-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_ctw1500_v2.py
220 lines (185 loc) · 9.85 KB
/
train_ctw1500_v2.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
import numpy as np
import torch
import argparse
import torch.nn as nn
import torch.nn.functional as F
import shutil
from torch.autograd import Variable
from torch.utils import data
from dataset import CTW1500Trainset_BoundE2E
from models import resnet50
from models.loss import dice_loss
from models.loss import get_pull_push_loss
from myutils import Logger
from myutils import AverageMeter
from myutils import RunningScore
from myutils import ohem_single, ohem_batch
from myutils import adjust_learning_rate_StepLR
from myutils import adjust_learning_rate_Poly
from myutils import PolynomialLR
import os
import sys
import time
# import pyclipper
# import Polygon as plg
def cal_kernel_score(kernels, gt_kernels, gt_texts, training_masks, running_metric_kernel):
mask = (gt_texts * training_masks).data.cpu().numpy()
kernel = kernels[:, -1, :, :]
gt_kernel = gt_kernels[:, -1, :, :]
pred_kernel = torch.sigmoid(kernel).data.cpu().numpy()
pred_kernel[pred_kernel <= 0.5] = 0
pred_kernel[pred_kernel > 0.5] = 1
pred_kernel = (pred_kernel * mask).astype(np.int32)
gt_kernel = gt_kernel.data.cpu().numpy()
gt_kernel = (gt_kernel * mask).astype(np.int32)
running_metric_kernel.update(gt_kernel, pred_kernel)
score_kernel, _ = running_metric_kernel.get_scores()
return score_kernel
def train(model, trainloader, criterion, optimizer, epoch, scheduler):
log_file = '/home/data1/zhm/ctw_purebound_checkpoint_poly_0417_baseline_predmask_nobound_moreaug_600e_log.txt'
print('Epoch:', epoch)
with open(log_file, 'a') as f:
f.write('Epoch: ' + str(epoch) + '\n')
model.train()
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
losses_text = AverageMeter()
losses_kernel = AverageMeter()
losses_top = AverageMeter()
losses_bot = AverageMeter()
losses_pull = AverageMeter()
losses_push = AverageMeter()
# losses_top_left = AverageMeter()
# losses_top_right = AverageMeter()
# losses_bot_right = AverageMeter()
# losses_bot_left = AverageMeter()
running_metric_kernel = RunningScore(2)
current_time = time.time()
# for batch_idx, (imgs, gt_texts, gt_kernels, gt_top_lefts, gt_top_rights, gt_bot_rights, gt_bot_lefts, training_masks) in enumerate(trainloader):
for batch_idx, (imgs, gt_texts, gt_kernels, gt_tops, gt_bots, gt_texts_labeled, gt_kernels_labeled, gt_tops_labeled, gt_bots_labeled, training_masks) in enumerate(trainloader):
data_time.update(time.time() - current_time)
imgs = Variable(imgs.cuda())
gt_texts = Variable(gt_texts.cuda())
gt_kernels = Variable(gt_kernels.cuda())
gt_tops = Variable(gt_tops.cuda())
gt_bots = Variable(gt_bots.cuda())
gt_texts_labeled = Variable(gt_texts_labeled.cuda())
gt_kernels_labeled = Variable(gt_kernels_labeled.cuda())
gt_tops_labeled = Variable(gt_tops_labeled.cuda())
gt_bots_labeled = Variable(gt_bots_labeled.cuda())
training_masks = Variable(training_masks.cuda())
# i_channels = Variable(i_channels.cuda())
# j_channels = Variable(j_channels.cuda())
outputs = model(imgs)
output_texts = outputs[:, 0, :, :]
output_kernels = outputs[:, 1, :, :]
output_tops = outputs[:, 2, :, :]
output_bots = outputs[:, 3, :, :]
output_sim_vectors = outputs[:, 4:, :, :]
# attention: -----------------generating training masks for each part---------------------
selected_text_masks = ohem_batch(output_texts, gt_texts, training_masks)
selected_text_masks = Variable(selected_text_masks.cuda())
# TODO: think twice whether to use ohem or the method used in the original PSENet paper
mask_training = training_masks.data.cpu().numpy()
mask_gt_text = gt_texts.data.cpu().numpy()
mask_pred_text = output_texts.data.cpu().numpy()
selected_kernel_masks = ((mask_training > 0.5) & (mask_pred_text > 0.5)).astype('float32')
selected_kernel_masks = torch.from_numpy(selected_kernel_masks).float()
selected_kernel_masks = Variable(selected_kernel_masks.cuda())
