-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
272 lines (220 loc) · 10.8 KB
/
train.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
import os
import datetime
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import argparse
import random
from tqdm import tqdm
from torchvision import transforms
import torch.utils.data as data
from model import *
from utils import *
from config import *
import wandb
from copy import deepcopy
def set_seed(global_seed):
torch.manual_seed(7)
torch.cuda.manual_seed_all(999)
np.random.seed(global_seed)
random.seed(global_seed)
def train(net, args, train_loader, val_loaders, test_loaders, device, mode='synth'):
param_groups = net.parameters()
max_f1 = 0.0
start_epoch = 1
optimizer = torch.optim.Adam(param_groups, lr=args.lr, weight_decay=args.weight_decay)
if args.rot_data:
sym_type = 'rotation'
else:
sym_type = 'reflection'
for epoch in range(start_epoch, args.num_epochs + 1):
adjust_learning_rate(optimizer, epoch)
net.train()
print('%s image train' % mode)
print(epoch)
running_sample = 0
running_loss = 0
running_theta_loss = 0
train_log_images = []
for idx, data in enumerate(tqdm(train_loader)):
if args.get_theta:
img, mask, axis, axis_gs, is_syn, a_lbl = data
a_lbl = a_lbl.to(device)
else:
img, mask, axis, axis_gs, is_syn, _ = data
a_lbl = None
_mask = (mask > 0).float().to(device)
axis_out, mask_out, total_loss, losses = net(img=img.to(device), lbl=axis_gs.float().to(device), mask=_mask, is_syn = is_syn.to(device), a_lbl=a_lbl)
loss = total_loss.mean()
if a_lbl is not None:
loss += losses[1].mean() * args.theta_loss_scale
running_theta_loss += losses[1].mean().item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
running_sample += 1
running_loss += loss.item()
if (idx == len(train_loader) - 1) and not args.wandb_off:
log_dict = {"Train Loss": running_loss / running_sample,}
if a_lbl is not None:
log_dict['Train theta Loss'] = running_theta_loss / running_sample
running_sample = 0
running_loss = 0
running_theta_loss = 0
train_log_images = [
wandb.Image(unnorm(img), caption='Image/train_%d_image' % idx),
wandb.Image(axis_gs.float(), caption='Image/train_%d_axisGT' % idx),
wandb.Image(axis_out.cpu(), caption='Image/train_%d_axisPred' % idx),
]
log_dict['train_log_images'] = train_log_images
wandb.log(log_dict)
if epoch % 5 == 0:
rec, prec, f1, f1_max = test(net, args, val_loaders, device, mode='val', sym_type=sym_type) # subset of testset
_max_f1 = f1_max[0]
max_f1 = max(max_f1, _max_f1)
checkpoint = {
'state_dict': net.state_dict(),
'optimizer': optimizer.state_dict(),
'args': args,
'epoch': epoch,
'max_f1': max_f1,
}
if not os.path.exists('./weights'):
os.makedirs('./weights')
torch.save(checkpoint, './weights/v_' + args.ver + '_last_checkpoint.pt')
print(max_f1)
if max_f1 == _max_f1:
print('best model renewed.')
best_ckpt = deepcopy(net.state_dict())
torch.save(checkpoint, './weights/v_' + args.ver + '_best_checkpoint.pt')
net.load_state_dict(best_ckpt)
rec, prec, f1, f1_max = test(net, args, test_loaders, device, mode='test', sym_type=sym_type)
checkpoint['stats'] = rec, prec, f1, f1_max
print(f1_max)
torch.save(checkpoint, './weights/v_' + args.ver + '_best_checkpoint.pt')
def test(net, args, test_loaders, device, mode='test', sym_type='reflection'):
net.eval()
val_sample, val_loss = 0, 0
n_thresh = 100 if mode in ['test', 'ref_test', 'rot_test'] else 10
recs, precs, f1s, f1_maxs= [], [], [], []
with torch.no_grad():
for i, test_loader in enumerate(test_loaders):
axis_eval = PointEvaluation(n_thresh, blur_pred=True, device=device)
val_log_images = []
for idx, data in enumerate(tqdm(test_loader)):
if args.get_theta:
img, mask, axis, axis_gs, is_syn, a_lbl = data
a_lbl.to(device)
else:
img, mask, axis, axis_gs, is_syn, a_lbl = data
# a_lbl = None
a_lbl = a_lbl.to(device)
_mask = (mask > 0).float().to(device)
axis_out, mask_out, total_loss, losses = net(img=img.to(device), lbl=axis_gs.float().to(device), mask=_mask, is_syn = is_syn, a_lbl=a_lbl, sym_type=sym_type)
axis_out = F.interpolate(axis_out, size=axis.size()[2:], mode='bilinear', align_corners=True)
img = F.interpolate(img, size=axis.size()[2:], mode='bilinear', align_corners=True)
axis_eval(axis_out, axis)
val_loss += total_loss.mean().item()
val_sample += 1
if not args.wandb_off:
_val_log_images = [
