-
Notifications
You must be signed in to change notification settings - Fork 0
/
Triplet_trainer.py
222 lines (182 loc) · 9.17 KB
/
Triplet_trainer.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
#!/usr/bin/env python
# -*- coding: UTF-8 -*-
'''=================================================
@Project -> File ：L5 -> Triplet_trainer
@IDE ：PyCharm
@Author ：DIPTE
@Date ：2020/8/3 0:28
@Desc ：
=================================================='''
import time
import datetime
import torch
import torch.nn as nn
import numpy as np
from loss_function import *
from average_meter import AverageMeter
class Trainer(object):
def __init__(self, epochs, dataloaders, model, optimizer, scheduler, device,
margin, ckpt_tag,criterion):
self.epochs = epochs
self.dataloaders = dataloaders
self.model = model
self.optimizer = optimizer
self.scheduler = scheduler
self.device = device
self.margin = margin
self.ckpt_tag = ckpt_tag
# save best model
self.best_val_acc = -100
self.criterion = criterion#torch.nn.CrossEntropyLoss().to(device)
# self.criterion = ContrastiveLoss(margin=0.8)
# self.criterion = FocalCrossEntropyLoss(3, 0.5)
# self.criterion = OHEMCrossEntropyLoss(0.75)
# self.criterion = torch.nn.CrossEntropyLoss().to(device)
# self.criterion = torch.nn.NLLLoss().to(device)
def train(self):
for epoch in range(self.epochs):
self.train_epoch(epoch, 'train')
self.eval_epoch(epoch, 'LFW')
print("Best acc on LFW: {}, best threshold: {}".format(self.best_val_acc,
self.best_threshold))
def train_epoch(self, epoch, phase):
self.model.train()
self.margin.train()
triplet_loss_sum = 0
num_valid_training_triplets = 0
margins=0.5
l2_distance = PairwiseDistance(2).cuda()
for batch_idx, sample in enumerate(self.dataloaders[phase]):
anc_img = sample['anc_img'].cuda()
pos_img = sample['pos_img'].cuda()
neg_img = sample['neg_img'].cuda()
# Forward pass - compute embeddings
anc_embedding, pos_embedding, neg_embedding = self.model(anc_img), self.model(pos_img), self.model(neg_img)
# Forward pass - choose hard negatives only for training
pos_dist = l2_distance.forward(anc_embedding, pos_embedding)
neg_dist = l2_distance.forward(anc_embedding, neg_embedding)
all = (neg_dist - pos_dist < margins).cpu().numpy().flatten()
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
anc_hard_embedding = anc_embedding[hard_triplets].cuda()
pos_hard_embedding = pos_embedding[hard_triplets].cuda()
neg_hard_embedding = neg_embedding[hard_triplets].cuda()
# Calculate triplet loss
triplet_loss = TripletLoss(margin=margins).forward(
anchor=anc_hard_embedding,
positive=pos_hard_embedding,
negative=neg_hard_embedding
).cuda()
# Calculating loss
triplet_loss_sum += triplet_loss.item()
num_valid_training_triplets += len(anc_hard_embedding)
# Backward pass
self.optimizer.zero_grad()
# optimizer_model.zero_grad()
triplet_loss.backward()
# optimizer_model.step()
self.optimizer.step()
# Model only trains on hard negative triplets
avg_triplet_loss = 0 if (
num_valid_training_triplets == 0) else triplet_loss_sum / num_valid_training_triplets
# Print training statistics and add to log
print('Epoch {}:\tAverage Triplet Loss: {:.4f}\tNumber of valid training triplets in epoch: {}'.format(
epoch + 1,
avg_triplet_loss,
num_valid_training_triplets
)
)
# if batch_idx % 40 == 0:
# print('Train Epoch: {} [{:08d}/{:08d} ({:02.0f}%)]\tLoss:{:.6f}\tAcc:{:.6f} LR:{:.7f}'.format(
# epoch, batch_idx * len(list(sample)), len(self.dataloaders[phase].dataset),
# 100. * batch_idx / len(self.dataloaders[phase]), triplet_loss.item(), avg_triplet_loss,
# self.optimizer.param_groups[0]['lr']))
self.scheduler.step()
