jadechoghari
commited on
Commit
•
94ccc87
1
Parent(s):
f5bb4af
Create invert.py
Browse files
invert.py
ADDED
@@ -0,0 +1,289 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import torch
|
3 |
+
from tqdm import tqdm
|
4 |
+
import os
|
5 |
+
from transformers import logging
|
6 |
+
|
7 |
+
from utils import load_config, save_config
|
8 |
+
from utils import get_controlnet_kwargs, get_latents_dir, init_model, seed_everything
|
9 |
+
from utils import load_video, prepare_depth, save_frames, control_preprocess
|
10 |
+
|
11 |
+
# suppress partial model loading warning
|
12 |
+
logging.set_verbosity_error()
|
13 |
+
|
14 |
+
|
15 |
+
class Inverter(nn.Module):
|
16 |
+
def __init__(self, pipe, scheduler, config):
|
17 |
+
super().__init__()
|
18 |
+
|
19 |
+
self.device = config.device
|
20 |
+
self.use_depth = config.sd_version == "depth"
|
21 |
+
self.model_key = config.model_key
|
22 |
+
|
23 |
+
self.config = config
|
24 |
+
inv_config = config.inversion
|
25 |
+
|
26 |
+
float_precision = inv_config.float_precision if "float_precision" in inv_config else config.float_precision
|
27 |
+
if float_precision == "fp16":
|
28 |
+
self.dtype = torch.float16
|
29 |
+
print("[INFO] float precision fp16. Use torch.float16.")
|
30 |
+
else:
|
31 |
+
self.dtype = torch.float32
|
32 |
+
print("[INFO] float precision fp32. Use torch.float32.")
|
33 |
+
|
34 |
+
self.pipe = pipe
|
35 |
+
self.vae = pipe.vae
|
36 |
+
self.tokenizer = pipe.tokenizer
|
37 |
+
self.unet = pipe.unet
|
38 |
+
self.text_encoder = pipe.text_encoder
|
39 |
+
if config.enable_xformers_memory_efficient_attention:
|
40 |
+
try:
|
41 |
+
pipe.enable_xformers_memory_efficient_attention()
|
42 |
+
except ModuleNotFoundError:
|
43 |
+
print("[WARNING] xformers not found. Disable xformers attention.")
|
44 |
+
|
45 |
+
self.control = inv_config.control
|
46 |
+
if self.control != "none":
|
47 |
+
self.controlnet = pipe.controlnet
|
48 |
+
|
49 |
+
self.controlnet_scale = inv_config.control_scale
|
50 |
+
|
51 |
+
scheduler.set_timesteps(inv_config.save_steps)
|
52 |
+
self.timesteps_to_save = scheduler.timesteps
|
53 |
+
scheduler.set_timesteps(inv_config.steps)
|
54 |
+
|
55 |
+
self.scheduler = scheduler
|
56 |
+
|
57 |
+
self.prompt=inv_config.prompt
|
58 |
+
self.recon=inv_config.recon
|
59 |
+
self.save_latents=inv_config.save_intermediate
|
60 |
+
self.use_blip=inv_config.use_blip
|
61 |
+
self.steps=inv_config.steps
|
62 |
+
self.batch_size = inv_config.batch_size
|
63 |
+
self.force = inv_config.force
|
64 |
+
|
65 |
+
self.n_frames = inv_config.n_frames
|
66 |
+
self.frame_height, self.frame_width = config.height, config.width
|
67 |
+
self.work_dir = config.work_dir
|
68 |
+
|
69 |
+
|
70 |
+
@torch.no_grad()
|
71 |
+
def get_text_embeds(self, prompt, negative_prompt=None, device="cuda"):
|
72 |
+
text_input = self.tokenizer(prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
|
73 |
+
