-
Notifications
You must be signed in to change notification settings - Fork 2
/
tune.py
248 lines (215 loc) · 12.6 KB
/
tune.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
import os
import gc
import time
import random
import argparse
from typing import List, Union
import torch
import wandb
import numpy as np
import albumentations as albu
from torch.cuda import device_count
from torch.utils.data import DataLoader
import segmentation_models_pytorch as smp
from tools.models import TuningModel
from tools.datasets import SegmentationDataset
from tools.supervisely_tools import read_supervisely_project
from tools.utils import DynamicWeighting, StaticWeighting
from tools.data_processing import split_data, convert_seconds_to_hms
def main(config=None):
with wandb.init(config=config):
config = wandb.config
run_name = wandb.run.name
print('\033[92m' + '\n********** Run: {:s} **********\n'.format(run_name) + '\033[0m')
# Build dataset
img_paths, ann_paths, dataset_names = read_supervisely_project(sly_project_dir=config.dataset_dir,
included_datasets=config.included_datasets,
excluded_datasets=config.excluded_datasets)
if args.data_fraction_used < 1:
assert 0 < args.data_fraction_used <= 1, 'Fraction of used data should be in range (0; 1]'
random.seed(11)
indexes_to_include = set(random.sample(list(range(len(img_paths))), int(args.data_fraction_used * len(img_paths))))
img_paths = [n for idx, n in enumerate(img_paths) if idx in indexes_to_include]
ann_paths = [n for idx, n in enumerate(ann_paths) if idx in indexes_to_include]
subsets = split_data(img_paths=img_paths,
ann_paths=ann_paths,
dataset_names=dataset_names,
class_name=config.class_name,
seed=11,
ratio=args.ratio,
normal_datasets=['rsna_normal', 'chest_xray_normal'])
preprocessing_params = smp.encoders.get_preprocessing_params(encoder_name=config.encoder_name,
pretrained=config.encoder_weights)
augmentation_params = albu.Compose([
albu.CLAHE(p=0.2),
albu.RandomSizedCrop(min_max_height=(int(0.7*config.input_size), int(0.9*config.input_size)),
height=config.input_size,
width=config.input_size,
w2h_ratio=1.0,
p=0.2),
albu.Rotate(limit=15, p=0.5),
albu.HorizontalFlip(p=0.5),
albu.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.2)
])
datasets = {}
for subset_name in subsets:
_augmentation_params = augmentation_params if subset_name == 'train' else None
dataset = SegmentationDataset(img_paths=subsets[subset_name][0],
ann_paths=subsets[subset_name][1],
input_size=config.input_size,
class_name=config.class_name,
augmentation_params=_augmentation_params,
transform_params=preprocessing_params)
datasets[subset_name] = dataset
# If debug is frozen, use num_workers = 0
num_workers = 8 * device_count()
train_loader = DataLoader(datasets['train'], batch_size=config.batch_size, num_workers=num_workers)
val_loader = DataLoader(datasets['val'], batch_size=config.batch_size, num_workers=num_workers)
test_loader = DataLoader(datasets['test'], batch_size=config.batch_size, num_workers=num_workers)
wandb.log({'train_images': len(train_loader.dataset.img_paths),
'val_images': len(val_loader.dataset.img_paths),
'test_images': len(test_loader.dataset.img_paths)},
commit=False)
aux_params = None
if args.use_cls_head:
aux_params = dict(pooling='avg',
dropout=0.20,
activation='sigmoid',
classes=1)
if not args.use_cls_head:
args.loss_cls = None
weights_strategy = StaticWeighting(w1=1.0, w2=1.0)
# weights_strategy = DynamicWeighting(alpha=0.05)
# Build model
model = TuningModel(model_name=config.model_name,
encoder_name=config.encoder_name,
encoder_weights=config.encoder_weights,
aux_params=aux_params,
batch_size=config.batch_size,
epochs=config.epochs,
input_size=config.input_size,
class_name=config.class_name,
loss_seg=config.loss_seg,
loss_cls=config.loss_cls,
weights_strategy=weights_strategy,
optimizer=config.optimizer,
es_patience=args.es_patience,
es_min_delta=args.es_min_delta,
monitor_metric=config.monitor_metric,
lr=config.lr)
start = time.time()
try:
model.train(train_loader, val_loader, test_loader, logging_loader=None)
except Exception:
print('Run status: CUDA out-of-memory error or HyperBand stop')
else:
print('Run status: Success')
finally:
print('Reset memory and clean garbage')
gc.collect()
torch.cuda.empty_cache()
end = time.time()
print('\033[92m' + '\n********** Run {:s} took {} **********\n'.format(run_name, convert_seconds_to_hms(end - start)) + '\033[0m')
wandb.join()
def get_values(min: int, max: int, step: int, dtype) -> Union[List[int], List[float]]:
if dtype == int:
_values = np.arange(start=min, stop=max + step, step=step, dtype=int)
values = _values.tolist()
elif dtype == float:
_values = np.round(np.arange(start=min, stop=max + step, step=step, dtype=float), 2)
values = _values.tolist()
else:
raise ValueError('Unrecognized dtype!')
