-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
293 lines (238 loc) · 10 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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import time
import copy
from tqdm import tqdm
import numpy as np
import copy
import utils
from grad_cam import GradCam, show_cam_on_image
from dataloader import UnNormalize
from tta import *
tta_list = [NoneAug(),
Hflip(),
# Vflip(),
Rotate(30),
Rotate(60),
# Rotate(90),
]
def mixup_data(x, y, alpha=1.0, device=None):
'''Returns mixed inputs, pairs of targets, and lambda'''
if alpha > 0:
lam = np.random.beta(alpha, alpha)
else:
lam = 1
batch_size = x.size()[0]
index = torch.randperm(batch_size).to(device)
mixed_x = lam * x + (1 - lam) * x[index, :]
y_a, y_b = y, y[index]
return mixed_x, y_a, y_b, lam
def mixup_criterion(criterion, pred, y_a, y_b, lam):
return lam * criterion(pred, y_a) + (1 - lam) * criterion(pred, y_b)
def train_model(dataloaders, model, criterion, optimizer, summary_writer,
scheduler=None, scheduler_name='multistep', num_epochs=20, device=None,
is_inception=False, mixup=False, alpha=1.0, val_dis=[]):
tic = time.time()
acc_history = []
best_acc = 0
best_model_wgt = None
step_per_epoch = len(dataloaders['train'].dataset) / dataloaders['train'].batch_size
for epoch in range(num_epochs):
print('epoch {}/{}'.format(epoch + 1, num_epochs))
for phase in ['train', 'val']:
if phase == 'train':
model.train()
running_loss = 0.0
else:
model.eval()
running_correct = 0.0
err = []
for i, (inputs, labels) in enumerate(dataloaders[phase]):
print(inputs.shape)
inputs = inputs.to(device)
labels = labels.to(device)
if phase == 'train' and mixup:
inputs, targets_a, targets_b, lam = mixup_data(inputs, labels, alpha, device)
if phase == 'train' and is_inception:
logits, aux_logits = model(inputs)
if mixup:
loss1 = mixup_criterion(criterion, logits, targets_a, targets_b, lam)
loss2 = mixup_criterion(criterion, aux_logits, targets_a, targets_b, lam)
loss = (loss1 + 0.4*loss2)
else:
loss1 = criterion(logits, labels)
loss2 = criterion(aux_logits, labels)
loss = (loss1 + 0.4*loss2)
else:
if phase == 'train' and mixup:
logits = model(inputs)
loss = mixup_criterion(criterion, logits, targets_a, targets_b, lam)
else:
with torch.no_grad():
logits = model(inputs)
loss = criterion(logits, labels)
if phase == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
summary_writer.add_scalar(tag='loss', scalar_value=loss.item(),
global_step=step_per_epoch*epoch+i)
running_loss += loss.item()
if scheduler is not None and (scheduler_name == 'cycle' or
scheduler_name == 'warmup' or
scheduler_name == 'cos' or
scheduler_name == 'cosw'):
scheduler.step()
_, preds = logits.max(1)
if phase == 'train' and mixup:
correct = (lam * preds.eq(targets_a.data).sum().float()
+ (1 - lam) * preds.eq(targets_b.data).sum().float())
else:
correct = (preds == labels).sum()
running_correct += correct
if phase == 'val':
# print('label:', labels)
# print('preds:', preds)
error_label = (preds != labels).long() * (labels + 1)
error_label = error_label[error_label>0]
err.append(error_label)
del inputs, labels, loss
torch.cuda.empty_cache()
epoch_acc = running_correct.double() / len(dataloaders[phase].dataset)
if phase == 'train':
epoch_loss = running_loss / len(dataloaders[phase].dataset)
print('train --> loss: {:.4f}, acc: {:.4f}'.format(epoch_loss, epoch_acc))
else:
print('val --> acc:{:.4f}'.format(epoch_acc))
if len(val_dis) >= 8:
errors = torch.cat(err, 0)
# total_err = errors.cpu().numpy().shape[0]
err_rate = np.bincount(errors.cpu().numpy(), minlength=9)[1:] / val_dis
[print('{}: {:.2f}'.format(utils.weather_classes[i], rate)) for i, rate in enumerate(err_rate)]
acc_history.append(epoch_acc.item())
summary_writer.add_scalar(tag='correct', scalar_value=epoch_acc, global_step=epoch)
summary_writer.add_scalar(tag='learning rate', scalar_value=optimizer.param_groups[0]["lr"],