torch.save(self.model.state_dict(),
'./checkpoints/{}_{}_Triplet_{:04d}.pth'.format(self.ckpt_tag,
str(self.margin), epoch))
torch.save(self.margin.state_dict(),
'./checkpoints/{}_512_{}_Triplet_{:04d}.pth'.format(self.ckpt_tag,
str(self.margin), epoch))
def eval_epoch(self, epoch, phase):
feature_ls = feature_rs = flags = folds = None
# sample = {'pair':[img_l, img_r], 'label': 1/-1}
for batch_idx, sample in enumerate(self.dataloaders[phase]):
img_l = sample['pair'][0].to(self.device)
img_r = sample['pair'][1].to(self.device)
flag = sample['label'].numpy()
fold = sample['fold'].numpy()
feature_l, feature_r = self.getDeepFeature(img_l, img_r)
feature_l, feature_r = feature_l.cpu().numpy(), feature_r.cpu().numpy()
if (feature_ls is None) and (feature_rs is None):
feature_ls = feature_l
feature_rs = feature_r
flags = flag
folds = fold
else:
feature_ls = np.concatenate((feature_ls, feature_l), 0)
feature_rs = np.concatenate((feature_rs, feature_r), 0)
flags = np.concatenate((flags, flag), 0)
folds = np.concatenate((folds, fold), 0)
accs, thresholds = self.evaluation_10_fold(feature_ls, feature_rs, flags, folds,
method='cos_distance')
print("Eval Epoch Average Acc: {:.4f}, Average Threshold: {:.4f}".format(
np.mean(accs), np.mean(thresholds)))
if np.mean(accs) > self.best_val_acc:
self.best_val_acc = np.mean(accs)
torch.save(self.model.state_dict(),
'./checkpoints/{}_{}_Triplet_best.pth'.format(self.ckpt_tag,
str(self.margin)))
torch.save(self.margin.state_dict(),
'./checkpoints/{}_512_{}_Triplet_best.pth'.format(self.ckpt_tag, str(self.margin)))
self.best_threshold = np.mean(thresholds)
def getDeepFeature(self, img_l, img_r):
self.model.eval()
with torch.no_grad():
feature_l = self.model(img_l)
feature_r = self.model(img_r)
return feature_l, feature_r
def evaluation_10_fold(self, feature_ls, feature_rs, flags, folds,
method='l2_distance'):
accs = np.zeros(10)
thresholds = np.zeros(10)
for i in range(10):
val_fold = (folds != i)
test_fold = (folds == i)
# minus by mean
mu = np.mean(np.concatenate((feature_ls[val_fold, :],
feature_rs[val_fold, :]),
0), 0)
feature_ls = feature_ls - mu
feature_rs = feature_rs - mu
# normalization
feature_ls = feature_ls / np.expand_dims(np.sqrt(np.sum(np.power(feature_ls, 2), 1)), 1)
feature_rs = feature_rs / np.expand_dims(np.sqrt(np.sum(np.power(feature_rs, 2), 1)), 1)
if method == 'l2_distance':
scores = np.sum(np.power((feature_ls - feature_rs), 2), 1)
elif method == 'cos_distance':
scores = np.sum(np.multiply(feature_ls, feature_rs), 1)
else:
raise NameError("Distance Method not supported")
thresholds[i] = self.getThreshold(scores[val_fold], flags[val_fold], 10000, method)
accs[i] = self.getAccuracy(scores[test_fold], flags[test_fold], thresholds[i], method)
return accs, thresholds
def getThreshold(self, scores, flags, thrNum, method='l2_distance'):
accs = np.zeros((2 * thrNum + 1, 1))
thresholds = np.arange(-thrNum, thrNum + 1) * 3 / thrNum
# print(thresholds)
# print(np.min(scores))
# print(np.max(scores))
for i in range(2 * thrNum + 1):
accs[i] = self.getAccuracy(scores, flags, thresholds[i], method)
max_index = np.squeeze(accs == np.max(accs))
best_threshold = np.mean(thresholds[max_index]) # multi best threshold
return best_threshold
def getAccuracy(self, scores, flags, threshold, method='l2_distance'):
if method == 'l2_distance':
pred_flags = np.where(scores < threshold, 1, -1)
elif method == 'cos_distance':
pred_flags = np.where(scores > threshold, 1, -1)
acc = np.sum(pred_flags == flags) / pred_flags.shape[0]
return acc