-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain_unsup.py
161 lines (140 loc) · 7.96 KB
/
train_unsup.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
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as T
import torch.backends.cudnn as cudnn
cudnn.benchmark=True
import sys
from tqdm import tqdm
import argparse
import sys
import os
import numpy as np
from collections import OrderedDict
from utils import *
from loss import TripletLoss
from logger import Logger
import models
from cmc import Self_Cmc, Vanilla_Cmc
import inference
def vanilla_validation(q_dataloader, q_txt, g_dataloader, g_txt, base_net, rank_size):
q_features = inference.inference(q_dataloader, base_net)
g_features = inference.inference(g_dataloader, base_net)
CMC, mAP = Vanilla_Cmc(q_features, q_txt, g_features, g_txt, rank_size=rank_size)
return CMC[0], mAP
def self_validation(dataloader, txt, base_net, rank_size):
features = inference.inference(dataloader, base_net)
CMC, mAP = Self_Cmc(features, txt, rank_size=rank_size)
return CMC[0], mAP
def train_loop(args, tgt_dataloader, base_net):
optimizer_base = optim.SGD(base_net.parameters(), lr=args.lr, momentum=0.9, weight_decay=1e-4)
scheduler = optim.lr_scheduler.StepLR(optimizer_base, args.lr_step_size, 0.1)
criterion_triplet = TripletLoss(margin=args.margin, batch_hard=args.batch_hard)
loss_logger = Logger(args.save_model_dir, 'loss.csv')
acc_logger = Logger(args.save_model_dir, 'acc.csv')
best_tgt_r1 = 0.0
for it in range(args.n_iters):
# Validation
if it % args.test_every_n_iter == 0:
tgt_train_r1, tgt_train_mAP = self_validation(tgt_dataloader['train'], args.tgt_train, base_net, args.rank_size)
print('(iter %d)[Train] Target r-1 accuracy: %.3f, mAP: %.3f' % (it, tgt_train_r1, tgt_train_mAP))
if args.tgt_val is not None:
tgt_r1, tgt_mAP = self_validation(tgt_dataloader['val'], args.tgt_val, base_net, args.rank_size)
else:
tgt_r1, tgt_mAP = vanilla_validation(tgt_dataloader['query'], args.tgt_query,
tgt_dataloader['gallery'], args.tgt_gallery, base_net, args.rank_size)
print('(iter %d)[Val] Target r-1 accuracy: %.3f, mAP: %.3f' % (it, tgt_r1, tgt_mAP))
log = OrderedDict([('tgt train r-1', tgt_train_r1),
('tgt train mAP', tgt_train_mAP),
('tgt val r-1', tgt_r1),
('tgt val mAP', tgt_mAP)])
acc_logger.logg(it, log)
loss_logger.write_log()
acc_logger.write_log()
if tgt_r1 > best_tgt_r1:
torch.save(base_net.state_dict(),os.path.join(args.save_model_dir,'model_best_base.ckpt'))
best_tgt_r1 = tgt_r1
log = {}
data = next(tgt_dataloader['train_triplet'])
b_img = data['img'].cuda()
pos_mask = data['pos_mask'].cuda()
neg_mask = data['neg_mask'].cuda()
#forward
pred_feat = base_net(b_img)
b_loss = criterion_triplet(pred_feat, pos_mask=pos_mask, neg_mask=neg_mask, mode='mask')
loss = b_loss.mean()
# backward
base_net.zero_grad()
loss.backward()
optimizer_base.step()
log['loss_tgt_triplet'] = b_loss.data.mean().item()
log['loss_tgt_triplet_max'] = b_loss.data.max().item()
log['loss_tgt_triplet_min'] = b_loss.data.min().item()
scheduler.step()
loss_logger.logg(it, log)
print('(iter %d)[Train] Target triplet loss = %.3e' %
(it, log['loss_tgt_triplet']))
def parse_args():
parser = argparse.ArgumentParser(description='Train Re-ID net', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--tgt_pkl', help='pkl for target dataset (train)')
parser.add_argument('--tgt_train', help='txt for target dataset (train)')
parser.add_argument('--tgt_val', help='txt for target dataset (val)')
parser.add_argument('--tgt_query', help='txt for target dataset (val)')
parser.add_argument('--tgt_gallery', help='txt for target dataset (val)')
