-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathbfnsolver_i.py
235 lines (193 loc) · 10.6 KB
/
bfnsolver_i.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
import torch
import os
from omegaconf import OmegaConf, DictConfig
from utils import seed_everything, make_config, get_nnet
from tqdm.auto import tqdm
import logging
logging.basicConfig(level=logging.DEBUG, format='%(asctime)s - %(levelname)s - %(message)s')
torch.set_float32_matmul_precision("high")
torch.backends.cudnn.benchmark = True
class ImageBFNSolver:
def __init__(self, unet: torch.nn.Module, num_steps: int = 100, min_variance: float = 1e-6, eta: float = 1e-5):
self.unet = unet
self.sigma_1 = min_variance**0.5
self.num_steps = num_steps
self.steps = torch.flip(torch.arange(num_steps+1), [0])
self.times = self.steps.to(torch.float64)/(num_steps) * (1 - eta)
self.delta_t = (1 - eta) / num_steps
self.beta_s = self.sigma_1**(-2 * (1 - self.times)) - 1
self.gamma_t = 1 - self.sigma_1**(2 * (1 - self.times))
self.alpha_t = 1 - self.sigma_1**(2 * (1 - self.times))
self.sigma_t = (self.alpha_t * (1 - self.alpha_t)).sqrt()
self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
self.f_t = 2 * torch.log(torch.tensor(self.sigma_1)) * (1 - self.gamma_t) / self.gamma_t
self.g_t = (-2 * torch.log(torch.tensor(self.sigma_1)) * (1 - self.gamma_t))**0.5
def sde_euler_update(self, x_s, step, last_drop=False):
# x_s -> x_t
t = torch.ones_like(x_s, device=x_s.device) * (1 - self.times[step])
# noise predict and x0 predict
with torch.no_grad():
noise_pred = self.unet(x_s, t).reshape(x_s.shape)
alpha_t, sigma_t = self.alpha_t[step], self.sigma_t[step]
x0_pred = (x_s - sigma_t * noise_pred) / alpha_t
# clip x0
x0_pred = x0_pred.clip(min=-1.0, max=1.0)
noise_pred = (x_s - x0_pred * alpha_t) / sigma_t
beta_t, beta_s = self.beta_s[step + 1], self.beta_s[step]
gamma_s = self.gamma_t[step]
f = self.f_t[step]
g = self.g_t[step]
noise = torch.randn_like(x_s, device=x_s.device)
if last_drop == True and step == self.num_steps - 1:
return x0_pred, x0_pred
else:
# bfn
# x_t = ((beta_t - beta_s) / ((beta_t + 1) * gamma_s) + (beta_s + 1)/(beta_t + 1)) * x_s - (beta_t - beta_s) / (beta_t + 1) * ((1-gamma_s)/gamma_s)**0.5 * noise_pred + (beta_t-beta_s)**0.5/(beta_t + 1) * noise
# sde
x_t = x_s - (f * x_s + g**2 * noise_pred / (gamma_s * (1-gamma_s))**0.5) * self.delta_t + g * self.delta_t**0.5 * noise
return x_t, x0_pred
def ode_bfnsolver1_update(self, x_s, step, last_drop=False):
# x_s -> x_t
t = torch.ones_like(x_s, device=x_s.device) * (1 - self.times[step])
# noise predict and x0 predict
with torch.no_grad():
noise_pred = self.unet(x_s, t).reshape(x_s.shape)
alpha_t, sigma_t = self.alpha_t[step], self.sigma_t[step]
x0_pred = (x_s - sigma_t * noise_pred) / alpha_t
# clip x0
x0_pred = x0_pred.clip(min=-1.0, max=1.0)
noise_pred = (x_s - x0_pred * alpha_t) / sigma_t
# get schedule
lambda_t, lambda_s = self.lambda_t[step + 1], self.lambda_t[step]
alpha_t, alpha_s = self.alpha_t[step + 1], self.alpha_t[step]
sigma_t, sigma_s = self.sigma_t[step + 1], self.sigma_t[step]
h = lambda_t - lambda_s
if last_drop == True and step == self.num_steps - 1:
