forked from AndyCao1125/SDDPM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_SDDPM.py
350 lines (291 loc) · 14.3 KB
/
main_SDDPM.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
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
import copy
import json
import os
import warnings
import numpy as np
import wandb
# from data import ImageNet,LSUNBed
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.datasets import CIFAR10
from torchvision.utils import make_grid, save_image
from torchvision import transforms
import torchvision
from spikingjelly.activation_based import neuron, functional, surrogate, layer
from tqdm import trange
import random
from diffusion import GaussianDiffusionTrainer,GaussianDiffusionSampler,LatentGaussianDiffusionTrainer,LatentGaussianDiffusionSampler
from model import Spk_UNet
from score.both import get_inception_and_fid_score
# os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3,4,5,6,7,8"
#os.environ["CUDA_VISIBLE_DEVICES"] = "0,1"
## argument parsing ##
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--seed', default=42, type=int, help='seed')
parser.add_argument('--train', action='store_true', default=False, help='train from scratch')
parser.add_argument('--eval', action='store_true', default=False, help='load ckpt.pt and evaluate FID and IS')
parser.add_argument('--dataset', type=str, default='cifar10', help='dataset name')
parser.add_argument('--sample_type', type=str, default='ddpm', help='Sample Type')
parser.add_argument('--wandb', action='store_true', default=False, help='use wandb to log training')
# Spiking UNet
parser.add_argument('--ch', default=128, type=int, help='base channel of UNet')
parser.add_argument('--ch_mult', default=[1, 2, 2, 4], help='channel multiplier')
parser.add_argument('--attn', default=[], help='add attention to these levels')
parser.add_argument('--num_res_blocks', default=2, type=int, help='# resblock in each level')
parser.add_argument('--img_size', default=32, type=int, help='image size')
parser.add_argument('--dropout', default=0.1, type=float, help='dropout rate of resblock')
parser.add_argument('--timestep', default=4, type=int, help='snn timestep')
parser.add_argument('--img_ch', type=int, default=3, help='image channel')
# Gaussian Diffusion
parser.add_argument('--beta_1', default=1e-4, type=float, help='start beta value')
parser.add_argument('--beta_T', default=0.02, type=float, help='end beta value')
parser.add_argument('--T', default=1000, type=int, help='total diffusion steps')
parser.add_argument('--mean_type', default='epsilon', help='predict variable:[xprev, xstart, epsilon]')
parser.add_argument('--var_type', default='fixedlarge', help='variance type:[fixedlarge, fixedsmall]')
# Training
parser.add_argument('--resume', default=False, help="load pre-trained model")
parser.add_argument('--resume_model', type=str, help='resume model path')
parser.add_argument('--lr', default=2e-4, help='target learning rate')
parser.add_argument('--grad_clip', default=1., help="gradient norm clipping")
parser.add_argument('--total_steps', type=int, default=500000, help='total training steps')
parser.add_argument('--warmup', default=5000, help='learning rate warmup')
parser.add_argument('--batch_size', type=int, default=32, help='batch size')
parser.add_argument('--num_workers', type=int, default=4, help='workers of Dataloader')
parser.add_argument('--ema_decay', default=0.9999, help="ema decay rate")
parser.add_argument('--parallel', default=True, help='multi gpu training')
# Logging & Sampling
parser.add_argument('--logdir', default='./log', help='log directory')
parser.add_argument('--sample_size', type=int,default=64, help="sampling size of images")
parser.add_argument('--sample_step', type=int,default=5000, help='frequency of sampling')
# Evaluation
parser.add_argument('--save_step', type=int,default=0, help='frequency of saving checkpoints, 0 to disable during training')
parser.add_argument('--eval_step', type=int,default=0, help='frequency of evaluating model, 0 to disable during training')
