forked from dome272/Diffusion-Models-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdiffusion.py
183 lines (157 loc) · 7.59 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
import os
import copy
import numpy as np
import torch
import torch.nn as nn
from tqdm import tqdm
from torch import optim
from utils import *
from modules import UNet_conditional, EMA
import logging
from torch.utils.tensorboard import SummaryWriter
logging.basicConfig(format="%(asctime)s - %(levelname)s: %(message)s", level=logging.INFO, datefmt="%I:%M:%S")
class Diffusion:
def __init__(self, noise_steps=1000, beta_start=1e-4, beta_end=0.02, img_size=256, device="cuda"):
self.noise_steps = noise_steps
self.beta_start = beta_start
self.beta_end = beta_end
self.img_size = img_size
self.device = device
self.beta = self.prepare_noise_schedule().to(device)
self.alpha = 1. - self.beta
self.alpha_hat = torch.cumprod(self.alpha, dim=0)
def prepare_noise_schedule(self):
return torch.linspace(self.beta_start, self.beta_end, self.noise_steps)
def noise_images(self, x, t):
sqrt_alpha_hat = torch.sqrt(self.alpha_hat[t])[:, None, None, None]
sqrt_one_minus_alpha_hat = torch.sqrt(1 - self.alpha_hat[t])[:, None, None, None]
Ɛ = torch.randn_like(x)
return sqrt_alpha_hat * x + sqrt_one_minus_alpha_hat * Ɛ, Ɛ
def sample_timesteps(self, n):
return torch.randint(low=1, high=self.noise_steps, size=(n,))
def sample(self, model, n, channels):
logging.info(f"Sampling {n} new images....")
model.eval()
with torch.no_grad():
x = torch.randn((n, channels, self.img_size, self.img_size)).to(self.device)
for i in tqdm(reversed(range(1, self.noise_steps)), position=0):
t = (torch.ones(n) * i).long().to(self.device)
predicted_noise = model(x, t)
alpha = self.alpha[t][:, None, None, None]
alpha_hat = self.alpha_hat[t][:, None, None, None]
beta = self.beta[t][:, None, None, None]
if i > 1:
noise = torch.randn_like(x)
else:
noise = torch.zeros_like(x)
x = 1 / torch.sqrt(alpha) * (x - ((1 - alpha) / (torch.sqrt(1 - alpha_hat))) * predicted_noise) + torch.sqrt(beta) * noise
model.train()
x = (x.clamp(-1, 1) + 1) / 2
x = (x * 255).type(torch.uint8)
return x
class ConditionalDiffusion:
def __init__(self, noise_steps=1000, beta_start=1e-4, beta_end=0.02, img_size=256, device="cuda"):
self.noise_steps = noise_steps
self.beta_start = beta_start
self.beta_end = beta_end
self.beta = self.prepare_noise_schedule().to(device)
self.alpha = 1. - self.beta
self.alpha_hat = torch.cumprod(self.alpha, dim=0)
self.img_size = img_size
self.device = device
def prepare_noise_schedule(self):
return torch.linspace(self.beta_start, self.beta_end, self.noise_steps)
def noise_images(self, x, t):
sqrt_alpha_hat = torch.sqrt(self.alpha_hat[t])[:, None, None, None]
sqrt_one_minus_alpha_hat = torch.sqrt(1 - self.alpha_hat[t])[:, None, None, None]
Ɛ = torch.randn_like(x)
return sqrt_alpha_hat * x + sqrt_one_minus_alpha_hat * Ɛ, Ɛ
def sample_timesteps(self, n):
return torch.randint(low=1, high=self.noise_steps, size=(n,))
def sample(self, model, n, channels, labels, cfg_scale=3):
logging.info(f"Sampling {n} new images....")
