-
Notifications
You must be signed in to change notification settings - Fork 37
/
solver.py
215 lines (181 loc) · 8.72 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
import glob
import os
import numpy as np
import torch
from nn_common_modules import losses as additional_losses
from torch.optim import lr_scheduler
import utils.common_utils as common_utils
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.Adam,
optim_args={},
loss_func=additional_losses.CombinedLoss(),
model_name='quicknat',
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 = optim(model.parameters(), **optim_args)
self.scheduler = lr_scheduler.StepLR(self.optim, step_size=lr_scheduler_step_size,
gamma=lr_scheduler_gamma)
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()
# TODO:Need to correct the CM and dice score calculation.
def train(self, train_loader, val_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, scheduler = self.model, self.optim, self.scheduler
dataloaders = {
'train': train_loader,
'val': val_loader
}
if torch.cuda.is_available():
torch.cuda.empty_cache()
model.cuda(self.device)
print('START TRAINING. : model name = %s, device = %s' % (
self.model_name, torch.cuda.get_device_name(self.device)))
current_iteration = self.start_iteration
for epoch in range(self.start_epoch, self.num_epochs + 1):
print("\n==== Epoch [ %d / %d ] START ====" % (epoch, self.num_epochs))
for phase in ['train', 'val']:
print("<<<= Phase: %s =>>>" % phase)
loss_arr = []
out_list = []
y_list = []
if phase == 'train':
model.train()
scheduler.step()
else:
model.eval()
for i_batch, sample_batched in enumerate(dataloaders[phase]):
X = sample_batched[0].type(torch.FloatTensor)
y = sample_batched[1].type(torch.LongTensor)
w = sample_batched[2].type(torch.FloatTensor)
if model.is_cuda:
X, y, w = X.cuda(self.device, non_blocking=True), y.cuda(self.device,
non_blocking=True), w.cuda(self.device,
non_blocking=True)
output = model(X)
loss = self.loss_func(output, y, w)
if phase == 'train':
optim.zero_grad()
loss.backward()
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(output, dim=1)
out_list.append(batch_output.cpu())
y_list.append(y.cpu())
del X, y, w, output, batch_output, loss
torch.cuda.empty_cache()
if phase == 'val':
if i_batch != len(dataloaders[phase]) - 1:
print("#", end='', flush=True)
else:
print("100%", flush=True)
with torch.no_grad():
out_arr, y_arr = torch.cat(out_list), torch.cat(y_list)
self.logWriter.loss_per_epoch(loss_arr, phase, epoch)
index = np.random.choice(len(dataloaders[phase].dataset.X), 3, replace=False)
self.logWriter.image_per_epoch(model.predict(dataloaders[phase].dataset.X[index], self.device),
dataloaders[phase].dataset.y[index], phase, epoch)
self.logWriter.cm_per_epoch(phase, out_arr, y_arr, epoch)
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
print("==== Epoch [" + str(epoch) + " / " + str(self.num_epochs) + "] DONE ====")
self.save_checkpoint({
'epoch': epoch + 1,
'start_iteration': current_iteration + 1,
'arch': self.model_name,
'state_dict': model.state_dict(),
'optimizer': optim.state_dict(),
'scheduler': scheduler.state_dict(),
'best_ds_mean': self.best_ds_mean,
'best_ds_mean_epoch': self.best_ds_mean_epoch
}, os.path.join(self.exp_dir_path, CHECKPOINT_DIR,
'checkpoint_epoch_' + str(epoch) + '.' + CHECKPOINT_EXTENSION))
print('FINISH.')
self.logWriter.close()
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 Model at Epoch: ' + str(self.best_ds_mean_epoch))
self.load_checkpoint(self.best_ds_mean_epoch)
torch.save(self.model, path)
def save_checkpoint(self, state, filename):
torch.save(state, filename)
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.model.load_state_dict(checkpoint['state_dict'])
self.optim.load_state_dict(checkpoint['optimizer'])
if 'best_ds_mean' in checkpoint.keys():
self.best_ds_mean = checkpoint['best_ds_mean']
self.best_ds_mean_epoch = checkpoint['best_ds_mean_epoch']
for state in self.optim.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.to(self.device)
self.scheduler.load_state_dict(checkpoint['scheduler'])
self.logWriter.log("=> loaded checkpoint '{}' (epoch {})".format(file_path, checkpoint['epoch']))