parser.add_argument('--num_images', type=int,default=50000, help='the number of generated images for evaluation')
parser.add_argument('--fid_use_torch', default=True, help='calculate IS and FID on gpu')
parser.add_argument('--fid_cache', default='./stats/cifar10.train.npz', help='FID cache')
parser.add_argument('--num_step', type=int,default=1000, help='number of sampling steps')
parser.add_argument('--pre_trained_path', default='./pth/1224_4T.pt', help='FID cache')
args = parser.parse_args()
device = torch.device('cuda:0')
def seed_everything(seed_value):
random.seed(seed_value)
np.random.seed(seed_value)
torch.manual_seed(seed_value)
os.environ['PYTHONHASHSEED'] = str(seed_value)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed_value)
torch.cuda.manual_seed_all(seed_value)
torch.backends.cudnn.deterministic = True
def ema(source, target, decay):
source_dict = source.state_dict()
target_dict = target.state_dict()
for key in source_dict.keys():
target_dict[key].data.copy_(
target_dict[key].data * decay +
source_dict[key].data * (1 - decay))
def infiniteloop(dataloader):
while True:
for x, y in iter(dataloader):
yield x
def warmup_lr(step):
return min(step, args.warmup) / args.warmup
def evaluate(sampler, model):
model.eval()
with torch.no_grad():
images = []
desc = "generating images"
for i in trange(0, args.num_images, args.batch_size, desc=desc):
batch_size = min(args.batch_size, args.num_images - i)
x_T = torch.randn((batch_size, args.img_ch, args.img_size, args.img_size))
batch_images = sampler(x_T.to(device)).cpu()
images.append((batch_images + 1) / 2)
grid = (make_grid(batch_images[:64,...]) + 1) / 2
save_image(grid, 'ddpm.png')
images = torch.cat(images, dim=0).numpy()
model.train()
(IS, IS_std), FID = get_inception_and_fid_score(
images, args.fid_cache, num_images=args.num_images,
use_torch=args.fid_use_torch, verbose=True)
return (IS, IS_std), FID, images
def train():
if args.dataset == 'cifar10':
dataset = CIFAR10(
root='/home/dataset/Cifar10', train=True, download=False,
transform=transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]))
elif args.dataset == 'celeba':
SetRange = torchvision.transforms.Lambda(lambda X: 2 * X - 1.)
transform = torchvision.transforms.Compose([
torchvision.transforms.RandomHorizontalFlip(),
torchvision.transforms.CenterCrop(148),
torchvision.transforms.Resize((64,64)),
torchvision.transforms.ToTensor(),
SetRange])
dataset = torchvision.datasets.ImageFolder(root='/home/dataset/CelebA/celeba',
transform=transform)
elif args.dataset == 'fashion-mnist':
SetRange = torchvision.transforms.Lambda(lambda X: 2 * X - 1.)
transform = torchvision.transforms.Compose([
torchvision.transforms.Resize((32,32)),
torchvision.transforms.ToTensor(),
SetRange])
dataset = torchvision.datasets.FashionMNIST(root='/home/dataset/FashionMnist',
train=True,
download=True,
transform=transform)
elif args.dataset == 'mnist':
SetRange = torchvision.transforms.Lambda(lambda X: 2 * X - 1.)
transform = torchvision.transforms.Compose([
torchvision.transforms.Resize((32,32)),
torchvision.transforms.ToTensor(),
SetRange])
dataset = torchvision.datasets.MNIST(root='/home/dataset/Mnist',
train=True,
download=True,
transform=transform)
elif args.dataset == 'lsun':
dataset = LSUNBed()
else:
raise NotImplementedError
dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.num_workers, drop_last=True)
datalooper = infiniteloop(dataloader)
print(f'-------Starting loading {args.dataset} Dataset!-------')
# model setup
net_model = Spk_UNet(
T=args.T, ch=args.ch, ch_mult=args.ch_mult, attn=args.attn,
num_res_blocks=args.num_res_blocks, dropout=args.dropout, timestep=args.timestep, img_ch=args.img_ch)
optim = torch.optim.Adam(net_model.parameters(), lr=args.lr)
if args.resume:
ckpt = torch.load(os.path.join(args.resume_model))
print(f'Loading Resume model from {args.resume_model}')
net_model.load_state_dict(ckpt['net_model'], strict=True)
else:
print('Training from scratch')
trainer = GaussianDiffusionTrainer(
net_model, float(args.beta_1), float(args.beta_T), args.T).to(device)
net_sampler = GaussianDiffusionSampler(
net_model, float(args.beta_1), float(args.beta_T), args.T, args.img_size,
