Linoy Tsaban
commited on
Commit
•
b9a325a
1
Parent(s):
4bfe3d8
Rename utils.py to inversion_utils.py
Browse files- inversion_utils.py +291 -0
- utils.py +0 -2
inversion_utils.py
ADDED
@@ -0,0 +1,291 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import os
|
3 |
+
from tqdm import tqdm
|
4 |
+
from PIL import Image, ImageDraw ,ImageFont
|
5 |
+
from matplotlib import pyplot as plt
|
6 |
+
import torchvision.transforms as T
|
7 |
+
import os
|
8 |
+
import yaml
|
9 |
+
import numpy as np
|
10 |
+
|
11 |
+
|
12 |
+
def load_512(image_path, left=0, right=0, top=0, bottom=0, device=None):
|
13 |
+
if type(image_path) is str:
|
14 |
+
image = np.array(Image.open(image_path).convert('RGB'))[:, :, :3]
|
15 |
+
else:
|
16 |
+
image = image_path
|
17 |
+
h, w, c = image.shape
|
18 |
+
left = min(left, w-1)
|
19 |
+
right = min(right, w - left - 1)
|
20 |
+
top = min(top, h - left - 1)
|
21 |
+
bottom = min(bottom, h - top - 1)
|
22 |
+
image = image[top:h-bottom, left:w-right]
|
23 |
+
h, w, c = image.shape
|
24 |
+
if h < w:
|
25 |
+
offset = (w - h) // 2
|
26 |
+
image = image[:, offset:offset + h]
|
27 |
+
elif w < h:
|
28 |
+
offset = (h - w) // 2
|
29 |
+
image = image[offset:offset + w]
|
30 |
+
image = np.array(Image.fromarray(image).resize((512, 512)))
|
31 |
+
image = torch.from_numpy(image).float() / 127.5 - 1
|
32 |
+
image = image.permute(2, 0, 1).unsqueeze(0).to(device)
|
33 |
+
|
34 |
+
return image
|
35 |
+
|
36 |
+
|
37 |
+
def load_real_image(folder = "data/", img_name = None, idx = 0, img_size=512, device='cuda'):
|
38 |
+
from PIL import Image
|
39 |
+
from glob import glob
|
40 |
+
if img_name is not None:
|
41 |
+
path = os.path.join(folder, img_name)
|
42 |
+
else:
|
43 |
+
path = glob(folder + "*")[idx]
|
44 |
+
|
45 |
+
img = Image.open(path).resize((img_size,
|
46 |
+
img_size))
|
47 |
+
|
48 |
+
img = pil_to_tensor(img).to(device)
|
49 |
+
|
50 |
+
if img.shape[1]== 4:
|
51 |
+
img = img[:,:3,:,:]
|
52 |
+
return img
|
53 |
+
|
54 |
+
def mu_tilde(model, xt,x0, timestep):
|
55 |
+
"mu_tilde(x_t, x_0) DDPM paper eq. 7"
|
56 |
+
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
|
57 |
+
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
|
58 |
+
alpha_t = model.scheduler.alphas[timestep]
|
59 |
+
beta_t = 1 - alpha_t
|
60 |
+
alpha_bar = model.scheduler.alphas_cumprod[timestep]
|
61 |
+
return ((alpha_prod_t_prev ** 0.5 * beta_t) / (1-alpha_bar)) * x0 + ((alpha_t**0.5 *(1-alpha_prod_t_prev)) / (1- alpha_bar))*xt
|
62 |
+
|
63 |
+
def sample_xts_from_x0(model, x0, num_inference_steps=50):
|
64 |
+
"""
|
65 |
+
Samples from P(x_1:T|x_0)
|
66 |
+
"""
|
67 |
+
# torch.manual_seed(43256465436)
|
68 |
+
alpha_bar = model.scheduler.alphas_cumprod
|
69 |
+
sqrt_one_minus_alpha_bar = (1-alpha_bar) ** 0.5
|
70 |
+
alphas = model.scheduler.alphas
|
71 |
+
betas = 1 - alphas
|
72 |
+
variance_noise_shape = (
|
73 |
+
num_inference_steps,
|
74 |
+
model.unet.in_channels,
|
75 |
+
model.unet.sample_size,
|
76 |
+
model.unet.sample_size)
|
77 |
+
|
78 |
+
