-
Notifications
You must be signed in to change notification settings - Fork 24
/
train.py
350 lines (294 loc) · 20.1 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
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
from __future__ import print_function, absolute_import
import argparse
import os.path as osp
import os
import numpy as np
import copy
import sys
import torch
from torch.backends import cudnn
from reid import datasets
from reid import models
from reid.models.cal_cls_params import CalClsParams
from reid.utils.data import create_test_data_loader, create_train_data_loader
from reid.utils.data.cluster import Cluster
from reid.utils.logging import Logger
from reid.optim.pcb_trainer import PCBTrainer
from reid.eval.rerankor import Rerankor
from reid.utils.osutils import load_mat
def main(args):
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
cudnn.benchmark = True
# Redirect print to both console and log file
sys.stdout = Logger(osp.join(args.log_dir, 'log.txt'))
args.num_classes = 1
# Create train data loaders
ori_dataset = datasets.create(args.name, args.data_dir)
cur_dataset = copy.deepcopy(ori_dataset) # initialize the current target dataset as the whole target dataset
dist = load_mat(args.rerank_dist_file, args.name, args.rerank_dist_file, 'dist')
if args.dbscan_use:
cluster = Cluster(args, ori_dataset.train)
if args.start_dbscan == -1:
cur_dataset.train, cur_dataset.train_indices = cluster.hdbscancluster(dist, iteration=-1)
train_loader_CTL, train_loader_RTL = create_train_data_loader(args, args.name, cur_dataset, dist=dist,
istrain=True, savepath=args.save_dir)
# Create test data loaders
test_dataset = {}
test_dataset['dataset'] = datasets.create(args.name, args.data_dir)
test_dataset['train_loader'], test_dataset['query_loader'], test_dataset['gallery_loader'] \
= create_test_data_loader(args, args.name, test_dataset['dataset'])
if not args.num_train_loader:
args.num_train_loader = len(train_loader_CTL)
###### initial trainer #######
if args.idloss_use:
t_f = load_mat(args.init_t_t_f, args.name, args.init_t_t_f, 'features')
calclsparams = CalClsParams(args.name, cur_dataset, t_f)
cls_params = calclsparams.cal_cls_params()
trainer = PCBTrainer(args, cls_params=cls_params)
trainer.model = trainer.update_cls_params(trainer.model, cls_params)
print(trainer.model.module.cls_list)
else:
cls_params = None
trainer = PCBTrainer(args, cls_params=cls_params)
############### train and test ################
best_source = [np.zeros(3), np.zeros(1), np.zeros(1), np.zeros(1)]
best_state = 'Conservative Stage'
## initial rerankor ##
rerankor = Rerankor()
######################### start iter ########################
for iteration in range(args.start_iters, args.iters):
################ Start Conservative Stage Training ##############################
print("--------------------------Iter: {} Starting Conservative Stage Training----------------------".format(iteration))
stage = 'Conservative Stage'
for epoch in range(args.start_epochs, args.epochs):
trainer.train_conservative(iteration, epoch, train_loader_CTL, train_loader_RTL, print_freq=args.print_freq, cls_list=None, stage=stage)
torch.cuda.empty_cache()
print('Iteration: {}/{} Epoch: {}/{} has down'.format(iteration, args.iters, epoch, args.epochs))
evaluator = trainer.test()
if (epoch+1) % args.save_freq == 0 or epoch == args.epochs - 1:
# save model of every save_freq epochs
trainer.may_save_ckpt(name=args.name)
######################## test model of every 10 * save_freq epochs###########################
scores = {}
scores['cmc_scores'], scores['mAP'], _, _, _ = \
evaluator.evaluate(args.name, test_dataset['query_loader'], test_dataset['gallery_loader'],
test_dataset['dataset'].query, test_dataset['dataset'].gallery, istrain=False,
isevaluate=True, issave=True, savepath=args.save_dir)
# save best model of target
if scores['mAP'] >= best_source[1] and scores['cmc_scores'][0] >= best_source[0][0]:
trainer.may_save_ckpt(name='Best_' + args.name)
best_source[0] = copy.deepcopy(scores['cmc_scores'])
best_source[1] = copy.deepcopy(scores['mAP'])
best_source[2] = copy.deepcopy(epoch)
best_source[3] = copy.deepcopy(iteration)
