-
Notifications
You must be signed in to change notification settings - Fork 1
/
hcd_fastai.py
158 lines (126 loc) · 5.86 KB
/
hcd_fastai.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
import os, time, json, re
import itertools, argparse, pickle, random
import numpy as np
import pandas as pd
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import fastai
from fastai.vision import *
from fastai.callbacks import *
from fastai.metrics import error_rate
from torchvision.models import *
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import StratifiedKFold
SEED = 2019
path = '../input/histopathologic-cancer-detection/'
output_path = './'
parser = argparse.ArgumentParser()
parser.add_argument('--n-splits', type=int, default=5,
help='splits of n-fold cross validation')
parser.add_argument('--batch-size', type=int, default=128)
parser.add_argument('--img-size', type=int, default=196)
parser.add_argument('--lr', type=float, default=0.01)
parser.add_argument('--epochs', type=int, default=3,
help='number of training epochs')
args = parser.parse_args()
def kfold_wsi(nb_folds, wsi_df, missing_df, random_state=2019):
## cross validation split by wsi id
np.random.seed(random_state)
# wsi attributions
wsi_id, wsis = wsi_df['wsi'], wsi_df['wsi'].unique()
wsi_count = wsi_df.wsi.value_counts()
wsi_count.name = 'count'
wsi_mean = wsi_df.groupby('wsi')['label'].mean()
wsi_mean.name = 'label_mean'
wsi_attr = pd.concat([wsi_count, wsi_mean], axis=1)
wsi_attr.sort_values('count', axis=0, ascending=False, inplace=True)
# stratified split for cv
sample_per_fold = len(wsi_id) // nb_folds
val_wsis = []
for i in range(nb_folds-1):
# create valid set wsi id
print("collecting set %d..." %(i+1))
val_wsi_id2cnt = {}
nb_val = 0
index_to_select = wsi_attr.loc[wsis][wsi_attr.loc[wsis]['label_mean']>=0.5].index
while nb_val < sample_per_fold * 0.4:
w_id = np.random.choice(index_to_select,1)[0]
if w_id not in val_wsi_id2cnt:
val_wsi_id2cnt[w_id] = wsi_attr.loc[w_id]['count']
nb_val += val_wsi_id2cnt[w_id]
index_to_select = wsi_attr.loc[wsis][wsi_attr.loc[wsis]['label_mean']<0.5].index
while nb_val < sample_per_fold:
w_id = np.random.choice(index_to_select,1)[0]
if w_id not in val_wsi_id2cnt:
val_wsi_id2cnt[w_id] = wsi_attr.loc[w_id]['count']
nb_val += val_wsi_id2cnt[w_id]
val_wsis.append(list(val_wsi_id2cnt.keys()))
wsis = list(set(wsis) - set(val_wsis[i]))
print("collecting set {}...".format(nb_folds))
val_wsis.append(wsis)
# handling missing wsi_id image indices
missing_indices = np.array(missing_df.index)
skf = StratifiedKFold(n_splits=nb_folds, random_state=random_state, shuffle=False)
split_idxs = [(tr_idx, val_idx)
for tr_idx, val_idx in skf.split(np.zeros(missing_df.shape[0]), missing_df['label'].values)]
ms_trn_indices = [missing_indices[split_idxs[i][0]] for i in range(nb_folds)]
ms_val_indices = [missing_indices[split_idxs[i][1]] for i in range(nb_folds)]
# generate trn/val indices
train_indices = []
valid_indices = []
for i in range(nb_folds):
valid_indices.append( np.where(wsi_id.isin(val_wsis[i]))[0] )
tr_idx = np.array([])
for j in set(range(nb_folds))-{i}:
tr_idx = np.concatenate([ tr_idx, np.where(wsi_id.isin(val_wsis[j]))[0] ])
tr_idx.sort()
train_indices.append(tr_idx.astype('int'))
# combine 2 set of indices
train_indices = [np.concatenate([tr_idxs, ms_tr_idxs]) for tr_idxs, ms_tr_idxs in zip(train_indices, ms_trn_indices)]
valid_indices = [np.concatenate([val_idxs, ms_val_idxs]) for val_idxs, ms_val_idxs in zip(valid_indices, ms_val_indices)]
return train_indices, valid_indices
def load_meta_data():
train_df = pd.read_csv(path+'train_labels.csv')
wsi_df = pd.read_csv('../input/wsi-id/patch_id_wsi.csv')
df = train_df.merge(wsi_df, how='outer', on='id')
wsi_df = df[~df.wsi.isna()]
missing_df = df[df.wsi.isna()]
return train_df, wsi_df, missing_df
def main(args):
# load meta data
train_df, wsi_df, missing_df = load_meta_data()
# training preparation
train_targets = np.zeros(train_df.shape[0], dtype='int32')
train_preds = np.zeros(train_df.shape[0], dtype='float32') # matrix for the out-of-fold predictions
test_preds = np.zeros(57458, dtype='float32') # matrix for the predictions on the testset
cv_indices = kfold_wsi(args.n_splits, wsi_df, missing_df)
# start training
print()
for i, (trn_idx, val_idx) in enumerate(zip(*cv_indices)):
print(f'Fold {i + 1}')
# prepare data
src = (ImageList.from_df(df=train_df, path=path, folder='train', suffix='.tif')
.split_by_idxs(trn_idx, val_idx)
.label_from_df(cols='label'))
data = (src.transform(tfms=get_transforms(flip_vert=True), size=args.img_size)
.add_test_folder('test')
.databunch(bs=args.batch_size).normalize(imagenet_stats))
# prepare model
learn = cnn_learner(data, models.resnet50, metrics=[accuracy])
learn.model_dir = output_path
# train
learn.fit_one_cycle(args.epochs, max_lr=args.lr, callbacks=[SaveModelCallback(learn, name=f'rn-50-sz196-{i+1}')])
# inference
pred_val, y_val = learn.get_preds(ds_type=DatasetType.Valid)
train_targets[val_idx] = y_val.numpy()
train_preds[val_idx] = pred_val.numpy()[:,1]
pred_test, _ = learn.get_preds(ds_type=DatasetType.Test)
test_preds += pred_test.numpy()[:,1] / args.n_splits
print()
# make submission
print(f'val auc cv score is {roc_auc_score(train_targets, train_preds)}')
submit = pd.DataFrame({'id':[f.split('.')[0] for f in os.listdir(path+'test')], 'label':test_preds})
submit.to_csv('submission.csv', index=False)
if __name__ == '__main__':
main(args)