-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
315 lines (238 loc) · 12.9 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
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
import time
import os
import copy
import argparse
import pdb
import collections
import sys
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torch.autograd import Variable
from torchvision import datasets, models, transforms
import torchvision
from imgaug import augmenters as iaa
from imgaug.augmentables.bbs import BoundingBox, BoundingBoxesOnImage
import matplotlib.pyplot as plt
import matplotlib.patches as patches
import model
from anchors import Anchors
import losses
from dataloader import CSVDataset, collater, Resizer, AspectRatioBasedSampler, Augmenter, UnNormalizer, Normalizer
from torch.utils.data import Dataset, DataLoader
from torch.utils.tensorboard import SummaryWriter
import csv_eval
#assert torch.__version__.split('.')[1] == '4'
print('CUDA available: {}'.format(torch.cuda.is_available()))
def main(args=None):
"""
In current implementation, if test csv is provided, we use that as validation set and combine the val and train csv's
as the csv for training.
If train_all_labeled_data flag is use, then we combine all 3 (if test is provided) for training and use a prespecified learning rate step schedule.
"""
parser = argparse.ArgumentParser(description='Simple training script for training a RetinaNet network.')
parser.add_argument('--csv_train', help='Path to file containing training annotations (see readme)')
parser.add_argument('--csv_classes', help='Path to file containing class list (see readme)')
parser.add_argument('--csv_val', help='Path to file containing validation annotations (optional, see readme)', default=None)
parser.add_argument('--csv_test',
help='Path to file containing test annotations (optional, if provided, train & val will be combined for training and test will be used for evaluation)',
default=None)
parser.add_argument('--lr', type=float, default=2e-5)
parser.add_argument('--depth', help='Resnet depth, must be one of 18, 34, 50, 101, 152', type=int, default=101)
parser.add_argument('--epochs', help='Number of epochs', type=int, default=25)
parser.add_argument('--model_output_dir', type=str, default='models')
parser.add_argument('--train_all_labeled_data',
help='Combine train, val, and test into 1 training set. Will use prespecified learning rate scheduler steps',
action='store_true')
parser.add_argument('--resnet-backbone-normalization', choices=['batch_norm', 'group_norm'], type=str, default='batch_norm')
parser = parser.parse_args(args)
print('Learning Rate: {}'.format(parser.lr))
print("Normalization: ", parser.resnet_backbone_normalization)
# Create folder - will raise error if folder exists
assert(os.path.exists(parser.model_output_dir) == False)
os.mkdir(parser.model_output_dir)
if parser.csv_train is None:
raise ValueError('Must provide --csv_train when training,')
if parser.csv_classes is None:
raise ValueError('Must provide --csv_classes when training,')
if not parser.csv_val and parser.csv_test:
raise ValueError("Cannot specify test set without specifying validation set")
if parser.train_all_labeled_data:
csv_paths = [parser.csv_train, parser.csv_val, parser.csv_test]
train_csv = []
for path in csv_paths:
if isinstance(path, str):
train_csv.append(path)
val_csv = None
else:
if parser.csv_train and parser.csv_val and parser.csv_test:
train_csv = [parser.csv_train, parser.csv_val] # Combine train and val sets for training
val_csv = parser.csv_test
else:
train_csv = parser.csv_train
val_csv = parser.csv_val
print('loading train data')
print(train_csv)
dataset_train = CSVDataset(train_file=train_csv, class_list=parser.csv_classes,
transform=transforms.Compose([Normalizer(), Augmenter(), Resizer()]))
print(dataset_train.__len__())
if val_csv is None:
dataset_val = None
print('No validation annotations provided.')