# selected_kernel_masks = ohem_batch(output_kernels, gt_kernels, training_masks)
# selected_top_masks = ohem_batch(output_tops, gt_tops, training_masks)
# mask_training = training_masks.data.cpu().numpy()
# mask_gt_top = gt_tops.data.cpu().numpy()
# mask_gt_text = gt_texts.data.cpu().numpy()
# mask_pred_top = torch.sigmoid(output_tops).data.cpu().numpy()
# selected_top_masks = ((mask_training > 0.5) & ((mask_gt_top > 0.5) | (mask_gt_text > 0.5) | (mask_pred_top > 0.5))).astype('float32')
# selected_top_masks = torch.from_numpy(selected_top_masks).float()
# selected_top_masks = Variable(selected_top_masks.cuda())
# selected_bot_masks = ohem_batch(output_bots, gt_bots, training_masks)
# mask_training = training_masks.data.cpu().numpy()
# mask_gt_bot = gt_bots.data.cpu().numpy()
# mask_gt_text = gt_texts.data.cpu().numpy()
# mask_pred_bot = torch.sigmoid(output_bots).data.cpu().numpy()
# selected_bot_masks = ((mask_training > 0.5) & ((mask_gt_bot > 0.5) | (mask_gt_text > 0.5) | (mask_pred_bot > 0.5))).astype('float32')
# selected_bot_masks = torch.from_numpy(selected_bot_masks).float()
# selected_bot_masks = Variable(selected_bot_masks.cuda())
# TODO: to complete the whole project, an embedding vector and its corresponding loss is needed(should be done before 04.01)
loss_pull, loss_push = get_pull_push_loss(outputs, gt_texts_labeled, gt_kernels_labeled, gt_tops_labeled, gt_bots_labeled)
loss_text = criterion(output_texts, gt_texts, selected_text_masks)
loss_kernel = criterion(output_kernels, gt_kernels, selected_kernel_masks)
# loss_top = criterion(output_tops, gt_tops, selected_top_masks)
# loss_bot = criterion(output_bots, gt_bots, selected_bot_masks)
# loss = 0.2 * loss_kernel + 1.0 * loss_top + 1.0 * loss_bot
# loss = 1.0 * loss_text + 0.5 * loss_kernel + 1.0 * (loss_top + loss_bot) + 0.25 * (loss_pull + loss_push)
loss = 1.0 * loss_text + 0.5 * loss_kernel + 0.25 * (loss_pull + loss_push)
losses.update(loss.item(), imgs.shape[0])
losses_text.update(loss_text.item(), imgs.shape[0])
losses_kernel.update(loss_kernel.item(), imgs.shape[0])
# losses_top.update(loss_top.item(), imgs.shape[0])
# losses_bot.update(loss_bot.item(), imgs.shape[0])
losses_pull.update(loss_pull.item(), imgs.shape[0])
losses_push.update(loss_push.item(), imgs.shape[0])
optimizer.zero_grad()
loss.backward()
optimizer.step()
current_lr = optimizer.param_groups[0]['lr']
if isinstance(scheduler, PolynomialLR):
# print('updating')
scheduler.step()
# score_kernel = cal_kernel_score(output_kernels, gt_kernels, gt_texts, training_masks, running_metric_kernel)
batch_time.update(time.time() - current_time)
if (batch_idx + 1) % 20 == 0:
output_log = '({batch}/{size}) Batch: {dt:.3f}s {bt:.3f}s {lr} | TOTAL: {total:.0f}s | ETA: {eta:.0f}s | Loss: {loss:.4f} |' \
' {loss_text:.4f} | {loss_kernel:.4f} | {loss_top:.4f} | {loss_bot:.4f} | {loss_pull:.4f} | {loss_push:.4f}'.format(
batch=batch_idx + 1,
size=len(trainloader),
dt=data_time.avg,
bt=batch_time.avg,
lr=current_lr,
total=batch_time.avg * batch_idx,
eta=batch_time.avg * (len(trainloader) - batch_idx),
loss=losses.avg,
loss_text=losses_text.avg,
loss_kernel=losses_kernel.avg,
loss_top=0,
loss_bot=0,
loss_pull=losses_pull.avg,
loss_push=losses_push.avg)
print(output_log)
sys.stdout.flush()
with open(log_file, 'a') as f:
f.write(output_log + '\n')
current_time = time.time()
def main(args):
num_classes = 8 # gt_text, gt_kernel, gt_top, gt_bot, sim_vector(n_channels:4)
trainset = CTW1500Trainset_BoundE2E(with_coord=False)
trainloader = torch.utils.data.DataLoader(dataset=trainset,
batch_size=16,
shuffle=True,
num_workers=1,
drop_last=True,
pin_memory=True)
if args.backbone == 'res50':
model = resnet50(pretrained=True, num_classes=num_classes)
model = torch.nn.DataParallel(model).cuda()
optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=0.99, weight_decay=5e-4)
n_epoch = 600
scheduler = PolynomialLR(optimizer=optimizer, max_iter=n_epoch * len(trainloader), power=0.9)
for epoch in range(n_epoch):
# adjust_learning_rate_StepLR(args, optimizer, epoch)
# adjust_learning_rate_Poly(args, 1e-3, optimizer, epoch, n_epoch, 0.9)
# TODO: train func
_ = train(model, trainloader, dice_loss, optimizer, epoch, scheduler)
checkpoint_info = {'epoch': epoch + 1,
'state_dict': model.state_dict(),
'lr': args.lr,
'optimizer': optimizer.state_dict()}
torch.save(checkpoint_info, '/home/data1/zhm/ctw_purebound_checkpoint_poly_0417_baseline_predmask_nobound_moreaug_600e.pth.tar')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--backbone', nargs='?', type=str, default='res50')
# parser.add_argument('--schedule', nargs='+', type=int, default=[200, 400])
parser.add_argument('--lr', nargs='?', type=float, default=1e-3)
args = parser.parse_args()
main(args)