best_state = 'Conservative Stage'
print('Cross Domain CMC Scores')
print('Source\t Target\t Top1\t Top5\t Top10\t MAP\t Epoch\t Iteration\t Best_state')
print('{}->{}: {:6.2%} {:6.2%} {:6.2%} ({:.2%}) {} {} {}'.format(args.s_name, args.name,
scores['cmc_scores'][0], scores['cmc_scores'][1], scores['cmc_scores'][2],
scores['mAP'], epoch, iteration, best_state))
if (epoch+1) % args.dbscan_iter == 0:
_, _, t_f, _, _ = \
evaluator.evaluate(args.name, test_dataset['train_loader'], test_dataset['train_loader'],
test_dataset['dataset'].train, test_dataset['dataset'].train,
savepath=args.save_dir, issave=True, istrain=True, isevaluate=False)
if epoch != args.epochs - 1:
dist = rerankor.rerank(t_f, t_f,
savepath=os.path.join(args.save_dir, 'rerank'),
save=True, isevaluate=args.rerank_eval,
dataset=test_dataset['dataset'])
################## train_loader, dataset is after cluster ###############################
if args.dbscan_use:
cur_dataset.train, cur_dataset.train_indices = cluster.hdbscancluster(dist, iteration=iteration)
train_loader_CTL, train_loader_RTL = create_train_data_loader(args, args.name, cur_dataset,
dist=dist, istrain=True, savepath=args.save_dir)
print('Iteration: {} Epoch: {} train loader has changed.'.format(iteration, epoch))
del dist
else:
del train_loader_CTL
del train_loader_RTL
try:
del dist
except:
pass
print("--------------------------Iter: {} Ending Conservative Stage Training----------------------".format(iteration))
################# start train supervision model ################
if args.idloss_use and (iteration % args.idloss_iter == 0):
calclsparams = CalClsParams(args.name, cur_dataset, t_f)
cls_params = calclsparams.cal_cls_params()
trainer.model = trainer.update_cls_params(trainer.model, cls_params)
print(trainer.model.module.cls_list)
else:
cls_params = None
## change to supervision lr_scheduler ##
train_loader_IDL, _ = create_train_data_loader(args, args.name, cur_dataset,
istrain=True, idloss_only=True,
savepath=os.path.join(args.save_dir))
trainer.create_lr_scheduler_promoting(len(train_loader_IDL))
print("--------------------------Iter: {} Starting Promoting Stage Training----------------------".format(iteration))
stage = 'Promoting Stage'
for epoch in range(args.epochs):
trainer.train_promoting(iteration, epoch, train_loader_IDL, print_freq=args.print_freq, cls_list=trainer.model.module.cls_list, stage=stage)
torch.cuda.empty_cache()
evaluator = trainer.test()
if epoch % args.save_freq == 0 or epoch == args.epochs - 1:
# save model of every save_freq epochs
trainer.may_save_ckpt(name=args.name)
######################## test ###########################
scores = {}
scores['cmc_scores'], scores['mAP'], _, _, _ = \
evaluator.evaluate(args.name, test_dataset['query_loader'], test_dataset['gallery_loader'],
test_dataset['dataset'].query, test_dataset['dataset'].gallery, istrain=False,
isevaluate=True, issave=True, savepath=args.save_dir)
# save best model of target
if scores['mAP'] >= best_source[1] and scores['cmc_scores'][0] >= best_source[0][0]:
trainer.may_save_ckpt(name='Best_' + args.name)
best_source[0] = copy.deepcopy(scores['cmc_scores'])
best_source[1] = copy.deepcopy(scores['mAP'])
best_source[2] = copy.deepcopy(epoch)
best_source[3] = copy.deepcopy(iteration)
best_state = 'Promoting Stage'
print('Cross Domain CMC Scores')
print('Source\t Target\t Top1\t Top5\t Top10\t MAP\t Epoch\t Iteration\t Best_state')
print('{}->{}: {:6.2%} {:6.2%} {:6.2%} ({:.2%}) {} {} {}'.format(args.s_name, args.name,
scores['cmc_scores'][0],
scores['cmc_scores'][1],
scores['cmc_scores'][2],
scores['mAP'], epoch, iteration,
best_state))
evaluator = trainer.test()
_, _, t_f, _, _ = \
evaluator.evaluate(args.name, test_dataset['train_loader'], test_dataset['train_loader'],
test_dataset['dataset'].train, test_dataset['dataset'].train,
savepath=args.save_dir, issave=True, istrain=True, isevaluate=False)
dist = rerankor.rerank(t_f, t_f,
savepath=os.path.join(args.save_dir, 'rerank'),
save=True, isevaluate=args.rerank_eval,
dataset=test_dataset['dataset'])
################## get new train_loader and dataset before next conservative ###############################
