-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
152 lines (132 loc) · 5.77 KB
/
main.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
import torch
import torch.nn.functional as F
from torch import nn
from tqdm import tqdm
import matplotlib.pyplot as plt
import argparse
import os
from survae.distributions import DataParallelDistribution
from data import get_data_loaders, reconstruct, save_plt_img
from model import get_model
from schedular import LinearWarmupScheduler
parser = argparse.ArgumentParser()
parser.add_argument('-p','--data_path', type=str, default='tiny-imagenet.zip')
parser.add_argument('--param_path', type=str, default='models/')
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--output_dir', type=str, default='imgs_out/')
parser.add_argument('--num_epoch', type=int, default=16)
parser.add_argument('--img_size', type=int, default=256)
parser.add_argument('--latent_size', type=int, default=2)
parser.add_argument('--exp_name', type=str, default='tmp')
parser.add_argument('--resume', type=str, default='')
parser.add_argument('--lr', type=float, default=1e-2)
parser.add_argument('--warmup', type=float, default=1000)
parser.add_argument('--vae', action='store_true')
parser.add_argument('--rej', action='store_true')
parser.add_argument('--adam', action='store_true')
parser.add_argument('--no_pretrain', action='store_true')
parser.add_argument('--vis_mode', type=str, default='tensorboard', help='one of [tensorboard, plt, wandb]')
parser.add_argument('--dataset', type=str, default='tinyImgNet', help='one of [tinyImgNet, tinyImgNetZip, COCO]')
device = 'cuda' if torch.cuda.is_available() else 'cpu'
def log_img(model, args, wandb, writer):
with torch.no_grad():
lab = torch.cat([X_test, model.sample(X_test)], 1)
img = reconstruct(lab)
if args.vis_mode == 'tensorboard':
writer.add_images("result", img.transpose(0, 3, 1, 2), gIter)
elif args.vis_mode == 'wandb':
wandb.log({'result': [wandb.Image(i) for i in img]})
else:
save_plt_img(img, title='result')
l = lab[:1, 0].repeat(64, 1, 1, 1)
z = model.module.prior.sample(1).repeat(64,1)
z_ = torch.meshgrid(torch.linspace(-2, 2, 8), torch.linspace(-2, 2, 8))
z_ = torch.stack(z_, -1).flatten(0, 1).to(device)
z[:,:2] = z_
lab = torch.cat([l, model.module.transform(z, l)], 1)
img = reconstruct(lab)
if args.vis_mode == 'tensorboard':
writer.add_images("sample", img.transpose(0, 3, 1, 2), gIter)
elif args.vis_mode == 'wandb':
wandb.log({'sample':[wandb.Image(i) for i in img]})
else:
save_plt_img(img, title='sample')
log_iters = [25, 50, 100, 200, 400, 800, 1600]
if __name__=='__main__':
############
## Data ##
############
args = parser.parse_args()
os.makedirs(args.param_path, exist_ok=True)
torch.manual_seed(0)
tr_loader, va_loader = get_data_loaders(args.batch_size, args.dataset, args.img_size)
#############
## Model ##
#############
model = get_model(vae=args.vae, rej=args.rej, latent_size=args.latent_size).to(device)
model = DataParallelDistribution(model)
if args.resume:
model.load_state_dict(torch.load(args.resume), strict=False)
if args.adam:
optim = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-5)
else:
optim = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=0.9, weight_decay=1e-4)
sched = LinearWarmupScheduler(optim, args.warmup, [
args.num_epoch * 7 * len(tr_loader) // 10, args.num_epoch * 9 * len(tr_loader) // 10])
###############
## Logging ##
###############
if args.vis_mode == 'tensorboard':
from tensorboardX import SummaryWriter
writer = SummaryWriter(flush_secs=30)
wandb = None
elif args.vis_mode == 'wandb':
import wandb
wandb.init(project='colorvae')
wandb.config.update(args)
wandb.watch(model)
writer = None
gIter = 0
X_test, y_test = next(iter(va_loader))
X_test = X_test.to('cuda')
y_test = y_test.to('cuda')
for epoch in range(args.num_epoch):
cum_loss = 0.0
pbar = tqdm(tr_loader)
for i, (l, ab) in enumerate(pbar):
l = l.to(device)
ab = ab.to(device)
loss = -model.log_prob(ab, l).mean() / (args.img_size * args.img_size * 2)
optim.zero_grad()
loss.backward()
optim.step()
sched.step()
cum_loss += loss.item()
pbar.set_description_str(f"Epoch {epoch}, nll {cum_loss / (i+1):.4f}")
if args.vis_mode == 'tensorboard':
writer.add_scalar("Train/nll", loss, gIter)
if args.rej:
writer.add_scalar("Train/rej", model.module.rej_prob, gIter)
elif args.vis_mode == 'wandb':
logs = {"Train/nll": loss}
if args.rej:
logs.update({"Train/rej": model.module.rej_prob})
wandb.log(logs)
if gIter in log_iters:
log_img(model, args, wandb, writer)
gIter += 1
with torch.no_grad():
cum_loss = 0.0
pbar = tqdm(va_loader)
for i, (l, ab) in enumerate(pbar):
l = l.to(device)
ab = ab.to(device)
loss = -model.log_prob(ab, l).mean() / (args.img_size * args.img_size * 2)
cum_loss += loss.item()
pbar.set_description_str(f"Test nll {cum_loss / (i+1):.4f}")
if args.vis_mode == 'tensorboard':
writer.add_scalar("Val/nll", cum_loss / len(va_loader), gIter)
elif args.vis_mode == 'wandb':
wandb.log({"Val/nll": cum_loss / len(va_loader)})
log_img(model, args, wandb, writer)
torch.save(model.state_dict(), os.path.join(args.param_path, args.exp_name+'_model.pt'))