truncation=True, return_tensors='pt')
|
74 |
+
text_embeddings = self.text_encoder(text_input.input_ids.to(device))[0]
|
75 |
+
if negative_prompt is not None:
|
76 |
+
uncond_input = self.tokenizer(negative_prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
|
77 |
+
return_tensors='pt')
|
78 |
+
uncond_embeddings = self.text_encoder(
|
79 |
+
uncond_input.input_ids.to(device))[0]
|
80 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
81 |
+
return text_embeddings
|
82 |
+
|
83 |
+
@torch.no_grad()
|
84 |
+
def decode_latents(self, latents):
|
85 |
+
with torch.autocast(device_type=self.device, dtype=self.dtype):
|
86 |
+
latents = 1 / 0.18215 * latents
|
87 |
+
imgs = self.vae.decode(latents).sample
|
88 |
+
imgs = (imgs / 2 + 0.5).clamp(0, 1)
|
89 |
+
return imgs
|
90 |
+
|
91 |
+
@torch.no_grad()
|
92 |
+
def decode_latents_batch(self, latents):
|
93 |
+
imgs = []
|
94 |
+
batch_latents = latents.split(self.batch_size, dim = 0)
|
95 |
+
for latent in batch_latents:
|
96 |
+
imgs += [self.decode_latents(latent)]
|
97 |
+
imgs = torch.cat(imgs)
|
98 |
+
return imgs
|
99 |
+
|
100 |
+
@torch.no_grad()
|
101 |
+
def encode_imgs(self, imgs):
|
102 |
+
with torch.autocast(device_type=self.device, dtype=self.dtype):
|
103 |
+
imgs = 2 * imgs - 1
|
104 |
+
posterior = self.vae.encode(imgs).latent_dist
|
105 |
+
latents = posterior.mean * 0.18215
|
106 |
+
return latents
|
107 |
+
|
108 |
+
@torch.no_grad()
|
109 |
+
def encode_imgs_batch(self, imgs):
|
110 |
+
latents = []
|
111 |
+
batch_imgs = imgs.split(self.batch_size, dim = 0)
|
112 |
+
for img in batch_imgs:
|
113 |
+
latents += [self.encode_imgs(img)]
|
114 |
+
latents = torch.cat(latents)
|
115 |
+
return latents
|
116 |
+
|
117 |
+
@torch.no_grad()
|
118 |
+
def ddim_inversion(self, x, conds, save_path):
|
119 |
+
print("[INFO] start DDIM Inversion!")
|
120 |
+
timesteps = reversed(self.scheduler.timesteps)
|
121 |
+
with torch.autocast(device_type=self.device, dtype=self.dtype):
|
122 |
+
for i, t in enumerate(tqdm(timesteps)):
|
123 |
+
noises = []
|
124 |
+
x_index = torch.arange(len(x))
|
125 |
+
batches = x_index.split(self.batch_size, dim = 0)
|
126 |
+
for batch in batches:
|
127 |
+
noise = self.pred_noise(
|
128 |
+
x[batch], conds[batch], timesteps[i], batch_idx=batch)
|
129 |
+
noises += [noise]
|
130 |
+
noises = torch.cat(noises)
|
131 |
+
x = self.pred_next_x(x, noises, t, i, inversion=True)
|
132 |
+
if self.save_latents and t in self.timesteps_to_save:
|
133 |
+
torch.save(x, os.path.join(
|
134 |
+
save_path, f'noisy_latents_{t}.pt'))
|
135 |
+
|
136 |
+
# Save inverted noise latents
|
137 |
+
pth = os.path.join(save_path, f'noisy_latents_{t}.pt')
|
138 |
+
torch.save(x, pth)
|
139 |
+
print(f"[INFO] inverted latent saved to: {pth}")
|
140 |
+
return x
|
141 |
+
|
142 |
+
@torch.no_grad()
|
143 |
+
def ddim_sample(self, x, conds):
|
144 |
+
print("[INFO] reconstructing frames...")
