-
Notifications
You must be signed in to change notification settings - Fork 2
/
train_model.py
294 lines (283 loc) · 10.9 KB
/
train_model.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
import torch
import argparse
import json
import os
import logging
from pathlib import Path
from torch.nn.utils import clip_grad_norm_
from torch.utils.tensorboard import SummaryWriter
from timeit import default_timer as timer
from tqdm import tqdm
from models.map_splitter import reconstruct_maps
from models.unet import UNetRes
from utils.utils import (
load_data,
EarlyStopper,
pearson_cc,
logging_related,
peak_signal_to_noise_ratio,
process_config,
)
from models.loss_func import Composite_Loss
def train(conf):
RANDOM_SEED = int(conf.general.seed)
torch.manual_seed(RANDOM_SEED)
torch.cuda.manual_seed(RANDOM_SEED)
device = (
torch.device("cuda:{}".format(conf.general.gpu_id))
if torch.cuda.is_available()
else "cpu"
)
# load data
train_dataloader, val_dataloader = load_data(conf, training=True)
logging.info(
"Total train samples {}, val samples {}".format(
len(train_dataloader), len(val_dataloader)
)
)
# load model, optimizer, scheduler
model = UNetRes(n_blocks=conf.model.n_blocks, act_mode=conf.model.act_mode).to(
device
)
optimizer = torch.optim.AdamW(
model.parameters(), lr=conf.training.lr, weight_decay=conf.training.weight_decay
)
scheduler = torch.optim.lr_scheduler.StepLR(
optimizer,
step_size=conf.training.scheduler_step_size,
gamma=conf.training.lr_decay,
)
# if using PyTorch 2.0, use torch.compile to accelerate the training
if float(torch.__version__[:3]) >= 2.0:
logging.info("Using PyTorch 2.0, use torch.compile to accelerate the training")
model = torch.compile(model)
if conf.training.load_checkpoint:
logging.info(
"Resume training and load model from {}".format(
conf.training.load_checkpoint
)
)
checkpoint = torch.load(conf.training.load_checkpoint)
model.load_state_dict(checkpoint["model_state"])
optimizer.load_state_dict(checkpoint["optimizer"])
scheduler.load_state_dict(checkpoint["scheduler"])
# define loss function
criterion = Composite_Loss(
loss_1_type=conf.training.loss_1_type,
beta=conf.training.smooth_l1_beta,
cc_weight=conf.training.cc_weight,
)
batch_size = conf.training.batch_size
torch.backends.cudnn.benchmark = True
early_stopper = EarlyStopper(patience=6, mode="min")
best_models = []
for epoch in range(1, conf.training.epochs + 1):
model.train()
train_loss, pcc, psnr = 0.0, 0.0, 0.0
train_loss_l1, train_loss_cc = 0.0, 0.0
for x_train, y_train, original_shape, id in tqdm(train_dataloader):
x_train = x_train.squeeze()
y_train = y_train.squeeze()
y_pred = torch.tensor(())
for indx in range(0, x_train.shape[0], batch_size):
if indx + batch_size > x_train.shape[0]:
x_train_partial = x_train[indx:].unsqueeze(dim=1).to(device)
y_train_partial = y_train[indx:].unsqueeze(dim=1).to(device)
else:
x_train_partial = (
x_train[indx : indx + batch_size].unsqueeze(dim=1).to(device)
)
y_train_partial = (
y_train[indx : indx + batch_size].unsqueeze(dim=1).to(device)
)
optimizer.zero_grad(set_to_none=True)
y_pred_partial = model(x_train_partial)
_, _, loss_ = criterion(y_pred_partial, y_train_partial)
loss_.backward()
y_pred = torch.cat(
(y_pred, y_pred_partial.squeeze(dim=1).detach().cpu()), dim=0
)
clip_grad_norm_(model.parameters(), 2)
optimizer.step()
y_pred_recon = reconstruct_maps(
y_pred.numpy(),
original_shape,
box_size=conf.data.box_size,
core_size=conf.data.core_size,
)
y_train_recon = reconstruct_maps(
y_train.detach().cpu().numpy(),
original_shape,
box_size=conf.data.box_size,
core_size=conf.data.core_size,
)
tmp_pcc = pearson_cc(y_pred_recon, y_train_recon)
tmp_psnr = peak_signal_to_noise_ratio(y_pred_recon, y_train_recon)
tmp_loss_l1, tmp_loss_cc, tmp_loss = criterion(
torch.from_numpy(y_pred_recon).to(device),
torch.from_numpy(y_train_recon).to(device),
)
pcc += tmp_pcc
psnr += tmp_psnr
train_loss += tmp_loss.detach().cpu().numpy()