timesteps = model.scheduler.timesteps.to(model.device)
|
79 |
+
t_to_idx = {int(v):k for k,v in enumerate(timesteps)}
|
80 |
+
xts = torch.zeros(variance_noise_shape).to(x0.device)
|
81 |
+
for t in reversed(timesteps):
|
82 |
+
idx = t_to_idx[int(t)]
|
83 |
+
xts[idx] = x0 * (alpha_bar[t] ** 0.5) + torch.randn_like(x0) * sqrt_one_minus_alpha_bar[t]
|
84 |
+
xts = torch.cat([xts, x0 ],dim = 0)
|
85 |
+
|
86 |
+
return xts
|
87 |
+
|
88 |
+
def encode_text(model, prompts):
|
89 |
+
text_input = model.tokenizer(
|
90 |
+
prompts,
|
91 |
+
padding="max_length",
|
92 |
+
max_length=model.tokenizer.model_max_length,
|
93 |
+
truncation=True,
|
94 |
+
return_tensors="pt",
|
95 |
+
)
|
96 |
+
with torch.no_grad():
|
97 |
+
text_encoding = model.text_encoder(text_input.input_ids.to(model.device))[0]
|
98 |
+
return text_encoding
|
99 |
+
|
100 |
+
def forward_step(model, model_output, timestep, sample):
|
101 |
+
next_timestep = min(model.scheduler.config.num_train_timesteps - 2,
|
102 |
+
timestep + model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps)
|
103 |
+
|
104 |
+
# 2. compute alphas, betas
|
105 |
+
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
|
106 |
+
# alpha_prod_t_next = self.scheduler.alphas_cumprod[next_timestep] if next_ltimestep >= 0 else self.scheduler.final_alpha_cumprod
|
107 |
+
|
108 |
+
beta_prod_t = 1 - alpha_prod_t
|
109 |
+
|
110 |
+
# 3. compute predicted original sample from predicted noise also called
|
111 |
+
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
112 |
+
pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
|
113 |
+
|
114 |
+
# 5. TODO: simple noising implementatiom
|
115 |
+
next_sample = model.scheduler.add_noise(pred_original_sample,
|
116 |
+
model_output,
|
117 |
+
torch.LongTensor([next_timestep]))
|
118 |
+
return next_sample
|
119 |
+
|
120 |
+
|
121 |
+
def get_variance(model, timestep): #, prev_timestep):
|
122 |
+
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
|
123 |
+
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
|
124 |
+
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
|
125 |
+
beta_prod_t = 1 - alpha_prod_t
|
126 |
+
beta_prod_t_prev = 1 - alpha_prod_t_prev
|
127 |
+
variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
|
128 |
+
return variance
|
129 |
+
|
130 |
+
def inversion_forward_process(model, x0,
|
131 |
+
etas = None,
|
132 |
+
prog_bar = False,
|
133 |
+
prompt = "",
|
134 |
+
cfg_scale = 3.5,
|
135 |
+
num_inference_steps=50, eps = None):
|
136 |
+
|
137 |
+
if not prompt=="":
|
138 |
+
text_embeddings = encode_text(model, prompt)
|
139 |
+
uncond_embedding = encode_text(model, "")
|
140 |
+
timesteps = model.scheduler.timesteps.to(model.device)
|
141 |
+
variance_noise_shape = (
|
142 |
+
num_inference_steps,
|
143 |
+
model.unet.in_channels,
|
144 |
+
model.unet.sample_size,
|
145 |
+
model.unet.sample_size)
|
146 |
+
if etas is None or (type(etas) in [int, float] and etas == 0):
|
147 |
+
eta_is_zero = True
|
148 |
+
zs = None
|
149 |
+
else:
|
150 |
+
eta_is_zero = False
|
151 |
+