parser.add_argument('--lr',type=float,default=0.001,help='learning rate')
parser.add_argument('--lr_step_size',type=int,default=5000,help='step size for stepLR')
parser.add_argument('--batch_size',type=int,default=128,help='batch size number')
parser.add_argument('--n_iters',type=int,default=20,help='number of training iterations')
parser.add_argument('--load_ckpt',default=None,help='path to load ckpt')
parser.add_argument('--save_model_dir',default=None,help='path to save model')
parser.add_argument('--n_layer',type=int,default=18,help='number of Resnet layers')
parser.add_argument('--margin',type=str,default='0',help='margin of triplet loss ("soft" or float)')
parser.add_argument('--class_per_batch',type=int,default=32,help='# of class per batch for triplet training')
parser.add_argument('--image_per_class',type=int,default=4,help='# of images per class for triplet training')
parser.add_argument('--batch_hard',action='store_true',help='whether to use batch_hard for triplet loss')
parser.add_argument('--display_every_n_iter',type=int,default=1,help='display every n iterations')
parser.add_argument('--test_every_n_iter',type=int,default=200,help='test every n iterations')
parser.add_argument('--save_every_n_iter',type=int,default=1000,help='save model every n iterations')
parser.add_argument('--pretrain_model',type=str,default=None,help='load pretrained model')
parser.add_argument('--sample_mode',type=str,choices=['fix','unfix'],default='unfix',help='choose sampling model for CTM dataloader')
parser.add_argument('--drop_ctm', type=bool, default=False, help='do not train CTM')
parser.add_argument('--rank_size', type=int, default=100, help='rank size for CMC and mAP')
parser.add_argument('--fc_dim', type=int, default=0, help='adding a fc of # dim to resnet50')
args = parser.parse_args()
args.margin = args.margin if args.margin=='soft' else float(args.margin)
if args.class_per_batch * args.image_per_class != args.batch_size:
sys.exit('batch_size need to equal class_in_batch*image_per_class_in_batch')
if args.save_model_dir != None:
os.system('mkdir -p %s' % os.path.join(args.save_model_dir))
return args
if __name__ == '__main__':
# Parse arg
args = parse_args()
# Set transformation
train_transform = T.Compose([T.RandomHorizontalFlip(), T.Resize((224, 224)), T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
test_transform = T.Compose([T.Resize((224, 224)), T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
# Get target dataLoader
tgt_dataloader = {}
tgt_dataloader['train_triplet'] = Get_unsupervised_triplet_DataLoader(args.tgt_pkl, train_transform, args.batch_size,
args.sample_mode, args.image_per_class)
tgt_dataloader['train'] = Get_normal_DataLoader(args.tgt_train, test_transform, batch_size=64)
if args.tgt_val is not None:
tgt_dataloader['val'] = Get_normal_DataLoader(args.tgt_val, test_transform, batch_size=64)
else:
tgt_dataloader['query'] = Get_normal_DataLoader(args.tgt_query, test_transform, batch_size=64)
tgt_dataloader['gallery'] = Get_normal_DataLoader(args.tgt_gallery, test_transform, batch_size=64)
# Get Model
#base_net = models.new_FeatureResNet(n_layers=args.n_layer, pretrained=True, fc_dim=args.fc_dim)
base_net = models.FeatureResNet(n_layers=args.n_layer, pretrained=True)
if args.pretrain_model is not None:
pretrain_model = torch.load(args.pretrain_model)
for key in ['fc.weight','fc.bias']:
if key in pretrain_model:
del pretrain_model[key]
base_net.load_state_dict(pretrain_model)
print('pretrain model loaded.')
base_net = base_net.cuda()
# Train
train_loop(args, tgt_dataloader, base_net)