-
Notifications
You must be signed in to change notification settings - Fork 15
/
prior_distillation.py
198 lines (163 loc) · 9.1 KB
/
prior_distillation.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
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from torchvision import utils
import matplotlib.pyplot as plt
import numpy as np
import copy
import argparse
import math
from tqdm import tqdm
from model.content_encoder import AutoEncoder
from model.vgg import VGGLoss
from dataset import create_imagemasklabel_dataloader, natural_sort
class TrainOptions():
def __init__(self):
self.parser = argparse.ArgumentParser(description="Train Prior Distillation of GP-UNIT")
self.parser.add_argument("--task", type=str, default='prior_distillation', help="task name")
self.parser.add_argument("--lr", type=float, default=0.0002, help="learning rate")
self.parser.add_argument("--iter", type=int, default=45000, help="iterations")
self.parser.add_argument("--batch", type=int, default=16, help="batch size")
self.parser.add_argument("--lambda_app_rec", type=float, default=1.0, help="the weight of appearance distant loss")
self.parser.add_argument("--lambda_shape_rec", type=float, default=5.0, help="the weight of shape reconstruction loss")
self.parser.add_argument("--lambda_reg", type=float, default=0.001, help="the weight of regularization loss")
self.parser.add_argument("--lambda_class", type=float, default=1.0, help="the weight of classification loss")
self.parser.add_argument("--lambda_feat_dist", type=float, default=1.0, help="the weight of feature distant loss")
self.parser.add_argument("--lambda_shape_dist", type=float, default=5.0, help="the weight of shape distant loss")
self.parser.add_argument("--paired_data_root", type=str, help="the path to the synImageNet291")
self.parser.add_argument("--unpaired_data_root", type=str, help="the path to the ImageNet291 and CelebA-HQ")
self.parser.add_argument("--paired_mask_root", type=str, help="the path to the synImageNet291_mask")
self.parser.add_argument("--unpaired_mask_root", type=str, help="the path to the ImageNet291_mask and CelebA-HQ_mask")
self.parser.add_argument("--save_every", type=int, default=5000, help="interval of saving a checkpoint")
self.parser.add_argument("--save_begin", type=int, default=30000, help="when to start saving a checkpoint")
self.parser.add_argument("--visualize_every", type=int, default=500, help="interval of saving an intermediate result")
self.parser.add_argument("--model_path", type=str, default='./checkpoint/', help="path to the saved models")
def parse(self):
self.opt = self.parser.parse_args()
args = vars(self.opt)
print('Load options')
for name, value in sorted(args.items()):
print('%s: %s' % (str(name), str(value)))
return self.opt
def train(args, udataloader, pdataloader, netAE, optimizer_AE, vgg_loss, device='cuda'):
pbar = tqdm(range(args.iter), initial=0, smoothing=0.01, ncols=120, dynamic_ncols=False)
netAE.train()
piterator = iter(pdataloader)
uiterator = iter(udataloader)
for idx in pbar:
try:
pdata = next(piterator)
except StopIteration:
piterator = iter(pdataloader)
pdata = next(piterator)
try:
udata = next(uiterator)
except StopIteration:
uiterator = iter(udataloader)
udata = next(uiterator)
la, xa, lb, xb, ma, mb = pdata['label'], pdata['image'], pdata['labelB'], pdata['imageB'], pdata['mask'], pdata['maskB']
l, x, m = udata['label'], udata['image'], udata['mask']
la, lb, l = la.to(device), lb.to(device), l.to(device)
xa, xb, x = xa.to(device), xb.to(device), x.to(device)
m = F.interpolate(m.to(device), size=16, mode='bilinear')
ma = F.interpolate(ma.to(device), size=16, mode='bilinear')
mb = F.interpolate(mb.to(device), size=16, mode='bilinear')
imgs = torch.cat((x,xa,xb), dim=0)
labels = torch.cat((l,la,lb), dim=0)
masks = torch.cat((m,ma,mb), dim=0)
loss_dict = {}
# encode with random flip
noisy_content_feat, content_feat, style_feat = netAE.encode(imgs, 0.5)
