forked from Domben93/PMSEsegmentation
-
Notifications
You must be signed in to change notification settings - Fork 0
/
random_search.py
323 lines (254 loc) · 13.2 KB
/
random_search.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
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
import importlib
import math
import os.path
import random
from typing import List
import torch
import argparse
import datetime
import time
import utils.loss
from datetime import datetime, timedelta
from utils import transforms as t
from utils.dataset import get_dataloader
from utils.utils import *
from models.utils import *
from torch.utils.tensorboard import SummaryWriter
from utils.metrics import mIoU, SegMets
import seaborn as sns
import pandas as pd
import matplotlib.pyplot as plt
from torchmetrics.functional.classification import binary_jaccard_index
def main(args):
print(f'Preparing training of model')
config = load_yaml_as_dotmap(args.config_path)
device = (torch.device(config.gpu) if torch.cuda.is_available() else torch.device('cpu'))
augmentation_list = [t.RandomHorizontalFlip(0.25),
t.RandomVerticalFlip(0.25)]
train_pair_compose = t.PairCompose([
# [t.RandomHorizontalFlip(0.5)],
# [t.RandomContrastAdjust(0.5, (0.8, 1.2))],
# [t.RandomBrightnessAdjust(0.5, (0.5, 1.5))],
# [t.ObjectAugmentation(augmentation_list)],
[t.ConvertDtype(torch.float32), t.ConvertDtype(torch.float32)],
[t.Normalize((0, 1), (0, 255), return_type=torch.float32), None],
[t.QuasiResize(config.dataset.resize_shape, config.dataset.max_scale),
t.QuasiResize(config.dataset.resize_shape, config.dataset.max_scale)],
])
val_pair_compose = t.PairCompose([
[t.ConvertDtype(torch.float32), t.ConvertDtype(torch.float32)],
[t.Normalize((0, 1), (0, 255), return_type=torch.float32), None],
[t.QuasiResize(config.dataset.resize_shape, config.dataset.max_scale),
t.QuasiResize(config.dataset.resize_shape, config.dataset.max_scale)],
])
train_loader = get_dataloader(args.config_path, train_pair_compose)
if config.validation.validation_interval is not None:
validate_loader = get_dataloader(args.config_path, val_pair_compose, mode='validate')
else:
validate_loader = None
metrics = SegMets()
best_model_state = {}
best_optim_state = {}
print(f'Training on device: {device}')
pd_data = {}
search_res = {}
sample_classes = [['MAD6400_2008-07-02_arcd_60@vhf_400749', 'MAD6400_2009-06-10_arcd_60@vhf_422844',
'MAD6400_2009-07-16_manda_60@vhf_655441', 'MAD6400_2011-06-01_manda_59',
'MAD6400_2015-08-10_manda_59',
'MAD6400_2015-08-13_manda_59', 'MAD6400_2015-08-20_manda_59'],
['MAD6400_2008-06-30_manda_60@vhf_060693', 'MAD6400_2009-07-14_manda_60@vhf_684778',
'MAD6400_2009-07-17_manda_60@vhf_633279', 'MAD6400_2009-07-30_manda_60@vhf_779294',
'MAD6400_2010-07-08_manda_59',
'MAD6400_2010-07-09_manda_60@vhf_470083', 'MAD6400_2011-06-08_manda_59',
'MAD6400_2011-06-09_manda_59',
'MAD6400_2014-07-01_manda_48@vhf_178333', 'MAD6400_2015-08-12_manda_59'],
['MAD6400_2010-07-07_manda_60@vhf_576698']]
sample_classes = config.sample_classes.class_names
writer_name = f'{str(config.model.model_type)}_{str(config.model_init.init_features)}_' \
f'pretrain-{bool(config.model_init.pre_trained_weights)}_' \
f'freezed-{str(config.model_init.freeze_layers)}_' \
f'loss-{str(config.optimizer.loss_type)}_' \
f'optim-{str(config.optimizer.optim_type)}'
if not os.path.exists(os.path.join(config.model.save_path, writer_name)):
os.mkdir(os.path.join(config.model.save_path, writer_name))
if config.training.number_of_runs < 1:
raise ValueError(f'Number of runs must be bigger than 1. Got {config.training.number_of_runs}.')
