-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
271 lines (207 loc) · 10 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
import os
import numpy as np
import pandas as pd
import torch
from torch import nn
from torch import optim
from torch.optim import lr_scheduler
import time
import torchio
#import json
#import cv2
import gc
torch.cuda.empty_cache()
import datetime
from torch.utils.data import Dataset, DataLoader
# import torchvision
#from tqdm.auto import tqdms
import argparse
from src.data.torch_utils import MonkeyEyeballsDataset
from src.models.from_scratch import resnet_for_multimodal_regression as resnet
parser = argparse.ArgumentParser(description='PyTorch ImageNet Training')
parser.add_argument('--labels', default='data/monkey_data.csv', metavar='DF',
help='path to ICP/IOP dataframe')
parser.add_argument('--scans', default='data/torch_standard', metavar='DIR',
help='path to dataset folder')
parser.add_argument('--epochs', default=200, type=int, metavar='N',
help='number of total epochs to run')
parser.add_argument('--lr', default=3e-4, type=float, metavar='LR',
help='initial learning rate')
parser.add_argument('--save', default='models/run_{}'.format(datetime.datetime.today().strftime('%Y-%m-%d-%H-%M')),
type=str, metavar='SAVE_DIR',
help='path to save models and losses')
parser.add_argument('--batch', default=8, type=int, metavar='BATCH',
help='number of samples per mini-batch')
parser.add_argument('--warm_start_batch', default=0, type=int,
help='Batch number to warm start on')
parser.add_argument('--warm_start_epoch', default=0, type=int,
help='Epoch number to warm start on')
parser.add_argument('--warm_start_model', default=None, type=str,
help='Model filepath to warm start on')
def main():
args = parser.parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
seed = 12345
torch.manual_seed(seed)
def train(dataloader_train,
dataloader_val,
model,
optimizer,
scheduler,
val_interval,
save_interval,
save_folder,
warm_start_epoch=0,
warm_start_batch=0,
loss=nn.MSELoss(reduction='sum'),
total_epochs=100):
# settings
batches_per_epoch = len(dataloader_train)
print('{} epochs in total, {} batches per epoch'.format(total_epochs, batches_per_epoch))
if device == 'cuda':
loss = loss.to(device)
model.train()
train_time_sp = time.time()
temp_train = []
temp_val = []
train_loss_epoch = []
val_loss_epoch = []
for epoch in range(args.warm_start_epoch, total_epochs):
print('Start epoch {}'.format(epoch))
for batch_id, batch_data in enumerate(dataloader_train, start=args.warm_start_batch):
# getting data batch
batch_id_sp = epoch * batches_per_epoch + batch_id
icp = batch_data['icp'].float().unsqueeze(1).cuda()
iop = batch_data['iop'].float().cuda()
scan = batch_data['scan'].float().cuda()
if device == 'cuda':
scan = scan.to(device)
# standardize input
icp = (icp - 15) / 11
iop = (iop - 22) / 13
optimizer.zero_grad()
# add fake channel dimension as 5-D input is expected
preds = model(scan.unsqueeze(1),iop)
# calculating loss
loss_value = loss(preds, icp)
loss_value.backward()
optimizer.step()
avg_batch_time = (time.time() - train_time_sp) / (1 + batch_id_sp)
print(
'Batch: {}-{} ({}), loss = {:.3f}, avg_batch_time = {:.3f}'\
.format(epoch, batch_id, batch_id_sp, loss_value, avg_batch_time))
temp_train.append(loss_value.item())
#get validation loss
#if batch_id_sp % val_interval == 0:
if batch_id == len(dataloader_train)-1:
train_loss_epoch.append(np.mean(temp_train[:]))
temp_train.clear()
model.eval()
print('')
print('Validating...')