else:
dataset_val = CSVDataset(train_file=val_csv, class_list=parser.csv_classes,
transform=transforms.Compose([Normalizer(), Resizer()]))
print('putting data into loader')
sampler = AspectRatioBasedSampler(dataset_train, batch_size=2, drop_last=False)
dataloader_train = DataLoader(dataset_train, num_workers=3, collate_fn=collater, batch_sampler=sampler)
if dataset_val is not None:
sampler_val = AspectRatioBasedSampler(dataset_val, batch_size=1, drop_last=False)
dataloader_val = DataLoader(dataset_val, num_workers=3, collate_fn=collater, batch_sampler=sampler_val)
# Create the model
print('creating model')
if parser.depth == 18:
retinanet = model.resnet18(num_classes=dataset_train.num_classes(), pretrained=True, normalization=parser.resnet_backbone_normalization)
elif parser.depth == 34:
retinanet = model.resnet34(num_classes=dataset_train.num_classes(), pretrained=True, normalization=parser.resnet_backbone_normalization)
elif parser.depth == 50:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(), pretrained=True, normalization=parser.resnet_backbone_normalization)
elif parser.depth == 101:
retinanet = model.resnet101(num_classes=dataset_train.num_classes(), pretrained=True, normalization=parser.resnet_backbone_normalization)
elif parser.depth == 152:
retinanet = model.resnet152(num_classes=dataset_train.num_classes(), pretrained=True, normalization=parser.resnet_backbone_normalization)
else:
raise ValueError('Unsupported model depth, must be one of 18, 34, 50, 101, 152')
use_gpu = True
if use_gpu:
retinanet = retinanet.cuda()
retinanet = torch.nn.DataParallel(retinanet).cuda()
retinanet.training = True
optimizer = optim.Adam(retinanet.parameters(), lr=parser.lr)
lr_factor = 0.3
if not parser.train_all_labeled_data:
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=3, factor=lr_factor, verbose=True)
else:
# these milestones are for when using the lung masks - not for unmasked lung data
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[12, 16, 20, 24], gamma=lr_factor) # masked training
#scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[14, 18, 22, 26], gamma=lr_factor)
loss_hist = collections.deque(maxlen=500)
retinanet.train()
retinanet.module.freeze_bn()
#initialize tensorboard
writer = SummaryWriter(comment=parser.model_output_dir)
# Augmentation
seq = iaa.Sequential([
iaa.Fliplr(0.5),
iaa.Flipud(0.5),
iaa.Affine(
scale={"x": (1.0, 1.2), "y": (1.0, 1.2)},
rotate=(-20, 20),
shear=(-4, 4)
)
], random_order=True)
def augment(data, seq):
for n, img in enumerate(data['img']):
# imgaug needs dim in format (H, W, C)
image = data['img'][n].permute(1, 2, 0).numpy()
bbs_array = []
for ann in data['annot'][n]:
x1, y1, x2, y2, _ = ann
bbs_array.append(BoundingBox(x1=x1,y1=y1,x2=x2,y2=y2))
bbs = BoundingBoxesOnImage(bbs_array, shape=image.shape)
image_aug, bbs_aug = seq(image=image, bounding_boxes=bbs)
# save augmented image and chage dims to (C, H, W)
data['img'][n] = torch.tensor(image_aug.copy()).permute(2, 0, 1)
# save augmented annotations
for i, bbox in enumerate(bbs_aug.bounding_boxes):
x1, y1, x2, y2 = bbox.x1, bbox.y1, bbox.x2, bbox.y2
obj_class = data['annot'][n][i][-1]
data['annot'][n][i] = torch.tensor([x1, y1, x2, y2, obj_class])
return data
print('Num training images: {}'.format(len(dataset_train)))
dir_training_images = os.path.join(os.getcwd(), writer.log_dir, 'training_images')
os.mkdir(dir_training_images)
best_validation_loss = None
best_validation_map = None
for epoch_num in range(parser.epochs):