lr_range = [0.01, 0.0001]
weight_decay = [0.01, 0.0001]
beta1 = [0.1, 0.0001]
beta2 = config.optimizer.betas[1]
param1: List = lr_range
param2: List = weight_decay
number_of_searches = 25
validation_num = 1
stop_patience = 10
stop_min_delta = 0
total_best_loss = math.inf
for i in range(number_of_searches):
lr = random.uniform(param1[0], param1[1])
weight_decay = random.uniform(param2[0], param2[1])
pd_data[i] = {}
search_res[i] = {}
variance_states = [{'model_state': None, 'optim_state': None} for i in range(validation_num)]
found_best_flag = False
for n in range(validation_num):
best_loss_variance_run = math.inf
search_res[i][n] = {'lr': lr,
'beta1': weight_decay,
'loss': 0,
'miou': 0,
'dsc': 0,
'auc': 0,
'acc': 0,
'epoch': 0
}
pd_data[i][n] = {
'epoch': [],
'run': [],
'iou': [],
'dsc': [],
'acc': [],
'auc': [],
'loss': [],
'event': [],
'lr': lr,
'beta1': weight_decay,
}
model = load_model(args.config_path)
model.to(device)
optimizer, lr_scheduler = load_optimizer(config, model, grad_true_only=True, lr=lr,
weight_decay=weight_decay, betas=[0.9, 0.999])
loss_func = getattr(importlib.import_module('utils.loss'), config.optimizer.loss_type)()
print(f'Starting Training of {type(model).__name__} with \n'
f'--> Optimizer: {type(optimizer).__name__}\n'
f'--> Optimizer Params: {optimizer.defaults}'
f'--> loss function: {type(loss_func).__name__}\n'
f'--> learning schedule: Step size: {config.learning_scheduler.step_size},'
f' gamma: {config.learning_scheduler.gamma}, last epoch: {config.learning_scheduler.last_epoch} \n',
f'--> Number of same model run: {i + 1} of {number_of_searches}',
f'--> Number of same random values run: {n + 1} of {validation_num}')
early_stopping = EarlyStopper(patience=stop_patience, min_delta=stop_min_delta)
for epoch in range(1, config.training.epochs + 1):
model.train()
train_loss, val_loss = .0, .0
for num, data in enumerate(train_loader):
x, mask, info = data
x = x.to(device)
mask = mask.to(device)
optimizer.zero_grad()
with torch.set_grad_enabled(mode=True):
predicted = model(x)
if isinstance(predicted, tuple):
predicted = predicted[0]
loss = loss_func(predicted, mask)
info = remove_from_dataname(info['image_name'])
metrics(predicted.detach().cpu(), mask.detach().cpu(), info)
train_loss += loss.item()
loss.backward()
optimizer.step()
lr_scheduler.step()
loss_train = (float(train_loss / len(train_loader)))
miou_train = float(metrics.mIoU(sample_classes=sample_classes))
dsc_train = float(metrics.dice(sample_classes=sample_classes))
acc_train = float(metrics.dice(sample_classes=sample_classes))
auc_train = float(metrics.auc(sample_classes=sample_classes))
metrics.reset()
pd_data[i][n]['epoch'].append(epoch)
pd_data[i][n]['iou'].append(miou_train)
pd_data[i][n]['dsc'].append(dsc_train)
pd_data[i][n]['acc'].append(acc_train)
pd_data[i][n]['auc'].append(auc_train)
pd_data[i][n]['loss'].append(loss_train)
pd_data[i][n]['event'].append('Train')
pd_data[i][n]['run'].append(n)
if validate_loader and (epoch % config.training.eval_interval == 0 or epoch == 1):
model.eval()
for data in validate_loader:
x, mask, info = data
x = x.to(device)
mask = mask.to(device)
with torch.no_grad():
predicted = model(x)
if isinstance(predicted, tuple):
predicted = predicted[0]
loss = loss_func(predicted, mask)
