-
Notifications
You must be signed in to change notification settings - Fork 1
/
main.py
245 lines (203 loc) · 9.96 KB
/
main.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
from __future__ import print_function
import argparse
import os
# internal imports
from utils.file_utils import save_pkl
from utils.utils import *
from utils.core_utils_mtl import train as train_mtl
from utils.core_utils_mtl_stain import train as train_mtl_stain
from datasets.dataset_mtl import Generic_MIL_MTL_Dataset
# pytorch imports
import torch
import pandas as pd
import numpy as np
def main_mtl(args):
print("--------------------------")
print(" MTL ")
print("--------------------------")
# create results directory if necessary
if not os.path.isdir(args.results_dir):
os.mkdir(args.results_dir)
if args.k_start == -1:
start = 0
else:
start = args.k_start
if args.k_end == -1:
end = args.k
else:
end = args.k_end
# arrays to collect scores -- replace by generic one when refactoring
all_task1_test_auc = []
all_task1_val_auc = []
all_task1_test_acc = []
all_task1_val_acc = []
all_task2_test_auc = []
all_task2_val_auc = []
all_task2_test_acc = []
all_task2_val_acc = []
all_task3_test_auc = []
all_task3_val_auc = []
all_task3_test_acc = []
all_task3_val_acc = []
folds = np.arange(start, end)
for i in folds:
seed_torch(args.seed)
train_dataset, val_dataset, test_dataset = dataset.return_splits(from_id=False,
csv_path='{}/splits_{}.csv'.format(args.split_dir, i))
print('training: {}, validation: {}, testing: {}'.format(len(train_dataset), len(val_dataset), len(test_dataset)))
datasets = (train_dataset, val_dataset, test_dataset)
results, \
task1_test_auc, task1_val_auc, task1_test_acc, task1_val_acc, \
task2_test_auc, task2_val_auc, task2_test_acc, task2_val_acc, \
task3_test_auc, task3_val_auc, task3_test_acc, task3_val_acc = train(datasets, i, args)
all_task1_test_auc.append(task1_test_auc)
all_task1_val_auc.append( task1_val_auc )
all_task1_test_acc.append(task1_test_acc)
all_task1_val_acc.append( task1_val_acc )
all_task2_test_auc.append(task2_test_auc)
all_task2_val_auc.append( task2_val_auc )
all_task2_test_acc.append(task2_test_acc)
all_task2_val_acc.append( task2_val_acc )
all_task3_test_auc.append(task3_test_auc)
all_task3_val_auc.append( task3_val_auc )
all_task3_test_acc.append(task3_test_acc)
all_task3_val_acc.append( task3_val_acc )
#write results to pkl
filename = os.path.join(args.results_dir, 'split_{}_results.pkl'.format(i))
save_pkl(filename, results)
final_df = pd.DataFrame({'folds': folds,
'task1_test_auc': all_task1_test_auc, 'task1_val_auc': all_task1_val_auc,
'task1_test_acc': all_task1_test_acc, 'task1_val_acc': all_task1_val_acc,
'task2_test_auc': all_task2_test_auc, 'task2_val_auc': all_task2_val_auc,
'task2_test_acc': all_task2_test_acc, 'task2_val_acc': all_task2_val_acc,
'task3_test_auc': all_task3_test_auc, 'task3_val_auc': all_task3_val_auc,
'task3_test_acc': all_task3_test_acc, 'task3_val_acc': all_task3_val_acc})
if len(folds) != args.k:
save_name = 'summary_partial_{}_{}.csv'.format(start, end)
else:
save_name = 'summary.csv'
final_df.to_csv(os.path.join(args.results_dir, save_name))
# Training settings
parser = argparse.ArgumentParser(description='Configurations for WSI Training')
parser.add_argument('--data_root_dir', type=str, default='',
help='data directory')
parser.add_argument('--max_epochs', type=int, default=200,
help='maximum number of epochs to train (default: 200)')
parser.add_argument('--lr', type=float, default=1e-4,
help='learning rate (default: 0.0001)')
parser.add_argument('--label_frac', type=float, default=1.0,
help='fraction of training labels (default: 1.0)')
parser.add_argument('--bag_weight', type=float, default=0.7,
help='clam: weight coefficient for bag-level loss (default: 0.7)')
parser.add_argument('--reg', type=float, default=1e-5,
help='weight decay (default: 1e-5)')
parser.add_argument('--seed', type=int, default=1,
help='random seed for reproducible experiment (default: 1)')
parser.add_argument('--k', type=int, default=10, help='number of folds (default: 10)')
parser.add_argument('--k_start', type=int, default=-1, help='start fold (default: -1, last fold)')
parser.add_argument('--k_end', type=int, default=-1, help='end fold (default: -1, first fold)')
parser.add_argument('--results_dir', default='./results', help='results directory (default: ./results)')
parser.add_argument('--split_dir', type=str, default=None,