train_loss_l1 += tmp_loss_l1.detach().cpu().numpy()
train_loss_cc += tmp_loss_cc.detach().cpu().numpy()
logging.info(
"Epoch {}, running loss: {:.4f}, EMDB-{} psnr: {:.2f},\n"
"pcc: {:.4f}, l1 loss: {:.4f}, cc loss: {:.4f}".format(
epoch,
tmp_loss,
id[0],
tmp_psnr,
tmp_pcc,
tmp_loss_l1,
tmp_loss_cc,
)
)
scheduler.step()
lr = scheduler.get_last_lr()[0]
train_loss = train_loss / len(train_dataloader)
train_pcc = pcc / len(train_dataloader)
train_psnr = psnr / len(train_dataloader)
writer.add_scalars(
"train",
{
"loss": train_loss,
"pcc": train_pcc,
"psnr": train_psnr,
"lr": lr,
},
epoch,
)
"""
Validation
"""
model.eval()
with torch.no_grad():
val_loss, pcc, psnr = 0.0, 0.0, 0.0
val_loss_l1, val_loss_cc = 0.0, 0.0
for x_val, y_val, original_shape, id in val_dataloader:
x_val = x_val.squeeze()
y_val = y_val.squeeze()
y_val_pred = torch.tensor(())
for indx in range(0, x_val.shape[0], batch_size):
if indx + batch_size > x_val.shape[0]:
x_val_partial = x_val[indx:].unsqueeze(dim=1).to(device)
else:
x_val_partial = (
x_val[indx : indx + batch_size].unsqueeze(dim=1).to(device)
)
y_pred_partial = model(x_val_partial)
y_val_pred = torch.cat(
(y_val_pred, y_pred_partial.squeeze(dim=1).detach().cpu()),
dim=0,
)
y_val_pred_recon = reconstruct_maps(
y_val_pred.numpy(),
original_shape,
box_size=conf.data.box_size,
core_size=conf.data.core_size,
)
y_val_recon = reconstruct_maps(
y_val.numpy(),
original_shape,
box_size=conf.data.box_size,
core_size=conf.data.core_size,
)
tmp_pcc = pearson_cc(y_val_pred_recon, y_val_recon)
tmp_psnr = peak_signal_to_noise_ratio(y_val_pred_recon, y_val_recon)
tmp_loss_l1, tmp_loss_cc, tmp_loss = criterion(
torch.from_numpy(y_val_pred_recon).to(device),
torch.from_numpy(y_val_recon).to(device),
)
pcc += tmp_pcc
psnr += tmp_psnr
val_loss += tmp_loss.detach().cpu().numpy()
val_loss_l1 += tmp_loss_l1.detach().cpu().numpy()
val_loss_cc += tmp_loss_cc.detach().cpu().numpy()
logging.info(
"Epoch {}, running validation loss: {:.4f}, EMDB-{} psnr: {:.2f},\n"
"pcc: {:.4f}, l1 loss: {:.4f}, cc loss: {:.4f}".format(
epoch,
tmp_loss,
id[0],
tmp_psnr,
tmp_pcc,
tmp_loss_l1,
tmp_loss_cc,
)
)
val_loss = val_loss / len(val_dataloader)
val_psnr = psnr / len(val_dataloader)
val_pcc = pcc / len(val_dataloader)
writer.add_scalars(
"val",
{
"loss": val_loss,
"pcc": val_pcc,
"psnr": val_psnr,
},
epoch,
)
if early_stopper.early_stop(val_loss):
logging.info("Early stopping at epoch {}...".format(epoch))
break
if (
len(best_models) < 2
or val_loss < best_models[-1][0]
and not conf.general.debug
): # save the best top-2 models
state = {
"model_state": model.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
}
file_name = (
conf.output_path
+ "/"
+ "Epoch_{}".format(epoch)
+ "_psnr_{:.2f}".format(val_psnr)
+ "_pcc_{:.3f}".format(val_pcc)
)
logging.info("\n------------ Save the best model ------------")
torch.save(state, file_name + ".pt")
# Remove the lowest scoring model if we already have 2 models
if len(best_models) == 2:
_, old_file_name = best_models.pop()
os.remove(old_file_name + ".pt")
# Add the new model to the list and sort it
best_models.append((val_loss, file_name))
best_models.sort(key=lambda x: x[0])
logging.info(
"Epoch {} train loss: {:.4f}, val loss: {:.4f}, lr = {:.1e}\n"
"train psnr: {:.2f}, val psnr {:.2f}\n"
"train pcc: {:.4f}, val pcc: {:.4f}\n".format(
epoch,
train_loss,
val_loss,
lr,
train_psnr,
val_psnr,
train_pcc,
val_pcc,
)
)
if __name__ == "__main__":
start = timer()
parser = argparse.ArgumentParser()
parser.add_argument(
"--config", type=str, default=None, help="Name of training configuration file"
)
args = parser.parse_args()
with open(args.config, "r") as f:
conf = json.load(f)
conf = process_config(conf, config_name=Path(args.config).stem)
"""
logging related part
"""
logging_related(rank=0, output_path=conf.output_path)
writer = SummaryWriter(log_dir=conf.output_path)
train(conf)
writer.flush()
end = timer()
logging.info("Total time used: {:.1f}".format(end - start))