-
Notifications
You must be signed in to change notification settings - Fork 24
/
inversion.py
184 lines (165 loc) · 8.38 KB
/
inversion.py
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
# Null-text embedding Inversion, code from https://github.com/google/prompt-to-prompt/blob/main/null_text_w_ptp.ipynb
from typing import Union, List, Dict
import torch
import numpy as np
from diffusers import DDIMScheduler
import utils
from PIL import Image
from tqdm.notebook import tqdm
from torch.optim.adam import Adam
import torch.nn.functional as nnf
def load_512(image_path):
image = np.array(Image.open(image_path))[:, :, :3]
h, w, c = image.shape
if h < w:
offset = (w - h) // 2
image = image[:, offset:offset + h]
elif w < h:
offset = (h - w) // 2
image = image[offset:offset + w]
image = np.array(Image.fromarray(image).resize((512, 512)))
return image
class NullInversion:
def prev_step(self, model_output: Union[torch.FloatTensor, np.ndarray], timestep: int, sample: Union[torch.FloatTensor, np.ndarray]):
prev_timestep = timestep - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps
alpha_prod_t = self.scheduler.alphas_cumprod[timestep]
alpha_prod_t_prev = self.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.scheduler.final_alpha_cumprod
beta_prod_t = 1 - alpha_prod_t
pred_original_sample = (sample - beta_prod_t ** 0.5 * model_output) / alpha_prod_t ** 0.5
pred_sample_direction = (1 - alpha_prod_t_prev) ** 0.5 * model_output
prev_sample = alpha_prod_t_prev ** 0.5 * pred_original_sample + pred_sample_direction
return prev_sample
def next_step(self, model_output: Union[torch.FloatTensor, np.ndarray], timestep: int, sample: Union[torch.FloatTensor, np.ndarray]):
timestep, next_timestep = min(timestep - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps, 999), timestep
alpha_prod_t = self.scheduler.alphas_cumprod[timestep] if timestep >= 0 else self.scheduler.final_alpha_cumprod
alpha_prod_t_next = self.scheduler.alphas_cumprod[next_timestep]
beta_prod_t = 1 - alpha_prod_t
next_original_sample = (sample - beta_prod_t ** 0.5 * model_output) / alpha_prod_t ** 0.5
next_sample_direction = (1 - alpha_prod_t_next) ** 0.5 * model_output
next_sample = alpha_prod_t_next ** 0.5 * next_original_sample + next_sample_direction
return next_sample
def get_noise_pred_single(self, latents, t, context):
noise_pred = self.model.unet(latents, t, encoder_hidden_states=context)["sample"]
return noise_pred
def get_noise_pred(self, latents, t, is_forward=True, context=None):
latents_input = torch.cat([latents] * 2)
if context is None:
context = self.context
guidance_scale = 1 if is_forward else self.GUIDANCE_SCALE
noise_pred = self.model.unet(latents_input, t, encoder_hidden_states=context)["sample"]
noise_pred_uncond, noise_prediction_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
if is_forward:
latents = self.next_step(noise_pred, t, latents)
else:
latents = self.prev_step(noise_pred, t, latents)
return latents
@torch.no_grad()
def latent2image(self, latents, return_type='np'):
latents = 1 / 0.18215 * latents.detach()
image = self.model.vae.decode(latents)['sample']
if return_type == 'np':
image = (image / 2 + 0.5).clamp(0, 1)
image = image.cpu().permute(0, 2, 3, 1).numpy()[0]
image = (image * 255).astype(np.uint8)
return image
@torch.no_grad()
def image2latent(self, image):
with torch.no_grad():
if type(image) is Image:
image = np.array(image)
if type(image) is torch.Tensor and image.dim() == 4:
latents = image
else:
image = torch.from_numpy(image).float() / 127.5 - 1