global_step=epoch)
if epoch_acc >= best_acc:
best_acc = epoch_acc
if epoch > num_epochs//2:
best_model_wgt = copy.deepcopy(model.state_dict())
print()
if scheduler is not None:
if scheduler_name == 'multistep':
scheduler.step()
elif scheduler_name == 'plateau':
scheduler.step(epoch_loss)
summary_writer.close()
toc = time.time()
time_elapsed = toc - tic
print('training time -> %d:%.2f' % (time_elapsed // 60, time_elapsed % 60))
print('best_acc:', best_acc.item())
if best_model_wgt is not None:
model.load_state_dict(best_model_wgt)
return model, best_acc.item()
def clean_data(loader, model, device):
model.eval()
labels = []
im_names = []
err = []
not_correct = 0
for images, labels, names in tqdm(loader):
images = images.to(device)
labels = labels.to(device)
logits = torch.softmax(model(images), 1)
score, preds = logits.max(1)
not_correct += (preds != labels).sum()
# score_ = ((score>=0.9) | (0.4<score) * (score<0.6)).long()
error_label = (preds != labels).long() * (labels + 1) # *score_
for i, lab in enumerate(error_label):
# if lab>0 and (preds[i] in [5, 7] or labels[i] in [5, 7]):
if lab>0:
im_names.append((names[i].split('/')[-1], preds[i].item()+1, labels[i].item()+1, score[i].item()))
print("%d/%d" % (len(im_names), not_correct))
return im_names
def eval_model(loader, model, device):
model.eval()
labels = []
im_names = []
for images, names in tqdm(loader):
images = images.to(device)
logits = model(images)
_, preds = logits.max(1)
for p, n in zip(preds.cpu().numpy().tolist(), names):
im_names.append(n)
labels.append(p+1)
return im_names, labels
def eval_logits(loader, model, device=None):
model.eval()
labels = []
im_names = []
for images, names in tqdm(loader):
images = images.to(device)
logits = model(images)
for p, n in zip(logits.detach().cpu().numpy().tolist(), names):
im_names.append(n)
labels.append(p)
return im_names, labels
def eval_model_tta(loader, model, tta_augs=tta_list, device=None):
model.eval()
labels = []
im_names = []
for images, names in tqdm(loader):
images = TensorToPILs(images)
logits = []
for aug in tta_augs:
aug_imgs = PILsToTensor(aug(images)).to(device)
outputs = model(aug_imgs).detach().cpu().numpy().tolist()
logits.append(outputs)
# print(np.shape(logits))
logits = np.mean(np.array(logits), axis=0)
preds = np.argmax(logits, axis=1)
for p, n in zip(preds, names):
im_names.append(n)
labels.append(p+1)
return im_names, labels
def eval_logits_tta(loader, model, tta_augs=tta_list, device=None):
model.eval()
labels = []
im_names = []
for images, names in tqdm(loader):
images = TensorToPILs(images)
logits = []
for aug in tta_augs:
aug_imgs = PILsToTensor(aug(images)).to(device)
outputs = model(aug_imgs).detach().cpu().numpy().tolist()
logits.append(outputs)
# print(np.shape(logits))
logits = np.mean(np.array(logits), axis=0)
for p, n in zip(logits, names):
im_names.append(n)
labels.append(p)
return im_names, labels
def grad_cam(loader, model, device):
model.eval()
labels = []
im_names = []
use_cuda = device != None
model_no_fc = copy.deepcopy(model)
del model_no_fc.fc
cam = GradCam(model=model_no_fc, target_layer_names=['layer4'], use_cuda=use_cuda, org_model=model)
utils.mkdir('cam')
unorm = UnNormalize()
for i, (images, labels, org_imgs) in enumerate(tqdm(loader)):
images = images.to(device)
labels = labels.to(device)
logits = model(images)
_, preds = logits.max(1)
right = (preds == labels)
masks = cam(images, None)
# print(org_imgs.shape)
if right:
name = './cam/{}-{}-{}-{}.jpg'.format(i, utils.weather_classes[preds[0]], utils.weather_classes[labels[0]], right[0])
img = unorm(images[0]).cpu().numpy().transpose(1, 2, 0)
show_cam_on_image(img, masks, name)
# print(org_imgs[0].shape)
utils.save_image(org_imgs[0].numpy(), './cam/{}-{}.jpg'.format(i, utils.weather_classes[labels[0]]))
print('cam finished!')