for batch_id_val, batch_data_val in enumerate(dataloader_val):
icp_val = batch_data_val['icp'].float().unsqueeze(1).cuda()
iop_val = batch_data_val['iop'].float().cuda()
scan_val = batch_data_val['scan'].float().cuda()
# scan_val = (scan_val - 30) / 19
icp_val = (icp_val - 15) / 11
iop_val = (iop_val -22)/ 13
if device == 'cuda':
scan_val = scan_val.to(device)
preds_val = model(scan_val.unsqueeze_(1),iop_val)
loss_value_val = loss(preds_val, icp_val)
temp_val.append(loss_value_val.item())
val_loss_epoch.append(np.mean(temp_val[:]))
temp_val.clear()
print('VAL LOSS EPOCH-----------------------------------------------------')
print(val_loss_epoch)
print('TRAIN LOSS EPOCH-----------------------------------------------------')
print(train_loss_epoch)
np.save(os.path.join(args.save, "val_loss_epoch.npy"), np.asarray(val_loss_epoch))
np.save(os.path.join(args.save, "train_loss_epoch.npy"), np.asarray(train_loss_epoch))
torch.cuda.empty_cache()
gc.collect()
model.train()
# save model
if batch_id_sp != 0 and batch_id_sp % save_interval == 0:
model_save_path = os.path.join(save_folder, 'epoch_{}_batch_{}.pth.tar'\
.format(epoch, batch_id))
model_save_dir = os.path.dirname(model_save_path)
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
print('Save checkpoints: epoch = {}, batch_id = {}'.format(epoch, batch_id))
torch.save({
'epoch': epoch,
'batch_id': batch_id,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()},
model_save_path)
print('Saving to {}'.format(model_save_path))
print('lr = {}'.format(scheduler.get_lr()))
print('Finished training')
labels = pd.read_csv(args.labels)
labels = labels[labels['torch_present'] & ~labels['icp'].isnull() & ~labels['iop'].isnull() & labels['icp'] > 0]
labels['icp'] = labels['icp'].astype('float')
labels['iop'] = labels['iop'].astype('float')
# print(labels)
train_labels = labels[(labels['monkey_id'] != 14) & (labels['monkey_id'] != 9)]
# 8 handpicked examples
val_examples = [1751, 1754, 1761, 1766]
val_labels = labels[labels['id'].isin(val_examples)]
# # get train and val labels
# train_labels =labels.sample(frac=0.99,random_state=200)
# val_labels =labels.drop(train_labels.index)
# print(len(train_labels))
# print(len(val_labels))
#TRANSFORM###############################################################################################
transform = torchio.Compose([
torchio.RandomFlip(axes=2, p=0.5),
torchio.RandomAffine(
degrees=(0, 0, 10),
translation=1
),
torchio.RandomBlur(1, p=0.2),
torchio.RandomNoise(mean=0,std=1),
torchio.RandomGamma(),
torchio.RandomAffine(
scales=(1.2, 1.5)
)
])
#TRANSFORM###############################################################################################
med_train = MonkeyEyeballsDataset(args.scans, train_labels, transform=transform)
med_val = MonkeyEyeballsDataset(args.scans, val_labels)
dataloader_train = DataLoader(med_train, batch_size=args.batch, shuffle=True,pin_memory=True,num_workers=2 )
dataloader_val = DataLoader(med_val, batch_size=4, shuffle=False)
print(len(dataloader_train))
print(len(dataloader_val))
model = resnet.resnet50(sample_input_D=128, sample_input_H=128, sample_input_W=512).cuda()
OPTIMIZER = torch.optim.Adamax(model.parameters(), lr=args.lr)
SCHEDULER = lr_scheduler.ExponentialLR(OPTIMIZER, gamma=0.99)
LOSS = nn.MSELoss(reduction='mean')
if args.warm_start_model is not None:
warm_start = torch.load(args.warm_start_model)
model.load_state_dict(warm_start['state_dict'])
OPTIMIZER.load_state_dict(warm_start['optimizer'])
args.save = os.path.dirname(args.warm_start_model)
# # load in in case of warm start
# warm_start = torch.load('models/models/epoch_0_batch_100.pth.tar')
# model.load_state_dict(warm_start['state_dict'])
# OPTIMIZER.load_state_dict(warm_start['optimizer'])
# if warm_start.get('epoch') is not None:
# current_epoch = warm_start.get('epoch')
# else:
# current_epoch = 0
train(dataloader_train=dataloader_train,
dataloader_val=dataloader_val,
model=model,
optimizer=OPTIMIZER,
scheduler=SCHEDULER,
total_epochs=args.epochs,
warm_start_epoch=args.warm_start_epoch,
warm_start_batch=args.warm_start_batch,
save_interval=159,
save_folder=args.save, # change this for a new run or change to pass it in as command line arg
val_interval=10,
loss=LOSS)
if __name__ == '__main__':
main()