-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy path2-stage-ensemble-5fold.py
254 lines (221 loc) · 10.9 KB
/
2-stage-ensemble-5fold.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
# @Author: Xuan Cao <xuan>
# @Date: 2019-12-22, 1:27:19
# @Last modified by: xuan
# @Last modified time: 2019-12-22, 1:33:01
import warnings
warnings.filterwarnings("ignore")
from pathlib import Path
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import tqdm
from util.utils import *
def compute_detail_score(df, dice):
res = []
res.append(df[dice].mean())
#c1 -> c4 dice
for label in ['Fish', 'Flower', 'Gravel', 'Sugar']:
df_tmp = df[df['cls'] == label]
res.append(df_tmp[dice].mean())
# neg & pos dice
res.append(df[df['truth'] == ''][dice].mean())
res.append(df[df['truth'] != ''][dice].mean())
# c1 -> c4 pos
for label in ['Fish', 'Flower', 'Gravel', 'Sugar']:
df_tmp = df[df['cls'] == label]
res.append(df_tmp[df_tmp['truth'] != ''][dice].mean())
return res
def ensemble_rles(rles1, rles2, mode='intersect'):
res = []
for rle1, rle2 in tqdm.tqdm(zip(rles1, rles2)):
m1 = rle2mask(rle1, height=350, width=525, fill_value=1)
m2 = rle2mask(rle2, height=350, width=525, fill_value=1)
if mode == 'intersect':
mask = ((m1+m2) == 2).astype(int)
elif mode == 'union':
mask = ((m1+m2) > 0).astype(int)
else:
RuntimeError('%s not implemented.'%mode)
rle = mask2rle(mask)
res.append(rle)
return res
def load_stacking(seg_name, tta, ts=0.5):
df_seg_val = pd.read_csv('../output/'+seg_name+'/valid_5fold_tta%d.csv'%tta)
df_seg_test = pd.read_csv('../output/'+seg_name+'/test_5fold_tta%d.csv'%tta)
df_seg_val['s1'], df_seg_test['s1'] = np.nan, np.nan
df_seg_val['s1'].loc[df_seg_val.pred >= ts] = '1 1'
df_seg_test['s1'].loc[df_seg_test.pred >= ts] = '1 1'
return df_seg_val[['Image_Label', 's1']], df_seg_test[['Image_Label', 's1']]
def load_seg_pred(seg_name, name, tta):
#load val
df_val = []
try:
for fold in range(5):
if tta <= 1:
df_val.append(pd.read_csv('../output/'+ seg_name + '/' + 'valid_fold%d.csv'%fold))
else:
df_val.append(pd.read_csv('../output/'+ seg_name + '/' + 'valid_fold%d_tta%d.csv'%(fold, tta)))
df_val = pd.concat(df_val)
except:
df_val = pd.read_csv('../output/'+ seg_name + '/' + 'valid_5fold_tta%d.csv'%(tta))
df_val = df_val[['Image_Label', 'EncodedPixels']]
#df_val.rename(columns={'s3': 'EncodedPixels'}, inplace=True)
df_test = pd.read_csv('../output/'+ seg_name + '/' + 'test_5fold_tta%d.csv'%tta)
df_val.rename(columns={'EncodedPixels': name}, inplace=True)
df_test.rename(columns={'EncodedPixels': name}, inplace=True)
return df_val, df_test
def load_seg_cls_pred(seg_name, name, tta, ts):
#load val
df_val = []
try:
for fold in range(5):
if tta <= 1:
df_val.append(pd.read_csv('../output/'+ seg_name + '/' + 'valid_cls_fold%d.csv'%fold))
else:
df_val.append(pd.read_csv('../output/'+ seg_name + '/' + 'valid_cls_fold%d_tta%d.csv'%(fold, tta)))
df_val = pd.concat(df_val)
except:
df_val = pd.read_csv('../output/'+ seg_name + '/' + 'valid_5fold_tta%d.csv'%(tta))
df_val = df_val[['Image_Label', 'EncodedPixels']]
#df_val.rename(columns={'s3': 'EncodedPixels'}, inplace=True)
df_test = pd.read_csv('../output/'+ seg_name + '/' + 'test_cls_5fold_tta%d.csv'%tta)