args.mean_type, args.var_type).to(device)
if args.parallel:
trainer = torch.nn.DataParallel(trainer)
net_sampler = torch.nn.DataParallel(net_sampler).cuda()
# log setup
if not os.path.exists(os.path.join(args.logdir,'sample')):
os.makedirs(os.path.join(args.logdir, 'sample'))
x_T = torch.randn(int(args.sample_size), int(args.img_ch), int(args.img_size), int(args.img_size))
x_T = x_T.to(device)
grid = (make_grid(next(iter(dataloader))[0][:args.sample_size]) + 1) / 2
save_image(grid, os.path.join(args.logdir,'sample','groundtruth.png'))
# show model size
model_size = 0
for param in net_model.parameters():
model_size += param.data.nelement()
print('Model params: %.2f M' % (model_size / 1024 / 1024))
# start training
with trange(args.total_steps, dynamic_ncols=True) as pbar:
for step in pbar:
# train
optim.zero_grad()
x_0 = next(datalooper).to(device)
loss = trainer(x_0.float()).mean()
loss.backward()
if args.wandb:
wandb.log({'training loss': loss.item()})
torch.nn.utils.clip_grad_norm_(
net_model.parameters(), args.grad_clip)
optim.step()
pbar.set_postfix(loss='%.3f' % loss)
## reset SNN neuron
functional.reset_net(net_model)
# sample
# print(f'Sample at {step} step')
if args.sample_step > 0 and step % args.sample_step == 0:
net_model.eval()
with torch.no_grad():
x_0 = net_sampler(x_T)
grid = (make_grid(x_0) + 1) / 2
path = os.path.join(
args.logdir, 'sample', '%d.png' % step)
save_image(grid, path)
## log to wandb
if args.wandb:
wandb.log({'sample': [wandb.Image(grid, caption='sample')]})
net_model.train()
# save
# print(f'Save model at {step} step')
if args.save_step > 0 and step % args.save_step == 0 and step > 0:
ckpt = {
'net_model': net_model.state_dict(),
'optim': optim.state_dict(),
'step': step,
'x_T': x_T,
}
save_path = str(step) + 'ckpt.pt'
torch.save(ckpt, os.path.join(args.logdir,save_path))
# evaluate
# print(f'Evaluate at {step} step')
if args.eval_step > 0 and step % args.eval_step == 0 and step > 0:
net_IS, net_FID, _ = evaluate(net_sampler, net_model)
metrics = {
'IS': net_IS[0],
'IS_std': net_IS[1],
'FID': net_FID,
}
pbar.write(
"%d/%d " % (step, args.total_steps) +
", ".join('%s:%.3f' % (k, v) for k, v in metrics.items()))
with open(os.path.join(args.logdir, 'eval.txt'), 'a') as f:
metrics['step'] = step
f.write(json.dumps(metrics) + "\n")
def eval():
# model setup
model = Spk_UNet(
T=args.T, ch=args.ch, ch_mult=args.ch_mult, attn=args.attn,
num_res_blocks=args.num_res_blocks, dropout=args.dropout, timestep=args.timestep, img_ch=args.img_ch)
ckpt_path = args.pre_trained_path
ckpt1 = torch.load(ckpt_path)['net_model']
print(f'Successfully load checkpoint!')
model.load_state_dict(ckpt1)
model.eval()
sampler = GaussianDiffusionSampler(
model, float(args.beta_1), float(args.beta_T), args.T, img_size=int(args.img_size),
mean_type=args.mean_type, var_type=args.var_type,sample_type=args.sample_type,sample_steps=args.num_step).to(device)
if args.parallel:
sampler = torch.nn.DataParallel(sampler)
with torch.no_grad():
images = []
desc = "generating images"
for i in trange(0, args.num_images, args.batch_size, desc=desc):
batch_size = min(args.batch_size, args.num_images - i)
x_T = torch.randn((batch_size, int(args.img_ch), int(args.img_size), int(args.img_size)))
batch_images = sampler(x_T.to(device))
batch_images = batch_images.cpu()
images.append((batch_images + 1) / 2)
images = torch.cat(images, dim=0).numpy()
print(images.shape)
(IS, IS_std), FID = get_inception_and_fid_score(
images, args.fid_cache, num_images=args.num_images,
use_torch=args.fid_use_torch, verbose=True)
print(f'IS: {IS}, IS_std: {IS_std}, FID: {FID}')
def main():
if args.wandb:
## wandb init ##
wandb.init(project="spike_diffusion", name=str(args.dataset)+str(args.sample_type))
# suppress annoying inception_v3 initialization warning #
warnings.simplefilter(action='ignore', category=FutureWarning)
seed_everything(42)
if args.train:
train()
if args.eval:
eval()
if not args.train and not args.eval:
print('Add --train and/or --eval to execute corresponding tasks')
if __name__ == '__main__':
# app.run(main)
main()