if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
|
152 |
+
xts = sample_xts_from_x0(model, x0, num_inference_steps=num_inference_steps)
|
153 |
+
alpha_bar = model.scheduler.alphas_cumprod
|
154 |
+
zs = torch.zeros(size=variance_noise_shape, device=model.device)
|
155 |
+
|
156 |
+
t_to_idx = {int(v):k for k,v in enumerate(timesteps)}
|
157 |
+
xt = x0
|
158 |
+
op = tqdm(reversed(timesteps)) if prog_bar else reversed(timesteps)
|
159 |
+
|
160 |
+
for t in op:
|
161 |
+
idx = t_to_idx[int(t)]
|
162 |
+
# 1. predict noise residual
|
163 |
+
if not eta_is_zero:
|
164 |
+
xt = xts[idx][None]
|
165 |
+
|
166 |
+
with torch.no_grad():
|
167 |
+
out = model.unet.forward(xt, timestep = t, encoder_hidden_states = uncond_embedding)
|
168 |
+
if not prompt=="":
|
169 |
+
cond_out = model.unet.forward(xt, timestep=t, encoder_hidden_states = text_embeddings)
|
170 |
+
|
171 |
+
if not prompt=="":
|
172 |
+
## classifier free guidance
|
173 |
+
noise_pred = out.sample + cfg_scale * (cond_out.sample - out.sample)
|
174 |
+
else:
|
175 |
+
noise_pred = out.sample
|
176 |
+
|
177 |
+
if eta_is_zero:
|
178 |
+
# 2. compute more noisy image and set x_t -> x_t+1
|
179 |
+
xt = forward_step(model, noise_pred, t, xt)
|
180 |
+
|
181 |
+
else:
|
182 |
+
xtm1 = xts[idx+1][None]
|
183 |
+
# pred of x0
|
184 |
+
pred_original_sample = (xt - (1-alpha_bar[t]) ** 0.5 * noise_pred ) / alpha_bar[t] ** 0.5
|
185 |
+
|
186 |
+
# direction to xt
|
187 |
+
prev_timestep = t - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
|
188 |
+
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
|
189 |
+
|
190 |
+
variance = get_variance(model, t)
|
191 |
+
pred_sample_direction = (1 - alpha_prod_t_prev - etas[idx] * variance ) ** (0.5) * noise_pred
|
192 |
+
|
193 |
+
mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
|
194 |
+
|
195 |
+
z = (xtm1 - mu_xt ) / ( etas[idx] * variance ** 0.5 )
|
196 |
+
zs[idx] = z
|
197 |
+
|
198 |
+
# correction to avoid error accumulation
|
199 |
+
xtm1 = mu_xt + ( etas[idx] * variance ** 0.5 )*z
|
200 |
+
xts[idx+1] = xtm1
|
201 |
+
|
202 |
+
if not zs is None:
|
203 |
+
zs[-1] = torch.zeros_like(zs[-1])
|
204 |
+
|
205 |
+
return xt, zs, xts
|
206 |
+
|
207 |
+
|
208 |
+
def reverse_step(model, model_output, timestep, sample, eta = 0, variance_noise=None):
|
209 |
+
# 1. get previous step value (=t-1)
|
210 |
+
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
|
211 |
+
# 2. compute alphas, betas
|
212 |
+
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
|
213 |
+
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
|
214 |
+
beta_prod_t = 1 - alpha_prod_t
|
215 |
+
# 3. compute predicted original sample from predicted noise also called
|
216 |
+
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
217 |
+
pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
|
218 |
+
# 5. compute variance: "sigma_t(η)" -> see formula (16)
|
219 |
+
# σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
|
220 |
+
# variance = self.scheduler._get_variance(timestep, prev_timestep)
|
221 |
+