df_val['EncodedPixels'] = '1 1'
df_val['EncodedPixels'].loc[df_val['0'] < ts] = np.nan
df_test['EncodedPixels'] = '1 1'
df_test['EncodedPixels'].loc[df_test['0'] < ts] = np.nan
df_val.rename(columns={'EncodedPixels': name}, inplace=True)
df_test.rename(columns={'EncodedPixels': name}, inplace=True)
return df_val, df_test
def load_classifier(classifier, tta):
try:
df_cls_val = []
df_cls_test = []
for fold in range(5):
if tta <= 1:
df_cls_val.append(pd.read_csv('../output/'+ classifier + '/' + 'valid_cls_fold%d.csv'%fold))
df_cls_test.append(pd.read_csv('../output/'+ classifier + '/' + 'test_cls_fold%d.csv'%fold))
else:
df_cls_val.append(pd.read_csv('../output/'+ classifier + '/' + 'valid_cls_fold%d_tta%d.csv'%(fold, tta)))
df_cls_test.append(pd.read_csv('../output/'+ classifier + '/' + 'test_cls_fold%d_tta%d.csv'%(fold, tta)))
df_cls_val = pd.concat(df_cls_val)
df_tmp = df_cls_test[0]
for i in range(1, 5):
assert(np.sum(df_tmp['Image_Label'] != df_cls_test[i]['Image_Label']) == 0)
df_tmp['0'] += df_cls_test[i]['0']
df_tmp['0'] /= 5
df_cls_test = df_tmp
except:
df_cls_val = pd.read_csv('../output/'+ classifier + '/' + 'valid_cls_5fold_tta%d.csv'%tta)
df_cls_test = pd.read_csv('../output/'+ classifier + '/' + 'test_cls_5fold_tta%d.csv'%tta)
df_cls_val.rename(columns={'0': 'prob'}, inplace=True)
df_cls_test.rename(columns={'0': 'prob'}, inplace=True)
return df_cls_val, df_cls_test
df_train = pd.read_csv('../input/train_350.csv')
df_train.rename(columns={'EncodedPixels': 'truth'}, inplace=True)
_save=1
tta=3
seg1 = 'densenet121-FPN-BCE-warmRestart-10x3-bs16'
seg2 = 'b5-Unet-inception-FPN-b7-Unet-b7-FPN-b7-FPNPL'
classifier = 'efficientnetb1-cls-BCE-reduceLR-bs16-PL'
# load classifier results
if classifier:
if 'stacking' in classifier:
df_cls_val = pd.read_csv('../output/'+classifier+'/valid_5fold_tta%d.csv'%tta).rename(columns={'pred': 'prob'})
df_cls_test = pd.read_csv('../output/'+classifier+'/test_5fold_tta%d.csv'%tta).rename(columns={'pred': 'prob'})
else:
df_cls_val, df_cls_test = load_classifier(classifier, tta)
# load seg results
if isinstance(seg1, list):
df_seg1_val, df_seg1_test = load_seg_pred(seg1[0], 's1', tta)
for i in range(1, len(seg1)):
d1, d2 = load_seg_pred(seg1[i], 's1', tta)
df_seg1_val['s1'].loc[d1.s1.isnull()] = np.nan
df_seg1_test['s1'].loc[d2.s1.isnull()] = np.nan
elif 'stacking' in seg1:
df_seg1_val, df_seg1_test = load_stacking(seg1, 3, ts=0.54)
else:
df_seg1_val, df_seg1_test = load_seg_pred(seg1, 's1', 1)
df_seg2_val, df_seg2_test = load_seg_pred(seg2, 's2', tta)
# merge seg valid
df_seg_val = pd.merge(df_seg1_val, df_seg2_val, how='left')
df_seg_val = pd.merge(df_seg_val, df_train, how='left')
if classifier:
df_seg_val = pd.merge(df_seg_val, df_cls_val[['Image_Label', 'prob']], how='left')
df_seg_val['s3'] = df_seg_val['s1'].copy()
df_seg_val['s3'].loc[df_seg_val['s1'].notnull()] = df_seg_val['s2'].loc[df_seg_val['s1'].notnull()]
#df_seg_val['area'] = df_seg_val['s3'].apply(lambda x: rle2mask(x, height=350, width=525).sum())
df_seg_test = pd.merge(df_seg1_test, df_seg2_test, how='left')
df_seg_test = pd.merge(df_seg_test, df_cls_test[['Image_Label', 'prob']], how='left')
# Compute val score
print('Compute dice score.')