return x0_pred, x0_pred
else:
x_t = (alpha_t / alpha_s) * x_s - (sigma_t * (torch.exp(h) - 1.0)) * noise_pred
return x_t, x0_pred
def ode_bfnsolver2_multi_step_update(self, x_s0, step, x0_pred_last=None, last_drop=False):
lambda_t, lambda_s0 = self.lambda_t[step + 1], self.lambda_t[step],
alpha_t, alpha_s0 = self.alpha_t[step + 1], self.alpha_t[step]
sigma_t, sigma_s0 = self.sigma_t[step + 1], self.sigma_t[step]
h = lambda_t - lambda_s0
timestep_s0 = torch.ones_like(x_s0, device=x_s0.device) * (1 - self.times[step])
with torch.no_grad():
noise_pred_s0 = self.unet(x_s0, timestep_s0).reshape(x_s0.shape)
x0_pred_s0 = (x_s0 - sigma_s0 * noise_pred_s0) / alpha_s0
x0_pred_s0 = x0_pred_s0.clip(-1, 1)
if step == 0:
x_t = (sigma_t / sigma_s0) * x_s0 - alpha_t * (torch.exp(-h) - 1) * x0_pred_s0
return x_t, x0_pred_s0
elif last_drop == True and step == self.num_steps - 1:
return x0_pred_s0, x0_pred_s0
else:
lambda_s1 = self.lambda_t[step - 1]
h_0 = lambda_s0 - lambda_s1
r = h_0 / h
D = (1 + 1 / (2 * r)) * x0_pred_s0 - 1 / (2 * r) * x0_pred_last
x_t = sigma_t / sigma_s0 * x_s0 - alpha_t * (torch.exp(-h) - 1) * D
return x_t, x0_pred_s0
def sde_bfnsolver2_multi_step_update(self, x_s0, step, x0_pred_last=None, last_drop=False):
lambda_t, lambda_s0 = self.lambda_t[step + 1], self.lambda_t[step],
alpha_t, alpha_s0 = self.alpha_t[step + 1], self.alpha_t[step]
sigma_t, sigma_s0 = self.sigma_t[step + 1], self.sigma_t[step]
h = lambda_t - lambda_s0
timestep_s0 = torch.ones_like(x_s0, device=x_s0.device) * (1 - self.times[step])
with torch.no_grad():
noise_pred_s0 = self.unet(x_s0, timestep_s0).reshape(x_s0.shape)
x0_pred_s0 = (x_s0 - sigma_s0 * noise_pred_s0) / alpha_s0
x0_pred_s0 = x0_pred_s0.clip(-1, 1)
noise = torch.randn_like(x_s0, device=x_s0.device)
if step == 0:
x_t = (sigma_t / sigma_s0 * torch.exp(-h)) * x_s0 + \
alpha_t * (1 - torch.exp(-2.0 * h)) * x0_pred_s0 + sigma_t * torch.sqrt(
1.0 - torch.exp(-2 * h)) * noise
return x_t, x0_pred_s0
elif last_drop == True and step == self.num_steps - 1:
return x0_pred_s0, x0_pred_s0
else:
lambda_s1 = self.lambda_t[step - 1]
h_0 = lambda_s0 - lambda_s1 # 和paper相反，和diffusers一致
r = h_0 / h
D1 = (x0_pred_s0 - x0_pred_last) / r
x_t = (sigma_t / sigma_s0 * torch.exp(-h)) * x_s0 + alpha_t * (1 - torch.exp(-2.0 * h)) * x0_pred_s0 \
+ 0.5 * alpha_t * (1 - torch.exp(-2.0 * h)) * D1 + sigma_t * torch.sqrt(
1.0 - torch.exp(-2.0 * h)) * noise
return x_t, x0_pred_s0
@torch.inference_mode()
def sample(cfg: DictConfig):
train_cfg = make_config(cfg.config_file)
unet = get_nnet(**train_cfg.net)
if torch.cuda.is_available():
unet.cuda()
cfg = OmegaConf.merge(train_cfg.sampling, cfg)
state_dict = torch.load(cfg.load_model, map_location='cpu')
new_state_dict = {k[len("net."):]: v for k, v in state_dict.items() if k.startswith("net.")}
unet.load_state_dict(new_state_dict)
unet.eval()
logging.info(f"Loaded model from {cfg.load_model}")
seed_everything(cfg.seed)
logging.info(f"Seeded everything with seed {cfg.seed}")
logging.info(f"sample with {cfg.algorithm}")
logging.info(f"number of samples: {cfg.batch_size}, n_steps: {cfg.n_steps}, last_drop: {cfg.last_drop}, eta: {cfg.eta}, initial_dist: {cfg.initial_dist}")