info = remove_from_dataname(info['image_name'])
metrics(predicted.detach().cpu(), mask.detach().cpu(), info)
val_loss += loss.item()
loss_val = (float(val_loss / len(validate_loader)))
miou_val = float(metrics.mIoU(sample_classes=sample_classes))
dsc_val = float(metrics.dice(sample_classes=sample_classes))
acc_val = float(metrics.accuracy(sample_classes=sample_classes))
auc_val = float(metrics.auc(sample_classes=sample_classes))
metrics.reset()
pd_data[i][n]['epoch'].append(epoch)
pd_data[i][n]['iou'].append(miou_val)
pd_data[i][n]['loss'].append(loss_val)
pd_data[i][n]['dsc'].append(dsc_val)
pd_data[i][n]['acc'].append(acc_val)
pd_data[i][n]['auc'].append(auc_val)
pd_data[i][n]['event'].append('Validation')
pd_data[i][n]['run'].append(n)
if epoch == 1 or epoch % 1 == 0:
print(f'{datetime.now()}, Epoch {epoch} of {config.training.epochs}:\n'
f'Training: loss {loss_train:.6f}, mIoU {miou_train:.6f}, dsc {dsc_train:.6f}, AUC {auc_train:.6f}')
if validate_loader:
print(
f'Validation: loss {loss_val:.6f}, mIoU {miou_val:.6f}, dsc {dsc_val:.6f}, AUC {auc_val:.6f}')
if loss_val < best_loss_variance_run and config.training.save_best:
best_variance_run = loss_val
if total_best_loss > best_variance_run:
total_best_loss = best_variance_run
best_model_state = model.state_dict()
best_optim_state = optimizer.state_dict()
found_best_flag = True
variance_states[n]['model_state'] = model.state_dict()
variance_states[n]['optim_state'] = optimizer.state_dict()
search_res[i][n]['miou'] = miou_val
search_res[i][n]['dsc'] = dsc_val
search_res[i][n]['auc'] = auc_val
search_res[i][n]['acc'] = acc_val
search_res[i][n]['loss'] = loss_val
search_res[i][n]['epoch'] = epoch
if early_stopping.early_stop(val_loss):
print('Stopping early!...')
break
if found_best_flag:
pass
#for run, var_pair in enumerate(variance_states):
# save_model(model=model,
# optimizer=optimizer,
# epoch=-1,
# train_samples=-1,
# file_name=os.path.join(f'variance_run_{run}.pt'),
# save_path=os.path.join(config.model.save_path, writer_name),
# model_state_dict=var_pair['model_state'],
# optimizer_state_dict=var_pair['optim_state']
# )
if config.training.save_best:
save_model(model=model,
optimizer=optimizer,
epoch=-1,
train_samples=-1,
file_name=os.path.join('best_random_search.pt'),
save_path=os.path.join(config.model.save_path, writer_name),
model_state_dict=best_model_state,
optimizer_state_dict=best_optim_state)
df = pd.DataFrame(pd_data)
search = pd.DataFrame(search_res)
search.to_csv(f'../Test/results/{writer_name + "_random_search"}.csv', index=False)
df.to_csv(f'../Test/results/{writer_name}.csv', index=False)
if __name__ == '__main__':
torch.manual_seed(666)
parser = argparse.ArgumentParser(description='Training model for segmentation of PMSE signal')
parser.add_argument('--config-path', type=str, default='models\\options\\unet_config.ymal',
help='Path to confg.ymal file (Default unet_config.ymal)')
"""
unet_model = torch.hub.load('mateuszbuda/brain-segmentation-pytorch', 'unet',
in_channels=3, out_channels=1, init_features=32, pretrained=True)
from models.unets import UNet
unet = UNet(3, 1, 32)
unet.init_weights(unet_model.state_dict())
torch.save({'model_state': unet.state_dict()}, 'weights\\mateuszbuda-brain-segmentation-pytorch.pt')
"""
main(parser.parse_args())