if args.dbscan_use:
cur_dataset.train, cur_dataset.train_indices = cluster.hdbscancluster(dist, iteration=iteration+1)
train_loader_CTL, train_loader_RTL = create_train_data_loader(args, args.name, cur_dataset,
dist=dist, istrain=True, savepath=args.save_dir)
del train_loader_IDL
del dist
print("--------------------------Iter: {} Ending Promoting Stage Training----------------------".format(iteration))
print('[Best test] Cross Domain CMC Scores')
print('Source\t Target\t Top1\t Top5\t Top10\t MAP\t Epoch\t Iteration\t Best_state')
print('{}->{}: {:6.2%} {:6.2%} {:6.2%} ({:.2%}) {} {} {}'.format(args.s_name, args.name, best_source[0][0],
best_source[0][1], best_source[0][2],
best_source[1], best_source[2],
best_source[3], best_state))
if __name__ == '__main__':
parser = argparse.ArgumentParser(description="PAST")
# data
parser.add_argument('--data_dir', type=str, metavar='PATH',
default='./data/')
parser.add_argument('--name', type=str, default='market1501',
choices=datasets.names())
parser.add_argument('--log_dir', type=str, metavar='PATH',
default='./log')
parser.add_argument('--save_dir', type=str, metavar='PATH', default='./checkpoint',
help='Directory to store experiment output, including model checkpoint and tensorboard files, etc.')
parser.add_argument('--s_name', type=str, default='market1501', help='pretrained source dataset name')
# train parameters
parser.add_argument('--batch_size', type=int, default=16)
parser.add_argument('--workers', type=int, default=0)
parser.add_argument('--split', type=int, default=0)
parser.add_argument('--height', type=int, default=256,
help="input height, default: 256 for resnet*, "
"144 for inception")
parser.add_argument('--width', type=int, default=128,
help="input width, default: 128 for resnet*, "
"56 for inception")
parser.add_argument('--combine_trainval', action='store_true',
help="train and val sets together for training, "
"val set alone for validation")
parser.add_argument('--phase', type=str, default='normal',
choices=['scratch', 'normal', 'finetune', 'fix_finetune_layers'])
parser.add_argument('--print_freq', type=int, default=100)
parser.add_argument('--seed', type=int, default=1)
parser.add_argument('--continue_training', action='store_true',
help='whether continue training or not. if True, continue from the intermediate status')
parser.add_argument('--load_optimizer', action='store_true',
help='when continue training, whether load optimizer and scheduler or not.')
parser.add_argument('--save_freq', type=int, default=20, help='how many epochs to save checkpoint')
# test parameters
parser.add_argument('--dist_type', type=str, default='cosine', choices=['euclidean', 'cosine'])
parser.add_argument('--resume_file', type=str, default='', metavar='PATH')
parser.add_argument('--evaluate', action='store_true',
help="evaluation only")
# model
parser.add_argument('--model_name', type=str, default='resnet50', choices=models.names())
parser.add_argument('--embedding_dim', type=int, default=256)
parser.add_argument('--num_parts', type=int, default=6)
parser.add_argument('--dropout', type=float, default=0.5)
parser.add_argument('--backbone_pretrained', type=bool, default=True)
parser.add_argument('--backbone_pretrained_model_dir', type=str, default='initialization/pretrained_model/', metavar='PATH')
parser.add_argument('--last_conv_stride', type=int, default=1)
parser.add_argument('--local_conv', action='store_true', help='last_conv using linear or conv')
parser.add_argument('--max_or_avg', type=str, default='max')
parser.add_argument('--pool_type', type=str, default='PCBPool_nine', choices=['PCBPool', 'PCBPool_nine'])
parser.add_argument('--forward_type', type=str, default='reid')
# optimizer
parser.add_argument('--optimizer', type=str, default='sgd', choices=['sgd', 'adam'])
parser.add_argument('--weight_decay', type=float, default=5e-4)
parser.add_argument('--momentum', type=float, default=0.9, help='the parameter of SGD')
parser.add_argument('--nesterov', type=bool, default=False, help='the parameter of SGD')
parser.add_argument('--ft_lr', type=float, default=0.0001,
help="learning rate of finetune parameters")