samples_shape = (cfg.batch_size, train_cfg.data.image_size, train_cfg.data.image_size, train_cfg.data.image_channels)
min_variance = train_cfg.bayesian_flow.min_variance
# BFN solvers
bfnsolver = ImageBFNSolver(unet, num_steps=cfg.n_steps, min_variance=min_variance, eta=cfg.eta)
samples = batch_sample(cfg, train_cfg, unet, bfnsolver, samples_shape, min_variance)
cfg.save_path = f"./samples/cifar/{cfg.algorithm}/n_steps={cfg.n_steps}"
logging.info(f"save samples to {os.path.dirname(cfg.save_path)}")
os.makedirs(os.path.dirname(cfg.save_path), exist_ok=True)
save_batch(samples, cfg.save_path)
def batch_sample(cfg: DictConfig, train_cfg: DictConfig, unet: torch.nn.Module, bfnsolver: ImageBFNSolver, samples_shape: tuple, min_variance: float):
# Choices of the Initialization Distribution
if cfg.initial_dist == "zero_mean_and_std":
gamma_t = 1 - min_variance**cfg.eta
std_t = (gamma_t * (1 - gamma_t))**0.5
xt = torch.randn(*samples_shape).to(next(unet.parameters()).device) * std_t
elif cfg.initial_dist == "optimal_mean_and_std":
saved_tensors = torch.load(cfg.mean_std_path)
mean_x = saved_tensors['mean'].to(next(unet.parameters()).device)
std_x = saved_tensors['std'].to(next(unet.parameters()).device)
trace_x = std_x.square().sum()
gamma_t = 1 - min_variance**cfg.eta
mean_t = gamma_t * mean_x
std_t = (gamma_t * (1 - gamma_t) + gamma_t**2 * trace_x / (train_cfg.data.image_size * train_cfg.data.image_size * train_cfg.data.image_channels))**0.5
xt = torch.randn(*samples_shape).to(next(unet.parameters()).device) * std_t + mean_t
data_pred_last = None
for step in tqdm(range(cfg.n_steps)):
if cfg.algorithm == "ode_euler":
pass
if cfg.algorithm == "sde_euler":
xt, _ = bfnsolver.sde_euler_update(xt, step, last_drop=cfg.last_drop)
elif cfg.algorithm == "ode_bfnsolver1":
xt, _ = bfnsolver.ode_bfnsolver1_update(xt, step, last_drop=cfg.last_drop)
elif cfg.algorithm == 'ode_bfnsolver2_multi_step':
xt, data_pred_last = bfnsolver.ode_bfnsolver2_multi_step_update(xt, step, x0_pred_last=data_pred_last, last_drop=cfg.last_drop)
elif cfg.algorithm == 'sde_bfnsolver2_multi_step':
xt, data_pred_last = bfnsolver.sde_bfnsolver2_multi_step_update(xt, step, x0_pred_last=data_pred_last, last_drop=cfg.last_drop)
else:
raise NotImplementedError(cfg.algorithm)
samples = xt
return samples
def save_batch(samples: torch.Tensor, save_path: str) -> None:
import numpy as np
from PIL import Image
samples = (samples.float() + 1) / 2 * 255
# samples = ((samples.float() + 1) / 2 ).clamp(0, 1)* 255
np_images = samples.cpu().detach().numpy().astype(np.uint8)
if np_images.shape[0] <= 10:
new_image = Image.fromarray(np.concatenate(np_images, axis=1))
else:
img_size = np_images.shape[1]
new_image = Image.new('RGB', (10*img_size, np_images.shape[0] // 10*img_size))
locations = [(i % 10 * img_size, i // 10 * img_size) for i in range(np_images.shape[0])]
for (x, y), np_image in zip(locations, np_images):
img = Image.fromarray(np_image)
new_image.paste(img, (x, y))
new_image.save(f"{save_path}.png")
if __name__ == "__main__":
config = OmegaConf.from_cli()
sample(config)