-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
267 lines (237 loc) · 13.4 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
import argparse
from datetime import datetime
import logging
import os
from shutil import copyfile
import time
import numpy as np
import random
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
import torchvision
import torchvision.transforms as transforms
from data.data_iterators import SimpleISPRSVaihingenDataset, StrongWeakISPRSVaihingenDataset
from losses import SegmentationLoss, ClassLoss
import models.dummynet
import models.unet
from utils.visuals import make_grid
seed = 8671
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
dir_output = os.path.join(os.getenv('HOME'), 'data/crops/')
dir_train_strong = os.path.join(dir_output, 'train_strong')
dir_train_weak = os.path.join(dir_output, 'train_weak')
dir_train_strong_img = os.path.join(dir_output, 'train_strong', 'img')
dir_train_strong_seg = os.path.join(dir_output, 'train_strong', 'seg')
dir_train_weak_img = os.path.join(dir_output, 'train_weak', 'img')
dir_train_weak_seg = os.path.join(dir_output, 'train_weak', 'seg')
dir_val_img = os.path.join(dir_output, 'val', 'img')
dir_val_seg = os.path.join(dir_output, 'val', 'seg')
def train(opt):
start_time = time.time()
if opt.debug:
dir_logs = '{}_{}_DEBUG_{}'.format(opt.model, opt.tag, str(datetime.now().strftime("%Y%m%d_%H-%M-%S")))
else:
dir_logs = '{}_{}_{}'.format(opt.model, opt.tag, str(datetime.now().strftime("%Y%m%d_%H-%M-%S")))
dir_logs = os.path.join(os.getenv('HOME'), 'logs', dir_logs)
os.makedirs(dir_logs)
copyfile('train.py', os.path.join(dir_logs, 'train.py'))
copyfile('models/{}.py'.format(opt.model), os.path.join(dir_logs, '{}.py'.format(opt.model)))
logging.basicConfig(level=logging.INFO if not opt.debug else logging.DEBUG,
handlers=[logging.FileHandler(os.path.join(dir_logs, 'console_output.log')),
logging.StreamHandler()])
logging.getLogger('PIL').setLevel(logging.WARNING)
logger = logging.getLogger('train')
logger.info('Saving model logs into {}'.format(dir_logs))
tb_writer = SummaryWriter(log_dir=dir_logs)
logger.info('Starting training with opt {}'.format(opt))
device = torch.device("cuda:{}".format(opt.gpu) if int(opt.gpu) >= 0 else "cpu")
logger.info('Found available device {}'.format(device))
if opt.no_weak_supervision:
train_dataset = StrongWeakISPRSVaihingenDataset(dir_train_strong, None, transforms.ToTensor())
elif opt.no_strong_supervision:
train_dataset = StrongWeakISPRSVaihingenDataset(None, dir_train_weak, transforms.ToTensor())
else:
train_dataset = StrongWeakISPRSVaihingenDataset(dir_train_strong, dir_train_weak, transforms.ToTensor())
logger.info('Found {} images and segmaps in train_dataset'.format(len(train_dataset)))
train_data_loader = DataLoader(train_dataset, opt.batch_size, shuffle=True)
train_for_eval_dataset = SimpleISPRSVaihingenDataset(dir_train_strong_img, dir_train_strong_seg,
transforms.ToTensor(), preshuffle=True)
logger.info('Found {} images and segmaps in train_for_eval_dataset'.format(len(train_for_eval_dataset)))
train_for_eval_data_loader = DataLoader(train_for_eval_dataset, opt.batch_size, shuffle=False)
val_dataset = SimpleISPRSVaihingenDataset(dir_val_img, dir_val_seg,
transforms.ToTensor(), preshuffle=True)
logger.info('Found {} images and segmaps in val_dataset'.format(len(val_dataset)))
val_data_loader = DataLoader(val_dataset, opt.batch_size, shuffle=False)
if opt.model == 'dummynet':
model = models.dummynet.DummyNet()
elif opt.model == 'unet':