model.eval()
with torch.no_grad():
x = torch.randn((n, channels, self.img_size, self.img_size)).to(self.device)
for i in tqdm(reversed(range(1, self.noise_steps)), position=0):
t = (torch.ones(n) * i).long().to(self.device)
predicted_noise = model(x, t, labels)
if cfg_scale > 0:
uncond_predicted_noise = model(x, t, None)
predicted_noise = torch.lerp(uncond_predicted_noise, predicted_noise, cfg_scale)
alpha = self.alpha[t][:, None, None, None]
alpha_hat = self.alpha_hat[t][:, None, None, None]
beta = self.beta[t][:, None, None, None]
if i > 1:
noise = torch.randn_like(x)
else:
noise = torch.zeros_like(x)
x = 1 / torch.sqrt(alpha) * (x - ((1 - alpha) / (torch.sqrt(1 - alpha_hat))) * predicted_noise) + torch.sqrt(beta) * noise
model.train()
x = (x.clamp(-1, 1) + 1) / 2
x = (x * 255).type(torch.uint8)
return x
# def train(args):
# setup_logging(args.run_name)
# device = args.device
# dataloader = get_data(args)
# model = UNet_conditional(num_classes=args.num_classes).to(device)
# optimizer = optim.AdamW(model.parameters(), lr=args.lr)
# mse = nn.MSELoss()
# diffusion = Diffusion(img_size=args.image_size, device=device)
# logger = SummaryWriter(os.path.join("runs", args.run_name))
# l = len(dataloader)
# ema = EMA(0.995)
# ema_model = copy.deepcopy(model).eval().requires_grad_(False)
# for epoch in range(args.epochs):
# logging.info(f"Starting epoch {epoch}:")
# pbar = tqdm(dataloader)
# for i, (images, labels) in enumerate(pbar):
# images = images.to(device)
# labels = labels.to(device)
# t = diffusion.sample_timesteps(images.shape[0]).to(device)
# x_t, noise = diffusion.noise_images(images, t)
# if np.random.random() < 0.1:
# labels = None
# predicted_noise = model(x_t, t, labels)
# loss = mse(noise, predicted_noise)
# optimizer.zero_grad()
# loss.backward()
# optimizer.step()
# ema.step_ema(ema_model, model)
# pbar.set_postfix(MSE=loss.item())
# logger.add_scalar("MSE", loss.item(), global_step=epoch * l + i)
# if epoch % 10 == 0:
# labels = torch.arange(10).long().to(device)
# sampled_images = diffusion.sample(model, n=len(labels), labels=labels)
# ema_sampled_images = diffusion.sample(ema_model, n=len(labels), labels=labels)
# plot_images(sampled_images)
# save_images(sampled_images, os.path.join("results", args.run_name, f"{epoch}.jpg"))
# save_images(ema_sampled_images, os.path.join("results", args.run_name, f"{epoch}_ema.jpg"))
# torch.save(model.state_dict(), os.path.join("models", args.run_name, f"ckpt.pt"))
# torch.save(ema_model.state_dict(), os.path.join("models", args.run_name, f"ema_ckpt.pt"))
# torch.save(optimizer.state_dict(), os.path.join("models", args.run_name, f"optim.pt"))
# def launch():
# import argparse
# parser = argparse.ArgumentParser()
# args = parser.parse_args()
# args.run_name = "DDPM_conditional"
# args.epochs = 300
# args.batch_size = 14
# args.image_size = 64
# args.num_classes = 10
# args.dataset_path = r"C:\Users\dome\datasets\cifar10\cifar10-64\train"
# args.device = "cuda"
# args.lr = 3e-4
# train(args)
# if __name__ == '__main__':
# launch()
# # device = "cuda"
# # model = UNet_conditional(num_classes=10).to(device)
# # ckpt = torch.load("./models/DDPM_conditional/ckpt.pt")
# # model.load_state_dict(ckpt)
# # diffusion = Diffusion(img_size=64, device=device)
# # n = 8
# # y = torch.Tensor([6] * n).long().to(device)
# # x = diffusion.sample(model, n, y, cfg_scale=0)
# # plot_images(x)