wandb.Image(unnorm(img), caption='Image/%s%d_image%d' % (mode, i, idx)),
wandb.Image(axis_gs, caption='Image/%s%d_axisGT%d' % (mode, i, idx)),
wandb.Image(axis_out.cpu(), caption='Image/%s%d_axisPred%d' % (mode, i, idx)),
]
val_log_images += _val_log_images
rec, prec, f1 = axis_eval.f1_score()
recs.append(rec), precs.append(prec), f1s.append(f1), f1_maxs.append(f1.max())
print(f1.max())
if not args.wandb_off:
log_dict = {
"%s_%d Loss" % (mode, i): val_loss / val_sample,
"%s_%d new max f1" % (mode, i): f1.max(),
"%s_%d new max_f1_index" % (mode, i): f1.argmax()
}
log_dict['%s_%d_log_images' % (mode, i)] = val_log_images
wandb.log(log_dict)
return recs, precs, f1s, f1_maxs
def vis(net, args, test_loaders, device, mode='test', sym_type='reflection'):
net.eval()
with torch.no_grad():
for i, test_loader in enumerate(test_loaders):
if (i == 0) and (mode in ['test']):
print('skip nyu')
continue
val_log_images = []
for idx, data in enumerate(tqdm(test_loader)):
if args.get_theta:
img, mask, axis, axis_gs, is_syn, a_lbl = data
a_lbl.to(device)
else:
img, mask, axis, axis_gs, is_syn, _ = data
a_lbl = None
axis_out, _, _, _ = net(img=img.to(device), lbl=axis_gs.float().to(device), \
mask=(mask > 0).float().to(device), is_syn = is_syn, a_lbl=a_lbl, sym_type=sym_type)
axis_out = F.interpolate(axis_out, size=axis.size()[2:], mode='bilinear', align_corners=True)
img = F.interpolate(img, size=axis.size()[2:], mode='bilinear', align_corners=True)
if args.save_qual:
img_vis = (unnorm(img)[0])
axis_max = F.adaptive_max_pool2d(axis_out, (1, 1))
axis_min = - F.adaptive_max_pool2d(-axis_out, (1, 1))
axis_out = (axis_out - axis_min) / (axis_max - axis_min)
axis = axis_out[0][0].cpu()
axis = axis / axis.max()
heatmap = cv2.applyColorMap(np.uint8(255 * axis), cv2.COLORMAP_JET)
heatmap = cv2.cvtColor(heatmap, cv2.COLOR_BGR2RGB)
heatmap = np.float32(heatmap) / 255
img_ = np.float32(img_vis.cpu().permute(1, 2, 0)) #/ 255
_attended_img = heatmap + np.float32(img_)
attended_img = _attended_img / np.max(_attended_img)
plt.imsave("./vis_result/%s/%03d_pred.png" % (args.model_name, idx), np.clip(attended_img, 0, 1))
def adjust_learning_rate(optimizer, epoch):
lr = args.lr * (0.1 ** (epoch // (args.num_epochs * 0.5))) * (0.1 ** (epoch // (args.num_epochs * 0.75)))
for param_group in optimizer.param_groups:
param_group['lr'] = lr
if __name__ == '__main__':
set_seed(1)
args = get_parser()
args.sync_bn = True
comment = str(args.ver)
print(comment)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Let's use", torch.cuda.device_count(), "GPUs!")
if not args.wandb_off:
print('wandb on')
if args.test_only:
wandb.init(project="symdet_val", entity="symdet", notes=comment, save_code=True)
else:
wandb.init(project="symdet", entity="symdet", notes=comment, save_code=True)
wandb.config.update(args)
else:
print('wandb off')
if args.rot_data:
sym_type = 'rotation'
else:
sym_type = 'reflection'
net = SymmetryDetectionNetwork(args)
net = nn.DataParallel(net)
net.to(device)
from dendi_loader import NewSymmetryDatasets
trainset = NewSymmetryDatasets(sym_type=sym_type, split='train', input_size=[args.input_size, args.input_size], get_theta=args.get_theta, with_ref_circle=True, t_resize=False)
valset = NewSymmetryDatasets(sym_type=sym_type, split='val', input_size=[args.input_size, args.input_size], get_theta=args.get_theta, with_ref_circle=True, t_resize=False)
testset = NewSymmetryDatasets(sym_type=sym_type, split='test', input_size=[args.input_size, args.input_size], get_theta=1, with_ref_circle=True, t_resize=False)
train_loader = data.DataLoader(trainset, batch_size=args.bs_train, shuffle=True, num_workers=4, drop_last=True)
val_loader = data.DataLoader(valset, batch_size=1, shuffle=False, num_workers=4)
test_loader = data.DataLoader(testset, batch_size=1, shuffle=False, num_workers=4)
if args.test_only:
print('load pretrained model')
ckpt_path = './weights/v_' + args.ver + '_best_checkpoint.pt'
checkpoint = torch.load(ckpt_path)
net.load_state_dict(checkpoint['state_dict'], strict=True)
# net.module.export()
# net.to(device)
# print('export done')
# print(checkpoint['args'])
rec, prec, f1, f1_max = test(net, args, (test_loader,), device, mode='test', sym_type=sym_type)
checkpoint['stats'] = rec, prec, f1, f1_max
print(f1_max)
else:
train(net, args, train_loader, (val_loader, ), (test_loader, ), device)