-
Notifications
You must be signed in to change notification settings - Fork 0
/
diffusion.py
272 lines (172 loc) · 9.81 KB
/
diffusion.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
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
import numpy as np
try:
import cupy as cp
is_cupy_available = True
print('CuPy is available. Using CuPy for all computations.')
except:
is_cupy_available = False
print('CuPy is not available. Switching to NumPy.')
import sys
import os
import pickle as pkl
from pathlib import Path
sys.path[0] = str(Path(sys.path[0]).parent)
from tqdm import tqdm
from PIL import Image
from typing import Type, Union, List, Tuple, Dict, Optional, Callable
from diffusion.schedules import get_schedule
# https://arxiv.org/abs/2006.11239
# https://arxiv.org/abs/2102.09672
# https://huggingface.co/blog/annotated-diffusion
# https://lilianweng.github.io/posts/2021-07-11-diffusion-models/
# https://nn.labml.ai/diffusion/ddpm/index.html
class Diffusion():
def __init__(self, timesteps: int, beta_start: float, beta_end: float, criterion, optimizer, model = None, schedule = "linear"):
self.model = model
self.beta_start = beta_start
self.beta_end = beta_end
self.timesteps = timesteps
self.betas = get_schedule(schedule, beta_start, beta_end, timesteps)
self.sqrt_betas = np.sqrt(self.betas)
self.alphas = 1 - self.betas
self.inv_sqrt_alphas = 1 / np.sqrt(self.alphas)
self.alphas_cumprod = np.cumprod(self.alphas, axis = 0)
self.sqrt_alphas_cumprod = np.sqrt(self.alphas_cumprod)
self.sqrt_one_minus_alphas_cumprod = np.sqrt(1 - self.alphas_cumprod)
self.scaled_alphas = (1 - self.alphas) / self.sqrt_one_minus_alphas_cumprod
# self.sqrt_recip_alphas = np.sqrt(1.0 / self.alphas)
# self.alphas_cumprod_prev = np.concatenate([np.array([1]), self.alphas_cumprod[:-1]]) #np.pad(alphas_cumprod[:-1], (1, 0), value=1.0)
# self.posterior_variance = self.betas * (1. - self.alphas_cumprod_prev) / (1. - self.alphas_cumprod)
self.criterion = criterion
self.optimizer = optimizer
def load(self, path: str):
pickle_model = open(f'{path}/model.pkl', 'rb')
self.model = pkl.load(pickle_model)
pickle_model.close()
print(f'Loaded from "{path}"')
def save(self, path: str):
if not os.path.exists(path):
os.makedirs(path)
pickle_model = open(f'{path}/model.pkl', 'wb')
pkl.dump(self.model, pickle_model)
pickle_model.close()
print(f'Saved to "{path}"')
def forward(self, x: np.float32, t = None):
"""
https://arxiv.org/abs/2006.11239
Algorithm 1: Training; according to the paper
"""
timesteps_selection = np.random.randint(1, self.timesteps, (x.shape[0],))
noise = np.random.normal(size = x.shape)
x_t = self.sqrt_alphas_cumprod[timesteps_selection, None, None, None] * x + self.sqrt_one_minus_alphas_cumprod[timesteps_selection, None, None, None] * noise
x = self.model.forward(x_t, timesteps_selection / self.timesteps, training = True)
return x, noise
def ddpm_denoise_sample(self, sample_num = None, image_size = None, states_step_size = 1, x_t = None, orig_x = None, mask = None):
"""
https://arxiv.org/abs/2006.11239
Algorithm 2: Sampling; according to the paper
"""
if mask is not None:
assert orig_x is not None
assert orig_x.shape == mask.shape
if x_t is None:
if orig_x is None:
x_t = np.random.normal(size = (sample_num, *image_size))
else:
x_t = np.random.normal(size = orig_x.shape)
x_ts = []
for t in tqdm(reversed(range(0, self.timesteps)), desc = 'ddpm denoisinig samples', total = self.timesteps):
noise = np.random.normal(size = x_t.shape) if t > 1 else 0
epsilon = cp.asnumpy(self.model.forward(x_t, np.array([t]) / self.timesteps, training = False)).reshape(x_t.shape)
x_t = self.inv_sqrt_alphas[t] * (x_t - epsilon * self.scaled_alphas[t]) + self.sqrt_betas[t] * noise
# x_t = self.sqrt_recip_alphas[t] * (x_t - self.betas[t] * epsilon / self.sqrt_one_minus_alphas_cumprod[t]) + np.sqrt(self.posterior_variance[t]) * noise
if mask is not None:
orig_x_noise = np.random.normal(size = orig_x.shape)
orig_x_t = self.sqrt_alphas_cumprod[t] * orig_x + self.sqrt_one_minus_alphas_cumprod[t] * orig_x_noise
