/
train.py
268 lines (233 loc) · 9.52 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
import os
import torch
import numpy as np
import math
from torch.autograd import Variable
from datetime import datetime
from faster_rcnn import network
from faster_rcnn.network import init_data, data_to_variable
from faster_rcnn.network import train_net_params, print_weight_grad
from faster_rcnn.faster_rcnn_vgg import FasterRCNN as FasterRCNN_VGG
from faster_rcnn.faster_rcnn_res import FasterRCNN as FasterRCNN_RES
from faster_rcnn.utils.timer import Timer
from val import test, id_match_test
from faster_rcnn.roi_data_layer.sampler import sampler
from faster_rcnn.roi_data_layer.roidb import extract_roidb
from faster_rcnn.roi_data_layer.roibatchLoader import roibatchLoader
from faster_rcnn.fast_rcnn.config import cfg, cfg_from_file
try:
from termcolor import cprint
except ImportError:
cprint = None
try:
from pycrayon import CrayonClient
except ImportError:
CrayonClient = None
def log_print(text, color='blue', on_color=None, attrs=None):
if cprint is not None:
cprint(text, color=color, on_color=on_color, attrs=attrs)
else:
print(text)
# hyper-parameters
# ------------
imdb_name = 'voc_2007_trainval'
test_name = 'voc_2007_test'
# imdb_name = 'coco_2017_train'
# test_name = 'coco_2017_val'
# imdb_name = 'CaltechPedestrians_train'
# test_name = 'CaltechPedestrians_test'
cfg_file = 'experiments/cfgs/faster_rcnn_end2end.yml'
model_dir = 'data/pretrained_model/'
output_dir = 'models/saved_model3'
pre_model_name = 'voc_2007_trainval_14_vgg16_0.7_b1.h5'
pretrained_model = model_dir + pre_model_name
start_epoch = 1
end_epoch = 10
lr_decay_step = 5
lr_decay = 0.1
rand_seed = 1024
_DEBUG = True
use_tensorboard = True
remove_all_log = True # remove all historical experiments in TensorBoard
exp_name = None # the previous experiment name in TensorBoard
# ------------
if rand_seed is not None:
np.random.seed(rand_seed)
# load config
cfg_from_file(cfg_file)
fg_thresh = cfg.TRAIN.RPN_POSITIVE_OVERLAP
is_resnet = cfg.RESNET.IS_TRUE
batch_size = cfg.TRAIN.IMS_PER_BATCH
lr = cfg.TRAIN.LEARNING_RATE
momentum = cfg.TRAIN.MOMENTUM
disp_interval = cfg.TRAIN.DISPLAY
log_interval = cfg.TRAIN.LOG_IMAGE_ITERS
save_interval = cfg.TRAIN.SNAPSHOT_ITERS
# load data
imdb, roidb, ratio_list, ratio_index = extract_roidb(imdb_name)
test_imdb, test_roidb, _, _ = extract_roidb(test_name)
train_size = len(roidb)
sampler_batch = sampler(train_size, batch_size, cfg.TRIPLET.IS_TRUE)
dataset = roibatchLoader(imdb, roidb, ratio_list, ratio_index, batch_size,
imdb.num_classes, training=True)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size,
sampler=sampler_batch, num_workers=0)
# load net
if is_resnet:
model_name = cfg.RESNET.MODEL
cfg.TRAIN.DOUBLE_BIAS = False
cfg.TRAIN.WEIGHT_DECAY = 0.0001
net = FasterRCNN_RES(classes=imdb.classes, debug=_DEBUG)
net.init_module()
else:
model_name = 'vgg16'
net = FasterRCNN_VGG(classes=imdb.classes, debug=_DEBUG)
net.init_module()
if cfg.TRIPLET.IS_TRUE:
model_name += '_' + cfg.TRIPLET.LOSS
# network.load_net(pretrained_model, net)
# person_key = 15 (pascal_voc) user_defined_coco_set = 1
#network.load_net_pedestrians(pretrained_model, net, person_key=15)
blob = init_data(is_cuda=True)
# set net to be prepared to train
net.cuda()
params = train_net_params(net, cfg, lr)
optimizer = torch.optim.SGD(params, momentum=momentum)
def make_dir(output_dir):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
make_dir(output_dir)
# tensorboad
use_tensorboard = use_tensorboard and CrayonClient is not None
if use_tensorboard:
print('TENSORBOARD IS ON')
cc = CrayonClient(hostname='127.0.0.1')
if remove_all_log:
cc.remove_all_experiments()
if exp_name is None:
name = '{}_{}'.format(imdb_name, model_name)
exp_name = datetime.now().strftime(name+'_%m-%d_%H-%M')
exp = cc.create_experiment(exp_name)
else:
exp = cc.open_experiment(exp_name)
iters_per_epoch = int(train_size / batch_size)
# training
train_loss = 0
previous_precision = 0.
descend = 0
step_cnt = 0
cnt = 0
re_cnt = False
t = Timer()
t.tic()
from math import isnan
for epoch in range(start_epoch, end_epoch+1):
pf, tot = 0., 0
tp, fp, tn, fg, bg, tp_box, fg_box = 0., 0., 0., 0., 0., 0., 0.
rpn_cls, rpn_box, rcnn_cls, rcnn_box, sim_loss = 0., 0., 0., 0., 0.