model = models.unet.UNet(logger)
else:
logger.error('Unknown model name! {}'.format(opt.model))
model.to(device)
segmap_loss_func = SegmentationLoss()
class_loss_func = ClassLoss()
optimizer = optim.Adam(model.parameters(), lr=opt.lr)
best_val_loss_strong = np.inf
n_seen_datapoints_before_evaluation = 0
n_eval_iterations = 0
for epoch in range(opt.epochs):
epoch_start_time = time.time()
# Training
for i, data in enumerate(train_data_loader):
model.train()
image = data['image'].to(device)
segmap_gt = data['segmap'].to(device)
class_vec_gt = data['class_vec'].to(device)
supervision = data['strong_supervision'].to(device)
optimizer.zero_grad()
segmap_pred = model(image)
loss_strong = loss_weak = loss_flip = 0
if not opt.no_strong_supervision:
loss_strong = segmap_loss_func(segmap_pred, segmap_gt, supervision)
if not opt.no_weak_supervision:
loss_weak = class_loss_func(segmap_pred, class_vec_gt)
loss_weak *= opt.lambda_weak
if opt.lambda_flip > 0:
with torch.no_grad():
flip_dims = random.choice([[2], [3], (2, 3)])
image_flipped = image.flip(dims=flip_dims)
segmap_flipped_pred = model(image_flipped)
segmap_unflipped_pred = segmap_flipped_pred.flip(dims=flip_dims)
loss_flip = segmap_loss_func(segmap_pred, segmap_unflipped_pred)
loss_flip *= opt.lambda_flip
loss = loss_strong + loss_weak + loss_flip
loss.backward()
optimizer.step()
if opt.debug:
if i == 0:
grid = make_grid(image, segmap_pred, segmap_gt)
tb_writer.add_image('train_debug/images', grid, n_eval_iterations, dataformats='HWC')
tb_writer.close()
logger.debug('train iter {: >4d} | loss {: >2.4f} n_seen_datapoints_before_evaluation {} '.format(
i, loss.item(), n_seen_datapoints_before_evaluation))
n_seen_datapoints_before_evaluation += opt.batch_size
if n_seen_datapoints_before_evaluation >= opt.val_datapoints:
n_seen_datapoints_before_evaluation = 0
# Evaluate on train
model.eval()
train_loss_strong = 0
train_loss_weak = 0
train_loss_flip = 0
for i_eval, data in enumerate(train_for_eval_data_loader):
with torch.no_grad():
image = data['image'].to(device)
segmap_gt = data['segmap'].to(device)
class_vec_gt = data['class_vec'].to(device)
segmap_pred = model(image)
loss_strong = loss_weak = loss_flip = 0
loss_strong = segmap_loss_func(segmap_pred, segmap_gt).item()
loss_weak = class_loss_func(segmap_pred, class_vec_gt).item()
if opt.lambda_flip > 0:
flip_dims = random.choice([[2], [3], (2, 3)])
image_flipped = image.flip(dims=flip_dims)
segmap_flipped_pred = model(image_flipped)
segmap_unflipped_pred = segmap_flipped_pred.flip(dims=flip_dims)
loss_flip = segmap_loss_func(segmap_pred, segmap_unflipped_pred)
train_loss_strong += loss_strong
train_loss_weak += loss_weak
train_loss_flip += loss_flip
if i_eval == 0:
grid = make_grid(image, segmap_pred, segmap_gt)
tb_writer.add_image('train/images', grid, n_eval_iterations, dataformats='HWC')
tb_writer.close()
if opt.debug:
if i_eval >= 0:
break
train_loss_strong /= i_eval+1
train_loss_weak /= i_eval+1
train_loss_flip /= i_eval+1
train_loss = 0
if not opt.no_strong_supervision:
train_loss += train_loss_strong
if not opt.no_weak_supervision:
train_loss += train_loss_weak * opt.lambda_weak
if opt.lambda_flip > 0:
train_loss += train_loss_flip * opt.lambda_flip
# Evaluate on val
model.eval()
val_loss_strong = 0
val_loss_weak = 0
val_loss_flip = 0
for i_eval, data in enumerate(val_data_loader):
with torch.no_grad():
image = data['image'].to(device)
segmap_gt = data['segmap'].to(device)
class_vec_gt = data['class_vec'].to(device)
segmap_pred = model(image)
loss_strong = loss_weak = loss_flip = 0