label_feat = netAE.embed(labels)
recon_imgs, recon_masks = netAE.decode(style_feat, label_feat, noisy_content_feat)
# exchange content features of xa and xb for translation
style_feat2 = []
for s in style_feat:
style_feat2.append([s[0][args.batch:args.batch*3],s[1][args.batch:args.batch*3]])
_, trans_masks = netAE.decode(style_feat2, label_feat[args.batch:args.batch*3],
torch.cat((noisy_content_feat[args.batch*2:args.batch*3],
noisy_content_feat[args.batch:args.batch*2]), dim=0))
pred = netAE.classify(content_feat)
Larec = (F.mse_loss(recon_imgs, imgs) + vgg_loss(recon_imgs, imgs)) * args.lambda_app_rec
Lsrec = F.l1_loss(recon_masks, masks) * args.lambda_shape_rec
Lfdist = F.l1_loss(content_feat[args.batch:args.batch*2], content_feat[args.batch*2:args.batch*3].detach()) * args.lambda_feat_dist
Lsdist = F.l1_loss(trans_masks, masks[args.batch:args.batch*3]) * args.lambda_shape_dist
Lreg = torch.norm(content_feat, p=2) * args.lambda_reg + F.cross_entropy(pred, labels) * args.lambda_class
loss_dict['arec'] = Larec
loss_dict['srec'] = Lsrec
loss_dict['fdist'] = Lfdist
loss_dict['sdist'] = Lsdist
loss_dict['reg'] = Lreg
ae_loss = Larec + Lsrec + Lfdist + Lsdist + Lreg
optimizer_AE.zero_grad()
ae_loss.backward()
optimizer_AE.step()
message = ''
for k, v in loss_dict.items():
v = v.mean().float()
message += 'L%s: %.3f ' % (k, v)
pbar.set_description((message))
if ((idx+1) >= args.save_begin and (idx+1) % args.save_every == 0) or (idx+1) == args.iter:
torch.save(
{
"ae_ema": netAE.state_dict(),
"ae_optim": optimizer_AE.state_dict(),
#"args": args,
},
f"%s/%s-%05d.pt"%(args.model_path, args.task, idx+1),
)
if (idx+1) == args.iter:
torch.save(netAE.contentE.state_dict(),f"%s/content_encoder-%05d.pt"%(args.model_path, idx+1))
if idx == 0 or (idx+1) % args.visualize_every == 0 or (idx+1) == args.iter:
viznum = min(args.batch, 4)
masks = F.interpolate(masks, size=256)
recon_masks = F.interpolate(recon_masks, size=256)
trans_masks = F.interpolate(trans_masks, size=256)
sample = F.adaptive_avg_pool2d(torch.cat((imgs[0:viznum], masks[0:viznum],
recon_imgs[0:viznum], recon_masks[0:viznum],
imgs[args.batch:args.batch+viznum], masks[args.batch:args.batch+viznum],
recon_imgs[args.batch:args.batch+viznum], recon_masks[args.batch:args.batch+viznum],
imgs[args.batch*2:args.batch*2+viznum], masks[args.batch*2:args.batch*2+viznum],
trans_masks[0:viznum], trans_masks[args.batch:args.batch+viznum]), dim=0), 128).cpu()
utils.save_image(
sample,
f"log/%s/%05d.jpg"%(args.task, (idx+1)),
nrow=viznum*2,
normalize=True,
range=(-1, 1),
)
#plt.figure(figsize=(10,10), dpi=120)
#visualize(torchvision.utils.make_grid(sample, viznum*2, 2))
#plt.show()
if __name__ == "__main__":
parser = TrainOptions()
args = parser.parse()
print('*'*98)
if not os.path.exists("log/%s/"%(args.task)):
os.makedirs("log/%s/"%(args.task))
device = 'cuda'
netAE = AutoEncoder().to(device)
netAE.init_weights('kaiming', 0.02)
optimizer_AE = torch.optim.Adam(netAE.parameters(), lr=args.lr, betas=(0.9, 0.999))
print('Create models successfully!')
ufiles = os.listdir(args.unpaired_data_root)
natural_sort(ufiles)
udataset_sizes = [600] * len(ufiles)
udataset_sizes[-1] = 29000 # for faces
ulabels = list(range(len(ufiles)))
pfiles = os.listdir(args.paired_data_root)
natural_sort(pfiles)
pdataset_sizes = [600] * len(pfiles)
plabels = list(range(len(pfiles)))
# for unpaired data
udataloader = create_imagemasklabel_dataloader(args.unpaired_data_root, args.unpaired_mask_root,
ufiles, udataset_sizes, ulabels)
# for paired data
pdataloader = create_imagemasklabel_dataloader(args.paired_data_root, args.paired_mask_root,
pfiles, pdataset_sizes, plabels, pair=True)
print('Create dataloaders successfully!')
vgg_loss = VGGLoss()
vgg_loss.vgg = vgg_loss.vgg.to(device)
train(args, udataloader, pdataloader, netAE, optimizer_AE, vgg_loss, device)