-
Notifications
You must be signed in to change notification settings - Fork 0
/
model_utils.py
195 lines (164 loc) · 7.49 KB
/
model_utils.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
import time
from datetime import datetime
import os
import os.path as osp
import torch
import torch.optim as optim
import torch.nn as nn
from torch.utils.data import DataLoader
def train_step(model, train_loader, criterion, optimizer, epoch, device, scheduler, log_interval=10):
model.train()
train_loss = 0.
correct = 0
total = 0
train_loss_batch = 0
epoch_size = len(train_loader.dataset)
t_start = time.time()
for batch_idx, (inputs, targets) in enumerate(train_loader):
inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
scheduler.step(epoch)
train_loss += loss.item()
_, predicted = outputs.max(1)
total += targets.size(0)
if len(targets.size()) == 2:
# Labels could be a posterior probability distribution. Use argmax as a proxy.
target_probs, target_labels = targets.max(1)
else:
target_labels = targets
correct += predicted.eq(target_labels).sum().item()
prog = total / epoch_size
exact_epoch = epoch + prog - 1
acc = 100. * correct / total
train_loss_batch = train_loss / total
if (batch_idx + 1) % log_interval == 0:
print('[Train] Epoch: {:.2f} [{}/{} ({:.0f}%)]\tLoss: {:.6f}\tAccuracy: {:.1f} ({}/{})'.format(
exact_epoch, batch_idx * len(inputs), len(train_loader.dataset), 100. * batch_idx / len(train_loader),
loss.item(), acc, correct, total))
t_end = time.time()
t_epoch = int(t_end - t_start)
acc = 100. * correct / total
return train_loss_batch, acc
def test_step(model, test_loader, criterion, device, epoch=0., silent=False):
model.eval()
test_loss = 0.
correct = 0
total = 0
t_start = time.time()
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(test_loader):
#print(f"testing batch {batch_idx}")
inputs, targets = inputs.to(device), targets.to(device)
outputs = model(inputs)
loss = criterion(outputs, targets)
nclasses = outputs.size(1)
test_loss += loss.item()
_, predicted = outputs.max(1)
total += targets.size(0)
correct += predicted.eq(targets).sum().item()
t_end = time.time()
t_epoch = int(t_end - t_start)
acc = 100. * correct / total
test_loss /= total
if not silent:
print('[Test] Epoch: {}\tLoss: {:.6f}\tAcc: {:.1f}% ({}/{})'.format(epoch, test_loss, acc,
correct, total))
return test_loss, acc
def train_model(model, trainset, out_path, batch_size=64, criterion_train=None, criterion_test=None, testset=None,
device=None, num_workers=10, lr=0.1, momentum=0.5, lr_step=30, lr_gamma=0.1, resume=None,
epochs=100, log_interval=100, weighted_loss=False, checkpoint_suffix='', optimizer=None, scheduler=None,
callback=None,
**kwargs):
if device is None:
device = torch.device('cuda')
if not osp.exists(out_path):
print(f"Path {out_path} does not exist. Creating it...")
os.mkdir(out_path)
run_id = str(datetime.now())
# Data loaders
train_loader = DataLoader(trainset, batch_size=batch_size, shuffle=True, num_workers=num_workers)
if testset is not None:
test_loader = DataLoader(testset, batch_size=batch_size, shuffle=False, num_workers=num_workers)
else:
test_loader = None
if weighted_loss:
if not isinstance(trainset.samples[0][1], int):
print('Labels in trainset is of type: {}. Expected: {}.'.format(type(trainset.samples[0][1]), int))
class_to_count = dd(int)
for _, y in trainset.samples:
class_to_count[y] += 1
class_sample_count = [class_to_count[c] for c, cname in enumerate(trainset.classes)]
print('=> counts per class: ', class_sample_count)
weight = np.min(class_sample_count) / torch.Tensor(class_sample_count)
weight = weight.to(device)
print('=> using weights: ', weight)
else:
weight = None
# Optimizer
if criterion_train is None:
criterion_train = nn.CrossEntropyLoss(reduction='mean', weight=weight)
if criterion_test is None:
criterion_test = nn.CrossEntropyLoss(reduction='mean', weight=weight)
if optimizer is None:
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum, weight_decay=5e-4)
if scheduler is None:
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=lr_step, gamma=lr_gamma)
start_epoch = 1
best_train_acc, train_acc = -1., -1.
best_test_acc, test_acc, test_loss = -1., -1., -1.
# Resume if required
if resume is not None:
model_path = resume
if osp.isfile(model_path):
print("=> loading checkpoint '{}'".format(model_path))
checkpoint = torch.load(model_path)
start_epoch = checkpoint['epoch']
best_test_acc = checkpoint['best_acc']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(model_path))
# Initialize logging
log_path = osp.join(out_path, 'train{}.log.tsv'.format(checkpoint_suffix))
if not osp.exists(log_path): # Remove previous log
with open(log_path, 'w') as wf:
columns = ['run_id', 'epoch', 'split', 'loss', 'accuracy', 'best_accuracy']
wf.write('\t'.join(columns) + '\n')
model_out_path = osp.join(out_path, 'checkpoint{}.pth.tar'.format(checkpoint_suffix))
for epoch in range(start_epoch, epochs + 1):
#scheduler.step(epoch) # should call optimizer.step() before scheduler.stop(epoch)
train_loss, train_acc = train_step(model, train_loader, criterion_train, optimizer, epoch, device,
scheduler, log_interval=log_interval)
best_train_acc = max(best_train_acc, train_acc)
if test_loader is not None:
#print("Start testing")
test_loss, test_acc = test_step(model, test_loader, criterion_test, device, epoch=epoch)
best_test_acc = max(best_test_acc, test_acc)
# Checkpoint
if test_acc >= best_test_acc:
state = {
'epoch': epoch,
'arch': model.__class__,
'state_dict': model.state_dict(),
'best_acc': test_acc,
'optimizer': optimizer.state_dict(),
'created_on': str(datetime.now()),
}
torch.save(state, model_out_path)
# Log
with open(log_path, 'a') as af:
train_cols = [run_id, epoch, 'train', train_loss, train_acc, best_train_acc]
af.write('\t'.join([str(c) for c in train_cols]) + '\n')
test_cols = [run_id, epoch, 'test', test_loss, test_acc, best_test_acc]
af.write('\t'.join([str(c) for c in test_cols]) + '\n')
# Callback
if callback and test_acc >= callback:
with open(log_path, 'a') as af:
af.write(f'Validation accuracy reaches {callback}, so stop training.\n')
return model, train_loader
return model, train_loader