-
Notifications
You must be signed in to change notification settings - Fork 364
/
eval.py
executable file
·144 lines (126 loc) · 6.29 KB
/
eval.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
from __future__ import print_function
import numpy as np
import argparse
import torch
import torch.nn as nn
import pdb
import os
import pandas as pd
from utils.utils import *
from math import floor
import matplotlib.pyplot as plt
from dataset_modules.dataset_generic import Generic_WSI_Classification_Dataset, Generic_MIL_Dataset, save_splits
import h5py
from utils.eval_utils import *
# Training settings
parser = argparse.ArgumentParser(description='CLAM Evaluation Script')
parser.add_argument('--data_root_dir', type=str, default=None,
help='data directory')
parser.add_argument('--results_dir', type=str, default='./results',
help='relative path to results folder, i.e. '+
'the directory containing models_exp_code relative to project root (default: ./results)')
parser.add_argument('--save_exp_code', type=str, default=None,
help='experiment code to save eval results')
parser.add_argument('--models_exp_code', type=str, default=None,
help='experiment code to load trained models (directory under results_dir containing model checkpoints')
parser.add_argument('--splits_dir', type=str, default=None,
help='splits directory, if using custom splits other than what matches the task (default: None)')
parser.add_argument('--model_size', type=str, choices=['small', 'big'], default='small',
help='size of model (default: small)')
parser.add_argument('--model_type', type=str, choices=['clam_sb', 'clam_mb', 'mil'], default='clam_sb',
help='type of model (default: clam_sb)')
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('--fold', type=int, default=-1, help='single fold to evaluate')
parser.add_argument('--micro_average', action='store_true', default=False,
help='use micro_average instead of macro_avearge for multiclass AUC')
parser.add_argument('--split', type=str, choices=['train', 'val', 'test', 'all'], default='test')
parser.add_argument('--task', type=str, choices=['task_1_tumor_vs_normal', 'task_2_tumor_subtyping'])
parser.add_argument('--drop_out', type=float, default=0.25, help='dropout')
parser.add_argument('--embed_dim', type=int, default=1024)
args = parser.parse_args()
device=torch.device("cuda" if torch.cuda.is_available() else "cpu")
args.save_dir = os.path.join('./eval_results', 'EVAL_' + str(args.save_exp_code))
args.models_dir = os.path.join(args.results_dir, str(args.models_exp_code))
os.makedirs(args.save_dir, exist_ok=True)
if args.splits_dir is None:
args.splits_dir = args.models_dir
assert os.path.isdir(args.models_dir)
assert os.path.isdir(args.splits_dir)
settings = {'task': args.task,
'split': args.split,
'save_dir': args.save_dir,
'models_dir': args.models_dir,
'model_type': args.model_type,
'drop_out': args.drop_out,
'model_size': args.model_size}
with open(args.save_dir + '/eval_experiment_{}.txt'.format(args.save_exp_code), 'w') as f:
print(settings, file=f)
f.close()
print(settings)
if args.task == 'task_1_tumor_vs_normal':
args.n_classes=2
dataset = Generic_MIL_Dataset(csv_path = 'dataset_csv/tumor_vs_normal_dummy_clean.csv',
data_dir= os.path.join(args.data_root_dir, 'tumor_vs_normal_resnet_features'),
shuffle = False,
print_info = True,
label_dict = {'normal_tissue':0, 'tumor_tissue':1},
patient_strat=False,
ignore=[])
elif args.task == 'task_2_tumor_subtyping':
args.n_classes=3
dataset = Generic_MIL_Dataset(csv_path = 'dataset_csv/tumor_subtyping_dummy_clean.csv',
data_dir= os.path.join(args.data_root_dir, 'tumor_subtyping_resnet_features'),
shuffle = False,
print_info = True,
label_dict = {'subtype_1':0, 'subtype_2':1, 'subtype_3':2},
patient_strat= False,
ignore=[])
# elif args.task == 'tcga_kidney_cv':
# args.n_classes=3
# dataset = Generic_MIL_Dataset(csv_path = 'dataset_csv/tcga_kidney_clean.csv',
# data_dir= os.path.join(args.data_root_dir, 'tcga_kidney_20x_features'),
# shuffle = False,
# print_info = True,
# label_dict = {'TCGA-KICH':0, 'TCGA-KIRC':1, 'TCGA-KIRP':2},
# patient_strat= False,
# ignore=['TCGA-SARC'])
else:
raise NotImplementedError
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
if args.fold == -1:
folds = range(start, end)
else:
folds = range(args.fold, args.fold+1)
ckpt_paths = [os.path.join(args.models_dir, 's_{}_checkpoint.pt'.format(fold)) for fold in folds]
datasets_id = {'train': 0, 'val': 1, 'test': 2, 'all': -1}
if __name__ == "__main__":
all_results = []
all_auc = []
all_acc = []
for ckpt_idx in range(len(ckpt_paths)):
if datasets_id[args.split] < 0:
split_dataset = dataset
else:
csv_path = '{}/splits_{}.csv'.format(args.splits_dir, folds[ckpt_idx])
datasets = dataset.return_splits(from_id=False, csv_path=csv_path)
split_dataset = datasets[datasets_id[args.split]]
model, patient_results, test_error, auc, df = eval(split_dataset, args, ckpt_paths[ckpt_idx])
all_results.append(all_results)
all_auc.append(auc)
all_acc.append(1-test_error)
df.to_csv(os.path.join(args.save_dir, 'fold_{}.csv'.format(folds[ckpt_idx])), index=False)
final_df = pd.DataFrame({'folds': folds, 'test_auc': all_auc, 'test_acc': all_acc})
if len(folds) != args.k:
save_name = 'summary_partial_{}_{}.csv'.format(folds[0], folds[-1])
else:
save_name = 'summary.csv'
final_df.to_csv(os.path.join(args.save_dir, save_name))