|
207 |
+
print("[INFO] denoising frames...")
|
208 |
+
timesteps = self.scheduler.timesteps
|
209 |
+
noises = torch.zeros_like(x)
|
210 |
+
|
211 |
+
for i, t in enumerate(tqdm(timesteps, desc="Sampling")):
|
212 |
+
self.pre_iter(x, t)
|
213 |
+
|
214 |
+
# Split video into chunks and denoise
|
215 |
+
chunks = self.get_chunks(len(x))
|
216 |
+
for chunk in chunks:
|
217 |
+
torch.cuda.empty_cache()
|
218 |
+
noises[chunk] = self.pred_noise(
|
219 |
+
x[chunk], conds, t, batch_idx=chunk)
|
220 |
+
|
221 |
+
x = self.pred_next_x(x, noises, t, i, inversion=False)
|
222 |
+
|
223 |
+
self.post_iter(x, t)
|
224 |
+
return x
|
225 |
+
|
226 |
+
def pre_iter(self, x, t):
|
227 |
+
if self.use_pnp:
|
228 |
+
# Prepare PnP
|
229 |
+
register_time(self, t.item())
|
230 |
+
cur_latents = load_latent(self.latent_path, t=t, frame_ids = self.frame_ids)
|
231 |
+
self.cur_latents = cur_latents
|
232 |
+
|
233 |
+
def post_iter(self, x, t):
|
234 |
+
if self.merge_global:
|
235 |
+
# Reset global tokens
|
236 |
+
vidtome.update_patch(self.pipe, global_tokens = None)
|
237 |
+
|
238 |
+
@torch.no_grad()
|
239 |
+
def pred_noise(self, x, cond, t, batch_idx=None):
|
240 |
+
|
241 |
+
flen = len(x)
|
242 |
+
text_embed_input = cond.repeat_interleave(flen, dim=0)
|
243 |
+
|
244 |
+
# For classifier-free guidance
|
245 |
+
latent_model_input = torch.cat([x, x])
|
246 |
+
batch_size = 2
|
247 |
+
|
248 |
+
if self.use_pnp:
|
249 |
+
# Cat latents from inverted source frames for PnP operation
|
250 |
+
source_latents = self.cur_latents
|
251 |
+
if batch_idx is not None:
|
252 |
+
source_latents = source_latents[batch_idx]
|
253 |
+
latent_model_input = torch.cat([source_latents.to(x), latent_model_input])
|
254 |
+
batch_size += 1
|
255 |
+
|
256 |
+
# For sd-depth model
|
257 |
+
if self.use_depth:
|
258 |
+
depth = self.depths
|
259 |
+
if batch_idx is not None:
|
260 |
+
depth = depth[batch_idx]
|
261 |
+
depth = depth.repeat(batch_size, 1, 1, 1)
|
262 |
+
latent_model_input = torch.cat([latent_model_input, depth.to(x)], dim=1)
|
263 |
+
|
264 |
+
kwargs = dict()
|
265 |
+
# Compute controlnet outputs
|
266 |
+
if self.use_controlnet:
|
267 |
+
controlnet_cond = self.controlnet_images
|
268 |
+
if batch_idx is not None:
|
269 |
+
controlnet_cond = controlnet_cond[batch_idx]
|
270 |
+
controlnet_cond = controlnet_cond.repeat(batch_size, 1, 1, 1)
|
271 |
+
controlnet_kwargs = get_controlnet_kwargs(
|
272 |
+
self.controlnet, latent_model_input, text_embed_input, t, controlnet_cond, self.controlnet_scale)
|
273 |
+
kwargs.update(controlnet_kwargs)
