-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrainer.py
192 lines (149 loc) · 7.27 KB
/
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
import torch
import numpy as np
from torch.nn import TripletMarginLoss
from tqdm import tqdm
from torch.autograd import Variable
from torch.autograd import Function
from utils import PairwiseDistance
def fit(train_loader, val_loader, model, loss_fn, optimizer, scheduler, n_epochs, cuda, log_interval, metrics=[],
start_epoch=0):
"""
Loaders, model, loss function and metrics should work together for a given task,
i.e. The model should be able to process data output of loaders,
loss function should process target output of loaders and outputs from the model
Examples: Classification: batch loader, classification model, NLL loss, accuracy metric
Siamese network: Siamese loader, siamese model, contrastive loss
Online triplet learning: batch loader, embedding model, online triplet loss
"""
for epoch in range(0, start_epoch):
scheduler.step()
for epoch in range(start_epoch, n_epochs):
scheduler.step()
# Train stage
train_loss, metrics = train_epoch(train_loader, model, loss_fn, optimizer, cuda, log_interval, metrics)
message = 'Epoch: {}/{}. Train set: Average loss: {:.4f}'.format(epoch + 1, n_epochs, train_loss)
for metric in metrics:
message += '\t{}: {}'.format(metric.name(), metric.value())
val_loss, metrics = test_epoch(val_loader, model, loss_fn, cuda, metrics)
val_loss /= len(val_loader)
message += '\nEpoch: {}/{}. Validation set: Average loss: {:.4f}'.format(epoch + 1, n_epochs,
val_loss)
for metric in metrics:
message += '\t{}: {}'.format(metric.name(), metric.value())
print(message)
def train_epoch(train_loader, model, loss_fn, optimizer, cuda, log_interval, metrics):
for metric in metrics:
metric.reset()
model.train()
losses = []
total_loss = 0
for batch_idx, (data, target) in enumerate(train_loader):
target = target if len(target) > 0 else None
if not type(data) in (tuple, list):
data = (data,)
if cuda:
data = tuple(d.cuda() for d in data)
if target is not None:
target = target.cuda()
optimizer.zero_grad()
outputs = model(*data)
if type(outputs) not in (tuple, list):
outputs = (outputs,)
loss_inputs = outputs
if target is not None:
target = (target,)
loss_inputs += target
loss_outputs = loss_fn(*loss_inputs)
loss = loss_outputs[0] if type(loss_outputs) in (tuple, list) else loss_outputs
losses.append(loss.item())
total_loss += loss.item()
loss.backward()
optimizer.step()
for metric in metrics:
metric(outputs, target, loss_outputs)
if batch_idx % log_interval == 0:
message = 'Train: [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
batch_idx * len(data[0]), len(train_loader.dataset),
100. * batch_idx / len(train_loader), np.mean(losses))
for metric in metrics:
message += '\t{}: {}'.format(metric.name(), metric.value())
print(message)
losses = []
total_loss /= (batch_idx + 1)
return total_loss, metrics
def test_epoch(val_loader, model, loss_fn, cuda, metrics):
with torch.no_grad():
for metric in metrics:
metric.reset()
model.eval()
val_loss = 0
for batch_idx, (data, target) in enumerate(val_loader):
target = target if len(target) > 0 else None
if not type(data) in (tuple, list):
data = (data,)
if cuda:
data = tuple(d.cuda() for d in data)
if target is not None:
target = target.cuda()
outputs = model(*data)
if type(outputs) not in (tuple, list):
outputs = (outputs,)
loss_inputs = outputs
if target is not None:
target = (target,)
loss_inputs += target
loss_outputs = loss_fn(*loss_inputs)
loss = loss_outputs[0] if type(loss_outputs) in (tuple, list) else loss_outputs
val_loss += loss.item()
for metric in metrics:
metric(outputs, target, loss_outputs)
return val_loss, metrics
def train_epoch_some(train_loader, model, loss_fn, optimizer, cuda, log_interval, metrics):
for metric in metrics:
metric.reset()
model.train()
losses = []
total_loss = 0
pbar = tqdm(enumerate(train_loader))
labels, distances = [], []
l2_dist = PairwiseDistance(2)
for batch_idx, (data_a, data_p, data_n, label_p, label_n) in pbar:
data_a, data_p, data_n = data_a.cuda(), data_p.cuda(), data_n.cuda()
data_a, data_p, data_n = Variable(data_a), Variable(data_p), \
Variable(data_n)
# compute output
out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)
# Choose the hard negatives
d_p = l2_dist.forward(out_a, out_p)
d_n = l2_dist.forward(out_a, out_n)
all = (d_n - d_p < args.margin).cpu().data.numpy().flatten()
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
out_selected_a = Variable(torch.from_numpy(out_a.cpu().data.numpy()[hard_triplets]).cuda())
out_selected_p = Variable(torch.from_numpy(out_p.cpu().data.numpy()[hard_triplets]).cuda())
out_selected_n = Variable(torch.from_numpy(out_n.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_a = Variable(torch.from_numpy(data_a.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_p = Variable(torch.from_numpy(data_p.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_n = Variable(torch.from_numpy(data_n.cpu().data.numpy()[hard_triplets]).cuda())
selected_label_p = torch.from_numpy(label_p.cpu().numpy()[hard_triplets])
selected_label_n = torch.from_numpy(label_n.cpu().numpy()[hard_triplets])
triplet_loss = loss_fn.forward(out_selected_a, out_selected_p, out_selected_n)
cls_a = model.forward_classifier(selected_data_a)
cls_p = model.forward_classifier(selected_data_p)
cls_n = model.forward_classifier(selected_data_n)
cls_a = model.forward_classifier(selected_data_a)
cls_p = model.forward_classifier(selected_data_p)
cls_n = model.forward_classifier(selected_data_n)
criterion = nn.CrossEntropyLoss()
predicted_labels = torch.cat([cls_a, cls_p, cls_n])
true_labels = torch.cat(
[Variable(selected_label_p.cuda()), Variable(selected_label_p.cuda()), Variable(selected_label_n.cuda())])
cross_entropy_loss = criterion(predicted_labels.cuda(), true_labels.cuda())
loss = cross_entropy_loss + triplet_loss
# compute gradient and update weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
# update the optimizer learning rate
adjust_learning_rate(optimizer)