-
Notifications
You must be signed in to change notification settings - Fork 15
/
Copy pathsolver.py
303 lines (256 loc) · 12.4 KB
/
solver.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
import glob
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import lr_scheduler
import utils.common_utils as common_utils
from nn_additional_losses import losses
from utils.data_utils import split_batch
from utils.log_utils import LogWriter
CHECKPOINT_DIR = 'checkpoints'
CHECKPOINT_EXTENSION = 'pth.tar'
class Solver(object):
def __init__(self,
model,
exp_name,
device,
num_class,
optim=torch.optim.SGD,
optim_args={},
loss_func=losses.DiceLoss(),
model_name='OneShotSegmentor',
labels=None,
num_epochs=10,
log_nth=5,
lr_scheduler_step_size=5,
lr_scheduler_gamma=0.5,
use_last_checkpoint=True,
exp_dir='experiments',
log_dir='logs'):
self.device = device
self.model = model
self.model_name = model_name
self.labels = labels
self.num_epochs = num_epochs
if torch.cuda.is_available():
self.loss_func = loss_func.cuda(device)
else:
self.loss_func = loss_func
self.optim_c = optim(
[{'params': model.conditioner.parameters(), 'lr': 1e-3, 'momentum': 0.99, 'weight_decay': 0.0001}
], **optim_args)
self.optim_s = optim(
[{'params': model.segmentor.parameters(), 'lr': 1e-3, 'momentum': 0.99, 'weight_decay': 0.0001}
], **optim_args)
self.scheduler_s = lr_scheduler.StepLR(self.optim_s, step_size=10,
gamma=0.1)
self.scheduler_c = lr_scheduler.StepLR(self.optim_c, step_size=10,
gamma=0.001)
exp_dir_path = os.path.join(exp_dir, exp_name)
common_utils.create_if_not(exp_dir_path)
common_utils.create_if_not(os.path.join(exp_dir_path, CHECKPOINT_DIR))
self.exp_dir_path = exp_dir_path
self.log_nth = log_nth
self.logWriter = LogWriter(
num_class, log_dir, exp_name, use_last_checkpoint, labels)
self.use_last_checkpoint = use_last_checkpoint
self.start_epoch = 1
self.start_iteration = 1
self.best_ds_mean = 0
self.best_ds_mean_epoch = 0
if use_last_checkpoint:
self.load_checkpoint()
def train(self, train_loader, test_loader):
"""
Train a given model with the provided data.
Inputs:
- train_loader: train data in torch.utils.data.DataLoader
- val_loader: val data in torch.utils.data.DataLoader
"""
model, optim_c, optim_s, scheduler_c, scheduler_s = self.model, self.optim_c, self.optim_s, self.scheduler_c, self.scheduler_s
data_loader = {
'train': train_loader,
'val': test_loader
}
if torch.cuda.is_available():
torch.cuda.empty_cache()
model.cuda(self.device)
self.logWriter.log('START TRAINING. : model name = %s, device = %s' % (
self.model_name, torch.cuda.get_device_name(self.device)))
current_iteration = self.start_iteration
warm_up_epoch = 15
val_old = 0
change_model = False
current_model = 'seg'
for epoch in range(self.start_epoch, self.num_epochs + 1):
self.logWriter.log(
'train', "\n==== Epoch [ %d / %d ] START ====" % (epoch, self.num_epochs))
if epoch > warm_up_epoch:
if current_model == 'seg':
self.logWriter.log("Optimizing Segmentor")
optim = optim_s
elif current_model == 'con':
optim = optim_c
self.logWriter.log("Optimizing Conditioner")
for phase in ['train', 'val']:
self.logWriter.log("<<<= Phase: %s =>>>" % phase)
loss_arr = []
input_img_list = []
y_list = []
out_list = []
condition_input_img_list = []
condition_y_list = []
if phase == 'train':
model.train()
scheduler_c.step()
scheduler_s.step()
else:
model.eval()
for i_batch, sampled_batch in enumerate(data_loader[phase]):
X = sampled_batch[0].type(torch.FloatTensor)
y = sampled_batch[1].type(torch.LongTensor)
w = sampled_batch[2].type(torch.FloatTensor)
query_label = data_loader[phase].batch_sampler.query_label
input1, input2, y1, y2 = split_batch(
X, y, int(query_label))
condition_input = torch.cat(
(input1, y1.unsqueeze(1)), dim=1)
query_input = input2
y1 = y1.type(torch.LongTensor)
if model.is_cuda:
condition_input, query_input, y2, y1 = condition_input.cuda(self.device,
non_blocking=True), query_input.cuda(
self.device,
non_blocking=True), y2.cuda(
self.device, non_blocking=True), y1.cuda(
self.device, non_blocking=True)
weights = model.conditioner(condition_input)
output = model.segmentor(query_input, weights)
loss = self.loss_func(F.softmax(output, dim=1), y2)
optim_s.zero_grad()
optim_c.zero_grad()
loss.backward()
if phase == 'train':
if epoch <= warm_up_epoch:
optim_s.step()
optim_c.step()
elif epoch > warm_up_epoch and change_model:
optim.step()
if i_batch % self.log_nth == 0:
self.logWriter.loss_per_iter(
loss.item(), i_batch, current_iteration)
current_iteration += 1
loss_arr.append(loss.item())
_, batch_output = torch.max(
F.softmax(output, dim=1), dim=1)
out_list.append(batch_output.cpu())
input_img_list.append(input2.cpu())
y_list.append(y2.cpu())
condition_input_img_list.append(input1.cpu())
condition_y_list.append(y1)
del X, y, w, output, batch_output, loss, input1, input2, y2
torch.cuda.empty_cache()
if phase == 'val':
if i_batch != len(data_loader[phase]) - 1:
print("#", end='', flush=True)
else:
print("100%", flush=True)
if phase == 'train':
self.logWriter.log('saving checkpoint ....')