|
145 |
+
timesteps = self.scheduler.timesteps
|
146 |
+
with torch.autocast(device_type=self.device, dtype=self.dtype):
|
147 |
+
for i, t in enumerate(tqdm(timesteps)):
|
148 |
+
noises = []
|
149 |
+
x_index = torch.arange(len(x))
|
150 |
+
batches = x_index.split(self.batch_size, dim = 0)
|
151 |
+
for batch in batches:
|
152 |
+
noise = self.pred_noise(
|
153 |
+
x[batch], conds[batch], t, batch_idx=batch)
|
154 |
+
noises += [noise]
|
155 |
+
noises = torch.cat(noises)
|
156 |
+
x = self.pred_next_x(x, noises, t, i, inversion=False)
|
157 |
+
return x
|
158 |
+
|
159 |
+
@torch.no_grad()
|
160 |
+
def pred_noise(self, x, cond, t, batch_idx=None):
|
161 |
+
# For sd-depth model
|
162 |
+
if self.use_depth:
|
163 |
+
depth = self.depths
|
164 |
+
if batch_idx is not None:
|
165 |
+
depth = depth[batch_idx]
|
166 |
+
x = torch.cat([x, depth.to(x)], dim=1)
|
167 |
+
|
168 |
+
kwargs = dict()
|
169 |
+
# Compute controlnet outputs
|
170 |
+
if self.control != "none":
|
171 |
+
if batch_idx is None:
|
172 |
+
controlnet_cond = self.controlnet_images
|
173 |
+
else:
|
174 |
+
controlnet_cond = self.controlnet_images[batch_idx]
|
175 |
+
controlnet_kwargs = get_controlnet_kwargs(self.controlnet, x, cond, t, controlnet_cond, self.controlnet_scale)
|
176 |
+
kwargs.update(controlnet_kwargs)
|
177 |
+
|
178 |
+
eps = self.unet(x, t, encoder_hidden_states=cond, **kwargs).sample
|
179 |
+
return eps
|
180 |
+
|
181 |
+
@torch.no_grad()
|
182 |
+
def pred_next_x(self, x, eps, t, i, inversion=False):
|
183 |
+
if inversion:
|
184 |
+
timesteps = reversed(self.scheduler.timesteps)
|
185 |
+
else:
|
186 |
+
timesteps = self.scheduler.timesteps
|
187 |
+
alpha_prod_t = self.scheduler.alphas_cumprod[t]
|
188 |
+
if inversion:
|
189 |
+
alpha_prod_t_prev = (
|
190 |
+
self.scheduler.alphas_cumprod[timesteps[i - 1]]
|
191 |
+
if i > 0 else self.scheduler.final_alpha_cumprod
|
192 |
+
)
|
193 |
+
else:
|
194 |
+
alpha_prod_t_prev = (
|
195 |
+
self.scheduler.alphas_cumprod[timesteps[i + 1]]
|
196 |
+
if i < len(timesteps) - 1
|
197 |
+
else self.scheduler.final_alpha_cumprod
|
198 |
+
)
|
199 |
+
mu = alpha_prod_t ** 0.5
|
200 |
+
sigma = (1 - alpha_prod_t) ** 0.5
|
201 |
+
mu_prev = alpha_prod_t_prev ** 0.5
|
202 |
+
sigma_prev = (1 - alpha_prod_t_prev) ** 0.5
|
203 |
+
|
204 |
+
if inversion:
|
205 |
+
pred_x0 = (x - sigma_prev * eps) / mu_prev
|
206 |
+
x = mu * pred_x0 + sigma * eps
|
207 |
+
else:
|
208 |
+
pred_x0 = (x - sigma * eps) / mu
|
209 |
+
x = mu_prev * pred_x0 + sigma_prev * eps
|
210 |
+
|
211 |
+
return x
|
212 |
+
|
213 |
+
@torch.no_grad()
|
214 |
+
def prepare_cond(self, prompts, n_frames):
|
215 |
+
if isinstance(prompts, str):
|
216 |
+
prompts = [prompts] * n_frames
|
217 |
+
cond = self.get_text_embeds(prompts[0])
|
218 |
+
conds = torch.cat([cond] * n_frames)
|
219 |
+
elif isinstance(prompts, list):
|
220 |
+
cond_ls = []
|
221 |
+
for prompt in prompts:
|
222 |
+