df_seg_val.fillna('', inplace=True)
df_seg_val['cls'] = df_seg_val['Image_Label'].apply(lambda x: x.split('_')[-1])
df_seg_val['dice1'] = df_seg_val.apply(lambda x: dice_np_rle(x['s1'], x['truth']), axis=1)
df_seg_val['dice2'] = df_seg_val.apply(lambda x: dice_np_rle(x['s2'], x['truth']), axis=1)
df_seg_val['dice3'] = df_seg_val['dice1'].copy()
df_seg_val['dice3'].loc[df_seg_val['s1']!=''] = df_seg_val['dice2'].loc[df_seg_val['s1']!='']
if classifier:
print('Finding cls threshold for empty')
cls_ts = []
best_dice_channel = []
df_seg_val['dice_cls'] = df_seg_val['dice3'].copy()
for label in ['Fish', 'Flower', 'Gravel', 'Sugar']:
df_tmp = df_seg_val[df_seg_val['cls'] == label].reset_index(drop=True).copy()
best_dice, best_ts = 0, 0
for ts in tqdm.tqdm(range(0, 100, 1)):
ts /= 100
df_tmp['s3'].loc[df_tmp['prob'] <= ts] = ''
df_tmp['dice'] = df_tmp['dice_cls']
df_tmp['dice'].loc[(df_tmp['s3'] == '') & (df_tmp['truth'] =='')] = 1
df_tmp['dice'].loc[(df_tmp['s3'] != '') & (df_tmp['truth'] =='')] = 0
df_tmp['dice'].loc[(df_tmp['s3'] == '') & (df_tmp['truth'] !='')] = 0
current_dice = np.mean(df_tmp['dice'])
if current_dice > best_dice:
best_dice = current_dice
best_ts = ts
cls_ts.append(best_ts)
best_dice_channel.append(best_dice)
print(cls_ts)
df_seg_val['s4'] = df_seg_val['s3'].copy()
cls_mask = cls_ts * (df_seg_val.shape[0]//4)
# remove empty
df_seg_val['s4'].loc[df_seg_val['prob'] <= cls_mask] = ''
df_seg_val['dice4'] = df_seg_val['dice3'].copy()
df_seg_val['dice4'].loc[(df_seg_val['s4'] == '') & (df_seg_val['truth'] =='')] = 1
df_seg_val['dice4'].loc[(df_seg_val['s4'] != '') & (df_seg_val['truth'] =='')] = 0
df_seg_val['dice4'].loc[(df_seg_val['s4'] == '') & (df_seg_val['truth'] !='')] = 0
df_seg_val['dice4'].loc[(df_seg_val['s4'] != '') & (df_seg_val['truth'] !='')] = \
df_seg_val['dice2'].loc[(df_seg_val['s4'] != '') & (df_seg_val['truth'] !='')]
if _save:
df_seg_val.to_csv('../output/ensemble/valid_5fold_tta%d.csv'%(tta), index=None)
print(' dice | c1 | c2 | c3 | c4 | neg | pos | pos1 | pos2 | pos3 | pos4')
res = compute_detail_score(df_seg_val, 'dice1')
res = ','.join('%.5f'%x for x in res)
print('Seg-1 scores: %s'%res)
res = compute_detail_score(df_seg_val, 'dice2')
res = ','.join('%.5f'%x for x in res)
print('Seg-2 scores: %s'%res)
res = compute_detail_score(df_seg_val, 'dice3')
res = ','.join('%.5f'%x for x in res)
print('ensemble scores: %s'%res)
if classifier:
res = compute_detail_score(df_seg_val, 'dice4')
res = ','.join('%.5f'%x for x in res)
print(' cls scores: %s'%res)
removed = np.sum((df_seg_val['prob'] <= cls_mask) & (df_seg_val['s3'] != ''))
tn = np.sum((df_seg_val['prob'] <= cls_mask) & (df_seg_val['s3'] != '') & (df_seg_val['truth'] == ''))
fn = np.sum((df_seg_val['prob'] <= cls_mask) & (df_seg_val['s3'] != '') & (df_seg_val['truth'] != ''))
print('Remove %d channels. TN: %d; FN: %d. ' % (removed, tn, fn))
df_seg_test['EncodedPixels'] = df_seg_test['s1'].copy()
df_seg_test['EncodedPixels'].loc[df_seg_test['s1'].notnull()] = df_seg_test['s2'].loc[df_seg_test['s1'].notnull()]
if _save:
df_seg_test.to_csv('../output/ensemble/test_5fold_tta%d_details.csv'%(tta), index=None)
df_seg_test.to_csv('../output/ensemble/test_5fold_tta%d.csv'%(tta), index=None, columns=['Image_Label', 'EncodedPixels'])
if classifier:
cls_mask = cls_ts * (df_seg_test.shape[0]//4)
print('Test remove %d channels.' % (np.sum((df_seg_test['prob'] <= cls_mask) & (df_seg_test['EncodedPixels'].notnull()))))
df_seg_test['EncodedPixels'].loc[df_seg_test['prob'] <= cls_mask] = np.nan
if _save:
df_seg_test.to_csv('../output/ensemble/test_5fold_tta%d_cls_details.csv'%(tta), index=None)
df_seg_test.to_csv('../output/ensemble/test_5fold_tta%d_cls.csv'%(tta), index=None, columns=['Image_Label', 'EncodedPixels'])