self.save_checkpoint({
'epoch': epoch + 1,
'start_iteration': current_iteration + 1,
'arch': self.model_name,
'state_dict': model.state_dict(),
'optimizer_c': optim_c.state_dict(),
'scheduler_c': scheduler_c.state_dict(),
'optimizer_s': optim_s.state_dict(),
'best_ds_mean_epoch': self.best_ds_mean_epoch,
'scheduler_s': scheduler_s.state_dict()
}, os.path.join(self.exp_dir_path, CHECKPOINT_DIR,
'checkpoint_epoch_' + str(epoch) + '.' + CHECKPOINT_EXTENSION))
with torch.no_grad():
input_img_arr = torch.cat(input_img_list)
y_arr = torch.cat(y_list)
out_arr = torch.cat(out_list)
condition_input_img_arr = torch.cat(
condition_input_img_list)
condition_y_arr = torch.cat(condition_y_list)
current_loss = self.logWriter.loss_per_epoch(
loss_arr, phase, epoch)
if phase == 'val':
if epoch > warm_up_epoch:
self.logWriter.log(
"Diff : " + str(current_loss - val_old))
change_model = (current_loss - val_old) > 0.001
if change_model and current_model == 'seg':
self.logWriter.log("Setting to con")
current_model = 'con'
elif change_model and current_model == 'con':
self.logWriter.log("Setting to seg")
current_model = 'seg'
val_old = current_loss
index = np.random.choice(len(out_arr), 3, replace=False)
self.logWriter.image_per_epoch(out_arr[index], y_arr[index], phase, epoch, additional_image=(
input_img_arr[index], condition_input_img_arr[index], condition_y_arr[index]))
ds_mean = self.logWriter.dice_score_per_epoch(
phase, out_arr, y_arr, epoch)
if phase == 'val':
if ds_mean > self.best_ds_mean:
self.best_ds_mean = ds_mean
self.best_ds_mean_epoch = epoch
self.logWriter.log(
"==== Epoch [" + str(epoch) + " / " + str(self.num_epochs) + "] DONE ====")
self.logWriter.log('FINISH.')
self.logWriter.close()
def save_checkpoint(self, state, filename):
torch.save(state, filename)
def save_best_model(self, path):
"""
Save model with its parameters to the given path. Conventionally the
path should end with "*.model".
Inputs:
- path: path string
"""
print('Saving model... %s' % path)
print("Best Epoch... " + str(self.best_ds_mean_epoch))
self.load_checkpoint(self.best_ds_mean_epoch)
torch.save(self.model, path)
def load_checkpoint(self, epoch=None):
if epoch is not None:
checkpoint_path = os.path.join(self.exp_dir_path, CHECKPOINT_DIR,
'checkpoint_epoch_' + str(epoch) + '.' + CHECKPOINT_EXTENSION)
self._load_checkpoint_file(checkpoint_path)
else:
all_files_path = os.path.join(
self.exp_dir_path, CHECKPOINT_DIR, '*.' + CHECKPOINT_EXTENSION)
list_of_files = glob.glob(all_files_path)
if len(list_of_files) > 0:
checkpoint_path = max(list_of_files, key=os.path.getctime)
self._load_checkpoint_file(checkpoint_path)
else:
self.logWriter.log(
"=> no checkpoint found at '{}' folder".format(os.path.join(self.exp_dir_path, CHECKPOINT_DIR)))
def _load_checkpoint_file(self, file_path):
self.logWriter.log("=> loading checkpoint '{}'".format(file_path))
checkpoint = torch.load(file_path)
self.start_epoch = checkpoint['epoch']
self.start_iteration = checkpoint['start_iteration']
self.best_ds_mean_epoch = checkpoint['best_ds_mean_epoch']
self.model.load_state_dict(checkpoint['state_dict'])
self.optim_c.load_state_dict(checkpoint['optimizer_c'])
self.optim_s.load_state_dict(checkpoint['optimizer_s'])
for state in self.optim_c.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.to(self.device)
for state in self.optim_s.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.to(self.device)
self.scheduler_c.load_state_dict(checkpoint['scheduler_c'])
self.scheduler_s.load_state_dict(checkpoint['scheduler_s'])
self.logWriter.log("=> loaded checkpoint '{}' (epoch {})".format(
file_path, checkpoint['epoch']))