x_t = orig_x_t * mask + x_t * (1 - mask)
if t % states_step_size == 0:
x_ts.append(x_t)
return x_t, x_ts
def ddim_denoise_sample(self, sample_num = None, image_size = None, states_step_size = 1, eta = 1., perform_steps = 100, x_t = None, orig_x = None, mask = None):
"""
https://arxiv.org/abs/2010.02502
Denoising Diffusion Implicit Models (DDIM) sampling; according to the paper
"""
if mask is not None:
assert orig_x is not None
assert orig_x.shape == mask.shape
if x_t is None:
if orig_x is None:
x_t = np.random.normal(size = (sample_num, *image_size))
else:
x_t = np.random.normal(size = orig_x.shape)
x_ts = []
for t in tqdm(reversed(range(1, self.timesteps)[:perform_steps]), desc = 'ddim denoisinig samples', total = perform_steps):
noise = np.random.normal(size = x_t.shape) if t > 1 else 0
epsilon = cp.asnumpy(self.model.forward(x_t, np.array([t]) / self.timesteps, training = False)).reshape(x_t.shape)
x0_t = (x_t - epsilon * np.sqrt(1 - self.alphas_cumprod[t])) / np.sqrt(self.alphas_cumprod[t])
sigma = eta * np.sqrt((1 - self.alphas_cumprod[t - 1]) / (1 - self.alphas_cumprod[t]) * (1 - self.alphas_cumprod[t] / self.alphas_cumprod[t - 1]))
c = np.sqrt((1 - self.alphas_cumprod[t - 1]) - sigma ** 2)
x_t = np.sqrt(self.alphas_cumprod[t - 1]) * x0_t - c * epsilon + sigma * noise
if mask is not None:
orig_x_noise = np.random.normal(size = orig_x.shape)
orig_x_t = self.sqrt_alphas_cumprod[t] * orig_x + self.sqrt_one_minus_alphas_cumprod[t] * orig_x_noise
x_t = orig_x_t * mask + x_t * (1 - mask)
if t % states_step_size == 0:
x_ts.append(x_t)
return x_t, x_ts
def get_images_set(self, x_num: int, y_num: int, margin: int, images: np.float32, image_size: Tuple[int, int, int]):
def denormalize(x):
return (x - np.min(x)) / (np.max(x) - np.min(x)) * 255
channels, H_size, W_size = image_size
images_array = np.full((y_num * (margin + H_size), x_num * (margin + W_size), channels), 255, dtype = np.uint8)
num = 0
for i in range(y_num):
for j in range(x_num):
y = i * (margin + H_size)
x = j * (margin + W_size)
images_array[y :y + H_size, x: x + W_size] = denormalize(images[num].transpose(1, 2, 0))
num += 1
images_array = images_array[: (y_num - 1) * (H_size + margin) + H_size, : (x_num - 1) * (W_size + margin) + W_size]
if channels == 1:
return Image.fromarray(images_array.squeeze(axis = 2)).convert("L")
else:
return Image.fromarray(images_array)
def train(self, dataset, epochs, batch_size, save_every_epochs, image_path, save_path, image_size):
channels, H_size, W_size = image_size
self.model.set_optimizer(self.optimizer)
data_batches = np.array_split(dataset, np.arange(batch_size, len(dataset), batch_size))
loss_history = []
for epoch in range(epochs):
tqdm_range = tqdm(enumerate(data_batches), total = len(data_batches))
loss = []
for batch_num, (batch) in tqdm_range:
output, noise = self.forward(batch)
loss.append(self.criterion.loss(output, noise).mean())
error = self.criterion.derivative(output, noise)
error = self.model.backward(error)
self.model.update_weights()
tqdm_range.set_description(
f"loss: {loss[-1]:.7f} | epoch {epoch + 1}/{epochs}"
)
if batch_num == (len(data_batches) - 1):
if is_cupy_available:
epoch_loss = cp.mean(cp.array(loss))
else:
epoch_loss = np.mean(loss)
tqdm_range.set_description(
f"loss: {epoch_loss:.7f} | epoch {epoch + 1}/{epochs}"
)
if ((epoch + 1) % save_every_epochs == 0):
self.save(f"{save_path}")
margin = 10
x_num, y_num = 5, 5
samples, samples_in_time = self.ddpm_denoise_sample(x_num * y_num, (channels, H_size, W_size), step_size = 10)
images_grid = self.get_images_set(x_num, y_num, margin, samples, (channels, H_size, W_size))
images_grid.save(f"{image_path}/np_ddpm_{epoch + 1}.png")
images_grid_in_time = []
for sample in samples_in_time:
images_grid_in_time.append(self.get_images_set(x_num, y_num, margin, sample, (channels, H_size, W_size)))
images_grid_in_time[0].save(f"{image_path}/np_ddpm_in_time_{epoch + 1}.gif", save_all = True, append_images = images_grid_in_time[1:], duration = 50, loop = 0)
loss_history.append(epoch_loss)
return loss_history