cond = self.get_text_embeds(prompt)
|
223 |
+
cond_ls += [cond]
|
224 |
+
conds = torch.cat(cond_ls)
|
225 |
+
return conds, prompts
|
226 |
+
|
227 |
+
def check_latent_exists(self, save_path):
|
228 |
+
save_timesteps = [self.scheduler.timesteps[0]]
|
229 |
+
if self.save_latents:
|
230 |
+
save_timesteps += self.timesteps_to_save
|
231 |
+
for ts in save_timesteps:
|
232 |
+
latent_path = os.path.join(
|
233 |
+
save_path, f'noisy_latents_{ts}.pt')
|
234 |
+
if not os.path.exists(latent_path):
|
235 |
+
return False
|
236 |
+
return True
|
237 |
+
|
238 |
+
|
239 |
+
@torch.no_grad()
|
240 |
+
def __call__(self, data_path, save_path):
|
241 |
+
self.scheduler.set_timesteps(self.steps)
|
242 |
+
save_path = get_latents_dir(save_path, self.model_key)
|
243 |
+
os.makedirs(save_path, exist_ok = True)
|
244 |
+
if self.check_latent_exists(save_path) and not self.force:
|
245 |
+
print(f"[INFO] inverted latents exist at: {save_path}. Skip inversion! Set 'inversion.force: True' to invert again.")
|
246 |
+
return
|
247 |
+
|
248 |
+
frames = load_video(data_path, self.frame_height, self.frame_width, device = self.device)
|
249 |
+
|
250 |
+
frame_ids = list(range(len(frames)))
|
251 |
+
if self.n_frames is not None:
|
252 |
+
frame_ids = frame_ids[:self.n_frames]
|
253 |
+
frames = frames[frame_ids]
|
254 |
+
|
255 |
+
if self.use_depth:
|
256 |
+
self.depths = prepare_depth(self.pipe, frames, frame_ids, self.work_dir)
|
257 |
+
conds, prompts = self.prepare_cond(self.prompt, len(frames))
|
258 |
+
with open(os.path.join(save_path, 'inversion_prompts.txt'), 'w') as f:
|
259 |
+
f.write('\n'.join(prompts))
|
260 |
+
|
261 |
+
if self.control != "none":
|
262 |
+
images = control_preprocess(
|
263 |
+
frames, self.control)
|
264 |
+
self.controlnet_images = images.to(self.device)
|
265 |
+
|
266 |
+
latents = self.encode_imgs_batch(frames)
|
267 |
+
torch.cuda.empty_cache()
|
268 |
+
print(f"[INFO] clean latents shape: {latents.shape}")
|
269 |
+
|
270 |
+
inverted_x = self.ddim_inversion(latents, conds, save_path)
|
271 |
+
save_config(self.config, save_path, inv = True)
|
272 |
+
if self.recon:
|
273 |
+
latent_reconstruction = self.ddim_sample(inverted_x, conds)
|
274 |
+
|
275 |
+
torch.cuda.empty_cache()
|
276 |
+
recon_frames = self.decode_latents_batch(
|
277 |
+
latent_reconstruction)
|
278 |
+
|
279 |
+
recon_save_path = os.path.join(save_path, 'recon_frames')
|
280 |
+
save_frames(recon_frames, recon_save_path, frame_ids = frame_ids)
|
281 |
+
|
282 |
+
if __name__ == "__main__":
|
283 |
+
config = load_config()
|
284 |
+
pipe, scheduler, model_key = init_model(
|
285 |
+
config.device, config.sd_version, config.model_key, config.inversion.control, config.float_precision)
|
286 |
+
config.model_key = model_key
|
287 |
+
seed_everything(config.seed)
|
288 |
+
inversion = Inverter(pipe, scheduler, config)
|
289 |
+
inversion(config.input_path, config.inversion.save_path)
|