writer.add_scalar('Train/LR', optimizer.param_groups[0]['lr'], epoch_num)
retinanet.train()
retinanet.module.freeze_bn()
epoch_loss = []
for iter_num, data in enumerate(dataloader_train):
try:
optimizer.zero_grad()
data = augment(data, seq)
# save a few training images to see what augmentation looks like
if iter_num % 100 == 0 and epoch_num == 0:
x1, y1, x2, y2, _ = data['annot'][0][0]
fig, ax = plt.subplots(1)
ax.imshow(data['img'][0][1])
rect = patches.Rectangle((x1,y1),x2-x1,y2-y1,linewidth=1,edgecolor='r',facecolor='none', alpha=1)
ax.add_patch(rect)
fig.savefig(os.path.join(dir_training_images, '{}.png'.format(iter_num)))
plt.close()
classification_loss, regression_loss = retinanet([data['img'].cuda().float(), data['annot']])
classification_loss = classification_loss.mean()
regression_loss = regression_loss.mean()
loss = classification_loss + regression_loss
if bool(loss == 0):
continue
loss.backward()
if parser.resnet_backbone_normalization == 'batch_norm':
torch.nn.utils.clip_grad_norm_(parameters=retinanet.parameters(), max_norm=0.1)
else:
torch.nn.utils.clip_grad_norm_(parameters=retinanet.parameters(), max_norm=0.01) # Decrease norm to reduce risk of exploding gradients
optimizer.step()
loss_hist.append(float(loss))
epoch_loss.append(float(loss))
print('Epoch: {} | Iteration: {} | Classification loss: {:1.5f} | Regression loss: {:1.5f} | Running loss: {:1.5f}'.format(epoch_num, iter_num, float(classification_loss), float(regression_loss), np.mean(loss_hist)))
del classification_loss
del regression_loss
except Exception as e:
print(e)
continue
writer.add_scalar('Train/Loss', np.mean(epoch_loss), epoch_num)
if not parser.train_all_labeled_data:
print('Evaluating Validation Loss...')
with torch.no_grad():
retinanet.train()
val_losses, val_class_losses, val_reg_losses = [],[],[]
for val_iter_num, val_data in enumerate(dataloader_val):
try:
val_classification_loss, val_regression_loss = retinanet([val_data['img'].cuda().float(), val_data['annot']])
val_losses.append(float(val_classification_loss)+float(val_regression_loss))
val_class_losses.append(float(val_classification_loss))
val_reg_losses.append(float(val_regression_loss))
del val_classification_loss, val_regression_loss
except Exception as e:
print(e)
continue
print('VALIDATION Epoch: {} | Classification loss: {:1.5f} | Regression loss: {:1.5f} | Total loss: {:1.5f}'.format(epoch_num, np.mean(val_class_losses), np.mean(val_reg_losses), np.mean(val_losses)))
# Save model with best validation loss
if best_validation_loss is None:
best_validation_loss = np.mean(val_losses)
if best_validation_loss >= np.mean(val_losses):
best_validation_loss = np.mean(val_losses)
torch.save(retinanet.module, parser.model_output_dir + '/best_result_valloss.pt')
writer.add_scalar('Validation/Loss', np.mean(val_losses), epoch_num)
# Calculate Validation mAP
print('Evaluating validation mAP')
mAP = csv_eval.evaluate(dataset_val, retinanet)
print("Validation mAP: " + str(mAP[0][0]))
if best_validation_map is None:
best_validation_map = mAP[0][0]
elif best_validation_map < mAP[0][0]:
best_validation_map = mAP[0][0]
torch.save(retinanet.module, parser.model_output_dir + '/best_result_valmAP.pt')
writer.add_scalar('Validation/mAP', mAP[0][0], epoch_num)
if not parser.train_all_labeled_data:
scheduler.step(np.mean(val_losses))
else:
scheduler.step()
torch.save(retinanet.module, parser.model_output_dir + '/retinanet_{}.pt'.format(epoch_num))
retinanet.eval()
torch.save(retinanet, parser.model_output_dir + '/model_final.pt')
if __name__ == '__main__':
main()