help='manually specify the set of splits to use, '
+'instead of infering from the task and label_frac argument (default: None)')
parser.add_argument('--log_data', action='store_true', default=False, help='log data using tensorboard')
parser.add_argument('--testing', action='store_true', default=False, help='debugging tool')
parser.add_argument('--subtyping', action='store_true', default=False, help='subtyping problem')
parser.add_argument('--early_stopping', action='store_true', default=False, help='enable early stopping')
parser.add_argument('--opt', type=str, choices = ['adam', 'sgd'], default='adam')
parser.add_argument('--drop_out', action='store_true', default=False, help='enabel dropout (p=0.25)')
parser.add_argument('--inst_loss', type=str, choices=['svm', 'ce', None], default=None,
help='instance-level clustering loss function (default: None)')
parser.add_argument('--bag_loss', type=str, choices=['svm', 'ce'], default='ce',
help='slide-level classification loss function (default: ce)')
parser.add_argument('--model_type', type=str, choices=['attention_mil'], default='attention_mil',
help='type of model (default: attention_mil)')
parser.add_argument('--exp_code', type=str, help='experiment code for saving results')
parser.add_argument('--weighted_sample', action='store_true', default=False, help='enable weighted sampling')
parser.add_argument('--model_size', type=str, choices=['small', 'big'], default='small', help='size of model')
parser.add_argument('--patient_level', action='store_true', default=True, help='To enable computing scores at the patient-level. I.e. all patients slides are treated as a single bag with a single label')
parser.add_argument('--stain_level', action='store_true', default=False)
parser.add_argument('--fusion', type=str, default='tensor')
parser.add_argument('--mtl', action='store_true', default=False, help='flag to enable multi-task problem')
parser.add_argument('--task', type=str,
choices=['kidney-mtl'])
args = parser.parse_args()
device=torch.device("cuda" if torch.cuda.is_available() else "cpu")
if args.stain_level:
train = train_mtl_stain
else:
train = train_mtl
print('Stain Level', args.stain_level)
print('Patient Level', args.patient_level)
def seed_torch(seed=7):
import random
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if device.type == 'cuda':
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed) # if you are using multi-GPU.
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
seed_torch(args.seed)
encoding_size = 768
settings = {'num_splits': args.k,
'k_start': args.k_start,
'k_end': args.k_end,
'task': args.task,
'max_epochs': args.max_epochs,
'results_': args.results_dir,
'lr': args.lr,
'experiment': args.exp_code,
'reg': args.reg,
'label_frac': args.label_frac,
'inst_loss': args.inst_loss,
'bag_loss': args.bag_loss,
'bag_weight': args.bag_weight,
'seed': args.seed,
'model_type': args.model_type,
'model_size': args.model_size,
"use_drop_out": args.drop_out,
'weighted_sample': args.weighted_sample,
'opt': args.opt}
print('\nLoad Dataset')
if args.task == 'kidney-mtl':
args.n_classes=[2,2,3]
dataset = Generic_MIL_MTL_Dataset(csv_path = 'dataset_csv/KidneySimpleLabels_edited_all_slides.csv',
data_dir= os.path.join(args.data_root_dir, 'kidney-features'),
shuffle = False,
seed = args.seed,
print_info = True,
label_dicts = [{'no_cell':0, 'cell':1},
{'no_amr':0, 'amr':1},
{'mild_ifta':0, 'moderate_ifta':1, 'advanced_ifta':2}],
label_cols=['label_cell','label_amr','label_ifta'],
patient_strat=False,
ignore=[],
patient_level = args.patient_level,
stain_level = args.stain_level)
else:
raise NotImplementedError
if not os.path.isdir(args.results_dir):
os.mkdir(args.results_dir)
args.results_dir = os.path.join(args.results_dir, str(args.exp_code) + '_s{}'.format(args.seed))
if not os.path.isdir(args.results_dir):
os.mkdir(args.results_dir)
if args.split_dir is None:
args.split_dir = os.path.join('splits', args.task+'_{}'.format(int(args.label_frac*100)))
else:
args.split_dir = os.path.join('splits', args.split_dir)
assert os.path.isdir(args.split_dir)
settings.update({'split_dir': args.split_dir})
with open(args.results_dir + '/experiment_{}.txt'.format(args.exp_code), 'w') as f:
print(settings, file=f)
f.close()
print("################# Settings ###################")
for key, val in settings.items():
print("{}: {}".format(key, val))
if __name__ == "__main__":
if args.mtl:
results = main_mtl(args)
print("finished!")
print("end script")