|
274 |
+
# Pred noise!
|
275 |
+
eps = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input, **kwargs).sample
|
276 |
+
noise_pred_uncond, noise_pred_cond = eps.chunk(batch_size)[-2:]
|
277 |
+
# CFG
|
278 |
+
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
|
279 |
+
return noise_pred
|
280 |
+
|
281 |
+
@torch.no_grad()
|
282 |
+
def pred_next_x(self, x, eps, t, i, inversion=False):
|
283 |
+
if inversion:
|
284 |
+
timesteps = reversed(self.scheduler.timesteps)
|
285 |
+
else:
|
286 |
+
timesteps = self.scheduler.timesteps
|
287 |
+
alpha_prod_t = self.scheduler.alphas_cumprod[t]
|
288 |
+
if inversion:
|
289 |
+
alpha_prod_t_prev = (
|
290 |
+
self.scheduler.alphas_cumprod[timesteps[i - 1]]
|
291 |
+
if i > 0 else self.scheduler.final_alpha_cumprod
|
292 |
+
)
|
293 |
+
else:
|
294 |
+
alpha_prod_t_prev = (
|
295 |
+
self.scheduler.alphas_cumprod[timesteps[i + 1]]
|
296 |
+
if i < len(timesteps) - 1
|
297 |
+
else self.scheduler.final_alpha_cumprod
|
298 |
+
)
|
299 |
+
mu = alpha_prod_t ** 0.5
|
300 |
+
sigma = (1 - alpha_prod_t) ** 0.5
|
301 |
+
mu_prev = alpha_prod_t_prev ** 0.5
|
302 |
+
sigma_prev = (1 - alpha_prod_t_prev) ** 0.5
|
303 |
+
|
304 |
+
if inversion:
|
305 |
+
pred_x0 = (x - sigma_prev * eps) / mu_prev
|
306 |
+
x = mu * pred_x0 + sigma * eps
|
307 |
+
else:
|
308 |
+
pred_x0 = (x - sigma * eps) / mu
|
309 |
+
x = mu_prev * pred_x0 + sigma_prev * eps
|
310 |
+
|
311 |
+
return x
|
312 |
+
|
313 |
+
def init_pnp(self, conv_injection_t, qk_injection_t):
|
314 |
+
qk_injection_timesteps = self.scheduler.timesteps[:qk_injection_t] if qk_injection_t >= 0 else []
|
315 |
+
conv_injection_timesteps = self.scheduler.timesteps[:conv_injection_t] if conv_injection_t >= 0 else []
|
316 |
+
register_attention_control(
|
317 |
+
self, qk_injection_timesteps, num_inputs=self.batch_size)
|
318 |
+
register_conv_control(
|
319 |
+
self, conv_injection_timesteps, num_inputs=self.batch_size)
|
320 |
+
|
321 |
+
def check_latent_exists(self, latent_path):
|
322 |
+
if self.use_pnp:
|
323 |
+
timesteps = self.scheduler.timesteps
|
324 |
+
else:
|
325 |
+
timesteps = [self.scheduler.timesteps[0]]
|
326 |
+
|
327 |
+
for ts in timesteps:
|
328 |
+
cur_latent_path = os.path.join(
|
329 |
+
latent_path, f'noisy_latents_{ts}.pt')
|
330 |
+
if not os.path.exists(cur_latent_path):
|
331 |
+
return False
|
332 |
+
return True
|
333 |
+
|
334 |
+
@torch.no_grad()
|
335 |
+
def __call__(self, data_path, latent_path, output_path, frame_ids):
|
336 |
+
self.scheduler.set_timesteps(self.n_timesteps)
|
337 |
+
latent_path = get_latents_dir(latent_path, self.model_key)
|
338 |
+
assert self.check_latent_exists(
|
339 |
+
latent_path), f"Required latent not found at {latent_path}. \
|
340 |
+
Note: If using PnP as control, you need inversion latents saved \
|
341 |
+
at each generation timestep."
|
342 |
+
|
343 |
+
self.data_path = data_path
|
344 |
+
self.latent_path = latent_path
|
345 |
+
self.frame_ids = frame_ids
|
346 |
+
self.prepare_data(data_path, latent_path, frame_ids)
|
347 |
+
|
348 |
+
print(f"[INFO] initial noise latent shape: {self.init_noise.shape}")
|
349 |
+
|
350 |
+
for edit_name, edit_prompt in self.prompt.items():
|
351 |
+
print(f"[INFO] current prompt: {edit_prompt}")
|
352 |
+
conds = self.get_text_embeds_input(edit_prompt, self.negative_prompt)
|
353 |
+
# Comment this if you have enough GPU memory
|
354 |
+
clean_latent = self.ddim_sample(self.init_noise, conds)
|
355 |
+
torch.cuda.empty_cache()
|
356 |
+
clean_frames = self.decode_latents_batch(clean_latent)
|
357 |
+
cur_output_path = os.path.join(output_path, edit_name)
|
358 |
+
save_config(self.config, cur_output_path, gene = True)
|
359 |
+
save_video(clean_frames, cur_output_path, save_frame = self.save_frame)
|
360 |
+
|
361 |
+
|
362 |
+
|
363 |
+
|
364 |
+
|
365 |
+
if __name__ == "__main__":
|
366 |
+
config = load_config()
|
367 |
+
pipe, scheduler, model_key = init_model(
|
368 |
+
config.device, config.sd_version, config.model_key, config.generation.control, config.float_precision)
|
369 |
+
config.model_key = model_key
|
370 |
+
seed_everything(config.seed)
|
371 |
+
generator = Generator(pipe, scheduler, config)
|
372 |
+
frame_ids = get_frame_ids(
|
373 |
+
config.generation.frame_range, config.generation.frame_ids)
|
374 |
+
generator(config.input_path, config.generation.latents_path,
|
375 |
+
config.generation.output_path, frame_ids=frame_ids)
|