parser.add_argument('--new_params_lr', type=float, default=0.0002, help="learning rate of new parameters")
parser.add_argument('--lr_decay_epochs', nargs='*', type=int, default=25, help="epoch for lr descend")
parser.add_argument('--num_train_loader', type=bool, default=False, help="epoch for lr descend * len(train_loader)")
parser.add_argument('--start_epochs', type=int, default=0, help="start_epochs")
parser.add_argument('--epochs', type=int, default=60, help="max_epochs")
parser.add_argument('--ft_lr_promoting', type=float, default=0.00005,
help="learning rate of finetune parameters for supervision")
parser.add_argument('--new_params_lr_promoting', type=float, default=0.001,
help="learning rate of new parameters for supervision")
parser.add_argument('--lr_decay_iters', nargs='*', type=int, default=8, help="iters for lr descend")
parser.add_argument('--start_iters', type=int, default=0, help="start_iters")
parser.add_argument('--iters', type=int, default=10, help="max_iters")
parser.add_argument('--dbscan_type', type=str, default='hdbscan', choices=['dbscan', 'hdbscan'])
parser.add_argument('--dbscan_use', type=bool, default=True, help='whether use dbscan or not')
parser.add_argument('--dbscan_iter', type=int, default=20, help="dbscan_iter epochs to calculate once dbscan")
parser.add_argument('--dbscan_minsample', type=int, default=10, help="dbscan_minsample")
parser.add_argument('--start_dbscan', type=int, default=-1, help="start_epochs")
# loss
parser.add_argument('--idloss_use', action='store_true', help='whether use ID loss or not')
parser.add_argument('--idloss_weight', type=float, default=0.1, help='ID loss weight')
parser.add_argument('--idloss_name', type=str, default='idL', help='ID loss name')
parser.add_argument('--idloss_iter', type=int, default=1, help="idloss_iter epochs to calculate once idloss")
parser.add_argument('--triloss_use', action='store_true', help='whether use Triplet loss or not')
parser.add_argument('--triloss_weight', type=float, default=0.1, help='Triplet loss weight')
parser.add_argument('--triloss_rho_ctl', type=float, default=1.0, help='Triplet loss weight')
parser.add_argument('--triloss_rho_rtl', type=float, default=1.0, help='Triplet loss weight')
parser.add_argument('--triloss_name', type=str, default='triL', help='Triplet loss name')
parser.add_argument('--triloss_mean', action='store_true', help='part triplet loss sum or mean')
parser.add_argument('--triloss_part', type=int, default=9, help='Triplet loss part sum or mean')
parser.add_argument('--tri_sampler_type', type=str, default='CTL_RTL', help='Triplet sampler type', choices=['CTL','RTL','CTL_RTL'])
parser.add_argument('--k_nearest', type=int, default=20, help='k_nearest for tri_sampler_type==softmargintriplet')
parser.add_argument('--hard_type', type=str, default='tri_hard',
help='batch triplets selection when tri_sampler_type==RandomIdentitySampler')
parser.add_argument('--margin', type=float, default=0.3, help='Triplet loss margin')
parser.add_argument('--num_instances', type=int, default=4,
help="number of instances per identity (if use triplet loss)")
parser.add_argument('--norm_by_num_of_effective_triplets', type=bool, default=False,
help='Triplet loss norm_by_num_of_effective_triplets')
# rerank misc parameters
parser.add_argument('--rerank', action='store_true', help="rerank or not")
parser.add_argument('--dist_epoch', nargs='*', type=int, default=0,
help="epoch for rerank distmat, the sequence is the same with test_names")
parser.add_argument('--k1', type=int, default=20, help="rerank k1")
parser.add_argument('--k2', type=int, default=6, help="rerank k2")
parser.add_argument('--lambda_value', type=float, default=0.3, help='rerank lambda_value')
parser.add_argument('--rerank_eval', action='store_true', help="after rerank, whether evaluate or not")
parser.add_argument('--rerank_dist_file', type=str, default='', metavar='PATH',
help='initial rerank distmat file path')
parser.add_argument('--rho', type=float, default=1.6e-3,
help="rho percentage, default: 1.6e-3")
parser.add_argument('--init_t_t_f', type=str, default='', metavar='PATH',
help='initial target train features file path')
main(parser.parse_args())