image = image.permute(2, 0, 1).unsqueeze(0).to(self.model.device)
latents = self.model.vae.encode(image)['latent_dist'].mean
latents = latents * 0.18215
return latents
@torch.no_grad()
def init_prompt(self, prompt: str):
uncond_input = self.model.tokenizer(
[""], padding="max_length", max_length=self.model.tokenizer.model_max_length,
return_tensors="pt"
)
uncond_embeddings = self.model.text_encoder(uncond_input.input_ids.to(self.model.device))[0]
text_input = self.model.tokenizer(
[prompt],
padding="max_length",
max_length=self.model.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
text_embeddings = self.model.text_encoder(text_input.input_ids.to(self.model.device))[0]
self.context = torch.cat([uncond_embeddings, text_embeddings])
self.prompt = prompt
@torch.no_grad()
def ddim_loop(self, latent):
uncond_embeddings, cond_embeddings = self.context.chunk(2)
all_latent = [latent]
latent = latent.clone().detach()
for i in range(self.NUM_DDIM_STEPS):
t = self.model.scheduler.timesteps[len(self.model.scheduler.timesteps) - i - 1]
noise_pred = self.get_noise_pred_single(latent, t, cond_embeddings)
latent = self.next_step(noise_pred, t, latent)
all_latent.append(latent)
return all_latent
@property
def scheduler(self):
return self.model.scheduler
@torch.no_grad()
def ddim_inversion(self, image):
latent = self.image2latent(image)
image_rec = self.latent2image(latent)
ddim_latents = self.ddim_loop(latent)
return image_rec, ddim_latents
def null_optimization(self, latents, num_inner_steps, epsilon):
uncond_embeddings, cond_embeddings = self.context.chunk(2)
uncond_embeddings_list = []
latent_cur = latents[-1]
bar = tqdm(total=num_inner_steps * self.NUM_DDIM_STEPS)
for i in range(self.NUM_DDIM_STEPS):
uncond_embeddings = uncond_embeddings.clone().detach()
uncond_embeddings.requires_grad = True
optimizer = Adam([uncond_embeddings], lr=1e-2 * (1. - i / 100.))
latent_prev = latents[len(latents) - i - 2]
t = self.model.scheduler.timesteps[i]
with torch.no_grad():
noise_pred_cond = self.get_noise_pred_single(latent_cur, t, cond_embeddings)
for j in range(num_inner_steps):
noise_pred_uncond = self.get_noise_pred_single(latent_cur, t, uncond_embeddings)
noise_pred = noise_pred_uncond + self.GUIDANCE_SCALE * (noise_pred_cond - noise_pred_uncond)
latents_prev_rec = self.prev_step(noise_pred, t, latent_cur)
loss = nnf.mse_loss(latents_prev_rec, latent_prev)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_item = loss.item()
bar.update()
if loss_item < epsilon + i * 2e-5:
break
for j in range(j + 1, num_inner_steps):
bar.update()
uncond_embeddings_list.append(uncond_embeddings[:1].detach())
with torch.no_grad():
context = torch.cat([uncond_embeddings, cond_embeddings])
latent_cur = self.get_noise_pred(latent_cur, t, False, context)
bar.close()
return uncond_embeddings_list
def invert(self, image_path: str, prompt: str, num_inner_steps=10, early_stop_epsilon=1e-5):
self.init_prompt(prompt)
utils.register_attention_control(self.model, None)
image_gt = load_512(image_path)
image_rec, ddim_latents = self.ddim_inversion(image_gt)
uncond_embeddings = self.null_optimization(ddim_latents, num_inner_steps, early_stop_epsilon)
return ddim_latents[-1], uncond_embeddings
def __init__(self, model, ddim_steps, guidance_scale):
scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False,
set_alpha_to_one=False)
self.model = model
self.tokenizer = self.model.tokenizer
self.model.scheduler.set_timesteps(ddim_steps)
self.prompt = None
self.context = None
self.NUM_DDIM_STEPS = ddim_steps
self.GUIDANCE_SCALE = guidance_scale