loss_strong = segmap_loss_func(segmap_pred, segmap_gt)
loss_weak = class_loss_func(segmap_pred, class_vec_gt)
if opt.lambda_flip > 0:
flip_dims = random.choice([[2], [3], (2, 3)])
image_flipped = image.flip(dims=flip_dims)
segmap_flipped_pred = model(image_flipped)
segmap_unflipped_pred = segmap_flipped_pred.flip(dims=flip_dims)
loss_flip = segmap_loss_func(segmap_pred, segmap_unflipped_pred)
val_loss_strong += loss_strong
val_loss_weak += loss_weak
val_loss_flip += loss_flip
if i_eval == 0:
grid = make_grid(image, segmap_pred, segmap_gt)
tb_writer.add_image('val/images', grid, n_eval_iterations, dataformats='HWC')
tb_writer.close()
if opt.debug:
if i_eval >= 0:
break
val_loss_strong /= i_eval+1
val_loss_weak /= i_eval+1
val_loss_flip /= i_eval+1
val_loss = 0
if not opt.no_strong_supervision:
val_loss += val_loss_strong
if not opt.no_weak_supervision:
val_loss += val_loss_weak * opt.lambda_weak
if opt.lambda_flip > 0:
val_loss += val_loss_flip * opt.lambda_flip
if val_loss_strong < best_val_loss_strong:
best_val_loss_strong = val_loss_strong
logger.info('Saving best model with val loss {} at epoch {}'.format(best_val_loss_strong, epoch))
torch.save(model.state_dict(), os.path.join(dir_logs, 'model_weights_best.pth'))
logger.info('Epoch {: >3d} | Eval iter {: >3d} | Train loss {: >2.6f} = strong {: >2.6f} + weak {: >2.6f} + flip {: >2.6f} | Val loss {: >2.6f} = strong {: >2.6f} + weak {: >2.6f} + flip {: >2.6f} | Time: {:.0f}'.format(
epoch, n_eval_iterations, train_loss, train_loss_strong, train_loss_weak, train_loss_flip,
val_loss, val_loss_strong, val_loss_weak, val_loss_flip, time.time() - epoch_start_time))
tb_writer.add_scalar('train/loss', train_loss, n_eval_iterations)
tb_writer.add_scalar('train/loss_strong', train_loss_strong, n_eval_iterations)
tb_writer.add_scalar('train/loss_weak', train_loss_weak, n_eval_iterations)
tb_writer.add_scalar('train/loss_flip', train_loss_flip, n_eval_iterations)
tb_writer.add_scalar('val/loss', val_loss, n_eval_iterations)
tb_writer.add_scalar('val/loss_strong', val_loss_strong, n_eval_iterations)
tb_writer.add_scalar('val/loss_weak', val_loss_weak, n_eval_iterations)
tb_writer.add_scalar('val/loss_flip', val_loss_flip, n_eval_iterations)
tb_writer.close()
n_eval_iterations += 1
logger.info('Finished training in {:.0f}'.format(time.time() - start_time))
if __name__== "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--model', required=True, help='From models folder (e.g. dummynet)')
parser.add_argument('--gpu', default='0', help='GPU to run on (0 or 1) (-1 for cpu)')
parser.add_argument('--epochs', default=10, type=int, help='Number of epochs')
parser.add_argument('--batch_size', default=32, type=int, help='Batch size')
parser.add_argument('--lr', default=0.001, type=float, help='Learning rate')
parser.add_argument('--lambda_weak', default=0.1, type=float, help='Scaling factor of the weakly supervised loss')
parser.add_argument('--lambda_flip', default=0.0, type=float, help='Scaling factor of the flip loss')
parser.add_argument('--no_strong_supervision', action='store_true', help='Skip strong supervision loss')
parser.add_argument('--no_weak_supervision', action='store_true', help='Skip weak supervision loss')
parser.add_argument('--val_datapoints', default=2048, type=int, help='Validate every N images')
parser.add_argument('--tag', default='', help='Tag added to model logs folder name')
parser.add_argument('--debug', '-d', action='store_true', help='Run in debug mode')
opt = parser.parse_args()
train(opt)