variance = get_variance(model, timestep) #, prev_timestep)
|
222 |
+
std_dev_t = eta * variance ** (0.5)
|
223 |
+
# Take care of asymetric reverse process (asyrp)
|
224 |
+
model_output_direction = model_output
|
225 |
+
# 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
226 |
+
# pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * model_output_direction
|
227 |
+
pred_sample_direction = (1 - alpha_prod_t_prev - eta * variance) ** (0.5) * model_output_direction
|
228 |
+
# 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
|
229 |
+
prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
|
230 |
+
# 8. Add noice if eta > 0
|
231 |
+
if eta > 0:
|
232 |
+
if variance_noise is None:
|
233 |
+
variance_noise = torch.randn(model_output.shape, device=model.device)
|
234 |
+
sigma_z = eta * variance ** (0.5) * variance_noise
|
235 |
+
prev_sample = prev_sample + sigma_z
|
236 |
+
|
237 |
+
return prev_sample
|
238 |
+
|
239 |
+
def inversion_reverse_process(model,
|
240 |
+
xT,
|
241 |
+
etas = 0,
|
242 |
+
prompts = "",
|
243 |
+
cfg_scales = None,
|
244 |
+
prog_bar = False,
|
245 |
+
zs = None,
|
246 |
+
controller=None,
|
247 |
+
asyrp = False):
|
248 |
+
|
249 |
+
batch_size = len(prompts)
|
250 |
+
|
251 |
+
cfg_scales_tensor = torch.Tensor(cfg_scales).view(-1,1,1,1).to(model.device)
|
252 |
+
|
253 |
+
text_embeddings = encode_text(model, prompts)
|
254 |
+
uncond_embedding = encode_text(model, [""] * batch_size)
|
255 |
+
|
256 |
+
if etas is None: etas = 0
|
257 |
+
if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
|
258 |
+
assert len(etas) == model.scheduler.num_inference_steps
|
259 |
+
timesteps = model.scheduler.timesteps.to(model.device)
|
260 |
+
|
261 |
+
xt = xT.expand(batch_size, -1, -1, -1)
|
262 |
+
op = tqdm(timesteps[-zs.shape[0]:]) if prog_bar else timesteps[-zs.shape[0]:]
|
263 |
+
|
264 |
+
t_to_idx = {int(v):k for k,v in enumerate(timesteps[-zs.shape[0]:])}
|
265 |
+
|
266 |
+
for t in op:
|
267 |
+
idx = t_to_idx[int(t)]
|
268 |
+
## Unconditional embedding
|
269 |
+
with torch.no_grad():
|
270 |
+
uncond_out = model.unet.forward(xt, timestep = t,
|
271 |
+
encoder_hidden_states = uncond_embedding)
|
272 |
+
|
273 |
+
## Conditional embedding
|
274 |
+
if prompts:
|
275 |
+
with torch.no_grad():
|
276 |
+
cond_out = model.unet.forward(xt, timestep = t,
|
277 |
+
encoder_hidden_states = text_embeddings)
|
278 |
+
|
279 |
+
|
280 |
+
z = zs[idx] if not zs is None else None
|
281 |
+
z = z.expand(batch_size, -1, -1, -1)
|
282 |
+
if prompts:
|
283 |
+
## classifier free guidance
|
284 |
+
noise_pred = uncond_out.sample + cfg_scales_tensor * (cond_out.sample - uncond_out.sample)
|
285 |
+
else:
|
286 |
+
noise_pred = uncond_out.sample
|
287 |
+
# 2. compute less noisy image and set x_t -> x_t-1
|
288 |
+
xt = reverse_step(model, noise_pred, t, xt, eta = etas[idx], variance_noise = z)
|
289 |
+
if controller is not None:
|
290 |
+
xt = controller.step_callback(xt)
|
291 |
+
return xt, zs
|
utils.py
DELETED
@@ -1,2 +0,0 @@
|
|
1 |
-
def hi():
|
2 |
-
return "hi"
|
|
|
|
|
|