net.train()
if epoch > 1 and (epoch-1) % lr_decay_step == 0:
lr *= lr_decay
params = train_net_params(net, cfg, lr)
optimizer = torch.optim.SGD(params, momentum=momentum)
data_iter = iter(dataloader)
for step in range(iters_per_epoch):
# get one batch
data = next(data_iter)
(im_data, im_info, gt_boxes, num_boxes) = data_to_variable(blob, data)
# forward
net.zero_grad()
net(im_data, im_info, gt_boxes, num_boxes)
if _DEBUG:
tp += float(net.tp)
tn += float(net.tn)
fp += float(net.fp)
fg += net.fg_cnt
bg += net.bg_cnt
tp_box += float(net.rpn.tp)
fg_box += net.rpn.fg_box
rpn_box += net.rpn.cross_entropy.data.cpu().numpy()[0]
rpn_cls += net.rpn.loss_box.data.cpu().numpy()[0]
rcnn_box += net.loss_box.data.cpu().numpy()[0]
rcnn_cls += net.cross_entropy.data.cpu().numpy()[0]
sim_loss += net.triplet_loss.data.cpu().numpy()[0] if cfg.TRIPLET.IS_TRUE else 0.
loss = net.rpn.loss + net.loss
if isnan(loss):
print(gt_boxes)
print(net.rpn.loss, net.loss)
train_loss += loss.data[0]
step_cnt += 1
cnt += 1
# backward
optimizer.zero_grad() # clear grad
loss.backward()
network.clip_gradient(net, 10.)
# print_weight_grad(net)
optimizer.step()
if step % disp_interval == 0 and step > 0:
duration = t.toc(average=False)
fps = step_cnt / duration
log_text = 'step %d, loss: %.4f, fps: %.2f (%.2fs per batch) --[epoch %2d] --[iter %4d/%4d]' % (
step, train_loss / step_cnt, fps, 1./fps, epoch, step, iters_per_epoch)
log_print(log_text, color='green', attrs=['bold'])
if _DEBUG:
if fg == 0 or bg == 0:
pass
else:
tot += 1
pf += tp/fg*100
match_rate = net.match/net.set * 100. if cfg.TRIPLET.IS_TRUE else 0.
log_print('\tEP: %.2f%% PR: %.2f%% TP: %.2f%%, TF: %.2f%%, fg/bg=(%d/%d), TD: %.2f%%' %
(tp_box/fg_box*100, tp/(tp+fp)*100, tp/fg*100., tn/bg*100., fg/step_cnt, bg/step_cnt, match_rate))
log_print('\trpn_cls: %.4f, rpn_box: %.4f, rcnn_cls: %.4f, rcnn_box: %.4f, sim_loss: %.4f' % (
rpn_cls/step_cnt, rpn_box/step_cnt, rcnn_cls/step_cnt, rcnn_box/step_cnt, sim_loss/step_cnt )
)
re_cnt = True
if use_tensorboard and cnt % log_interval == 0 and cnt > 0:
exp.add_scalar_value('train_loss', train_loss / step_cnt, step=cnt)
exp.add_scalar_value('learning_rate', lr, step=cnt)
if _DEBUG:
match_rate = net.match / net.set * 100. if cfg.TRIPLET.IS_TRUE else 0.
triplet_loss = net.triplet_loss.data.cpu().numpy() if cfg.TRIPLET.IS_TRUE else 0.
exp.add_scalar_value('true_positive', tp/fg*100., step=cnt)
exp.add_scalar_value('true_negative', tn/bg*100., step=cnt)
exp.add_scalar_value('precision', tp / (tp+fp) * 100., step=cnt)
exp.add_scalar_value('true_distance', match_rate, step=cnt)
losses = {'rpn_cls': float(rpn_cls/step_cnt),
'rpn_box': float(rpn_box/step_cnt),
'rcnn_cls': float(rcnn_cls/step_cnt),
'rcnn_box': float(rcnn_box/step_cnt),
'sim_loss': float(sim_loss/step_cnt)}
exp.add_scalar_dict(losses, step=cnt)
if re_cnt:
train_loss = 0
tp, fp, tn, fg, bg, tp_box, fg_box = 0., 0., 0., 0, 0, 0., 0
rpn_cls, rpn_box, rcnn_cls, rcnn_box, sim_loss = 0., 0., 0., 0., 0.
net.reset_match_count()
step_cnt = 0
t.tic()
re_cnt = False
# if epoch % save_interval == 0 and cnt > 0:
save_dir = os.path.join(output_dir, model_name)
make_dir(save_dir)
save_name = os.path.join(save_dir, '{}_{}_{}_{}_b{}.h5'
.format(imdb_name, epoch, model_name, fg_thresh, batch_size))
network.save_net(save_name, net)
print('save model: {}'.format(save_name))
if pf/tot > 80:
print('Entering Test Phase ...')
f = open('PrecisionAndRecall.txt', 'a')
prec, rec = test(save_name, net, test_imdb, test_roidb)
match = id_match_test(save_name, net, test_imdb, test_roidb, cfg.TRIPLET.LOSS) if cfg.TRIPLET.IS_TRUE else 0.
f.write(save_name + ' ----[prec: {:.2f}%, rec: {:.2f}%] / {:.2f}%\n'.format(prec, rec, match))
f.close()
if previous_precision == 0.:
previous_precision = prec
else:
if previous_precision > prec:
print('Precision decreased {:.2f}% -> {:.2f}% ...' \
.format(previous_precision, prec))
import warnings
warnings.warn('test set Precision decreased. Keep Watching')
else:
previous_precision = prec