return values
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Tuning pipeline')
parser.add_argument('--dataset_dir', default='dataset/covid_segmentation_single_crop', type=str,
help='dataset/covid_segmentation, '
'dataset/covid_segmentation_single_crop, '
'dataset/covid_segmentation_double_crop,'
'dataset/lungs_segmentation')
parser.add_argument('--included_datasets', default=None, type=str)
parser.add_argument('--excluded_datasets', default=None, type=str)
parser.add_argument('--data_fraction_used', default=1.0, type=float)
parser.add_argument('--ratio', nargs='+', default=(0.8, 0.2, 0.0), type=float, help='(train_size, val_size, test_size)')
parser.add_argument('--tuning_method', default='random', type=str, help='grid, random, bayes')
parser.add_argument('--max_runs', default=300, type=int, help='number of trials to run')
parser.add_argument('--batch_size', default=4, type=int)
parser.add_argument('--es_patience', default=6, type=int)
parser.add_argument('--es_min_delta', default=0.01, type=float)
parser.add_argument('--monitor_metric', default='f1_seg', type=str)
parser.add_argument('--epochs', default=16, type=int)
parser.add_argument('--use_cls_head', action='store_true')
parser.add_argument('--wandb_project_name', default=None, type=str)
parser.add_argument('--wandb_api_key', default='b45cbe889f5dc79d1e9a0c54013e6ab8e8afb871', type=str)
args = parser.parse_args()
# Used only for debugging
# args.excluded_datasets = [
# 'covid-chestxray-dataset',
# 'COVID-19-Radiography-Database',
# 'Figure1-COVID-chestxray-dataset',
# 'rsna_normal',
# 'chest_xray_normal'
# ]
if 'covid' in args.dataset_dir:
args.class_name = 'COVID-19'
args.wandb_project_name = 'covid_segmentation_tuning' if not isinstance(args.wandb_project_name, str) else args.wandb_project_name
elif 'lungs' in args.dataset_dir:
args.class_name = 'Lungs'
args.wandb_project_name = 'lungs_segmentation_tuning' if not isinstance(args.wandb_project_name, str) else args.wandb_project_name
else:
raise ValueError('There is no class name for dataset {:s}'.format(args.dataset_dir))
print('\n\033[92m' + 'W&B project name: {:s}\n'.format(args.wandb_project_name) + '\033[0m')
goal = 'minimize' if 'loss' in args.monitor_metric else 'maximize'
os.environ['WANDB_API_KEY'] = args.wandb_api_key
os.environ['WANDB_SILENT'] = "true"
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
sweep_config = {
'method': args.tuning_method,
'metric': {'name': 'val/{:s}'.format(args.monitor_metric), 'goal': goal},
'early_terminate': {'type': 'hyperband', 's': 2, 'eta': 2, 'max_iter': 16}, # 8 (16/2), 4 (16/2/2)
# 'early_terminate': {'type': 'hyperband', 'min_iter': 2, 'eta': 2}, # 2, 4, 8, 16, 32 ...
'parameters': {
# Constant hyperparameters
'dataset_dir': {'value': args.dataset_dir},
'class_name': {'value': args.class_name},
'included_datasets': {'value': args.included_datasets},
'excluded_datasets': {'value': args.excluded_datasets},
'encoder_weights': {'value': 'imagenet'}, # Possible options: imagenet, ssl or sws
'batch_size': {'value': args.batch_size},
'epochs': {'value': args.epochs},
'use_cls_head': {'value': args.use_cls_head},
'monitor_metric': {'value': args.monitor_metric},
# Variable hyperparameters
'model_name': {'values': ['Unet', 'Unet++', 'DeepLabV3', 'DeepLabV3+', 'FPN', 'Linknet', 'PSPNet', 'PAN', 'MAnet']},
# 'model_name': {'values': ['Unet']},
'input_size': {'values': get_values(min=384, max=640, step=32, dtype=int)},
# 'input_size': {'values': [512]},
'loss_seg': {'values': ['Dice', 'Jaccard', 'BCE', 'BCEL', 'Lovasz', 'Focal']},
# 'loss_seg': {'values': ['Dice']},
'loss_cls': {'values': ['BCE', 'SL1', 'L1']},
# 'loss_cls': {'values': ['BCE']},
'optimizer': {'values': ['SGD', 'RMSprop', 'Adam', 'AdamW', 'Adam_amsgrad', 'AdamW_amsgrad']},
# 'optimizer': {'values': ['Adam_amsgrad']},
'lr': {'values': [0.01, 0.005, 0.001, 0.0005, 0.0001]},
# 'lr': {'values': [1e-3]},
# 'encoder_name': {'values': ['resnet18']}
'encoder_name': {'values': ['resnet50', 'resnet101', # ResNet
'resnext50_32x4d', 'resnext101_32x8d', # ResNeXt
'timm-regnetx_032', 'timm-regnetx_064', # RegNet(x)
'timm-regnety_032', 'timm-regnety_064', # RegNet(y)
'se_resnet50', 'se_resnet101', # SE-ResNet
'se_resnext50_32x4d', 'se_resnext101_32x4d', # SE-ResNeXt
'efficientnet-b0', 'efficientnet-b1', 'efficientnet-b2', # EfficientNet
'mobilenet_v2', # MobileNet
'timm-skresnet34', 'timm-skresnext50_32x4d', # SK ResNet
'dpn68', 'dpn98']} # DPN
}
}
sweep_id = wandb.sweep(sweep=sweep_config, entity='viacheslav_danilov', project=args.wandb_project_name)
wandb.agent(sweep_id=sweep_id, function=main, count=args.max_runs)
# If the tuning is interrupted, use a specific sweep_id to keep tuning on the next call
# wandb.agent(sweep_id='cvcok87o', function=main, count=args.max_runs, entity='viacheslav_danilov', project=args.wandb_project_name)
print('\n\033[92m' + '-' * 100 + '\033[0m')
print('\033[92m' + 'Tuning has finished!' + '\033[0m')
print('\033[92m' + '-' * 100 + '\033[0m')