-
Notifications
You must be signed in to change notification settings - Fork 14
/
level2.py
231 lines (208 loc) · 8.4 KB
/
level2.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
import argparse
from collections import defaultdict
import pickle
import re
import lightgbm as lgb
import pandas as pd
import numpy as np
import xgboost as xgb
from ..data_utils import SEG_FP, get_encoded_classes
from ..utils import print_metrics
from ..metric import get_metrics
from .blend import (
score_predictions_by_image_id, submission_from_predictions_by_image_id)
def main():
parser = argparse.ArgumentParser()
arg = parser.add_argument
arg('detailed_then_features', nargs='+',
help='detailed dataframes and the features in the same order')
arg('--use-xgb', type=int, default=1)
arg('--use-lgb', type=int, default=1)
arg('--num-boost-round', type=int, default=400)
arg('--lr', type=float, default=0.05, help='for lightgbm')
arg('--eta', type=float, default=0.15, help='for xgboost')
arg('--save-model')
arg('--load-model')
arg('--output')
arg('--n-folds', type=int, default=5)
arg('--seg-fp-adjust', type=float)
args = parser.parse_args()
if len(args.detailed_then_features) % 2 != 0:
parser.error('number of detailed and features must be equal')
n = len(args.detailed_then_features) // 2
detailed_paths, feature_paths = (args.detailed_then_features[:n],
args.detailed_then_features[n:])
if args.output:
if not args.load_model:
parser.error('--output needs --load-model')
elif len(feature_paths) == 1:
parser.error('need more than one feature df for train/valid split')
print('\n'.join(
f'{f} | {d}' for f, d in zip(detailed_paths, feature_paths)))
detailed_dfs = [pd.read_csv(path) for path in detailed_paths]
feature_dfs = [pd.read_csv(path) for path in feature_paths]
valid_df = feature_dfs[0]
assert valid_df.columns[0] == 'item'
assert valid_df.columns[-1] == 'y'
feature_cols = [
col for col in valid_df.columns[1:-1] if col not in {
'width', 'height', 'aspect',
'candidate_count', 'candidate_count_on_page',
'candidate_freq_on_page',
}]
top_cls_re = re.compile('^top_\d+_cls')
def build_features(df):
df = df[feature_cols].copy()
for col in feature_cols:
if top_cls_re.match(col):
df[f'{col}_is_candidate'] = df[col] == df['candidate_cls']
# del df[col]
print(' '.join(df.columns))
return df
classes = get_encoded_classes()
cls_by_idx = {idx: cls for cls, idx in classes.items()}
cls_by_idx[-1] = SEG_FP
y_preds = []
all_metrics = []
for fold_num in range(args.n_folds):
print(f'fold {fold_num}')
detailed = (detailed_dfs[fold_num if len(detailed_dfs) != 1 else 0]
.copy())
valid_df = feature_dfs[fold_num if len(feature_dfs) != 1 else 0].copy()
valid_features = build_features(valid_df)
xgb_valid_data = xgb.DMatrix(valid_features, label=valid_df['y'])
fold_path = lambda path, kind: f'{path}.{kind}.fold{fold_num}'
if args.load_model:
lgb_load_path = (fold_path(args.load_model, 'lgb')
if args.use_lgb else None)
xgb_load_path = (fold_path(args.load_model, 'xgb')
if args.use_xgb else None)
print(f'loading from {lgb_load_path}, {xgb_load_path}')
if lgb_load_path:
lgb_model = lgb.Booster(model_file=lgb_load_path)
if xgb_load_path:
with open(xgb_load_path, 'rb') as f:
xgb_model = pickle.load(f)
else:
train_df = pd.concat([df for i, df in enumerate(feature_dfs)
if i != fold_num])
train_features = build_features(train_df)
if args.use_lgb:
lgb_model = train_lgb(
train_features, train_df['y'],
valid_features, valid_df['y'],
lr=args.lr,
num_boost_round=args.num_boost_round)
if args.use_xgb:
xgb_model = train_xgb(
train_features, train_df['y'],
valid_features, valid_df['y'],
eta=args.eta,
num_boost_round=args.num_boost_round)
if args.save_model:
lgb_save_path = (fold_path(args.save_model, 'lgb')
if args.use_lgb else None)
xgb_save_path = (fold_path(args.save_model, 'xgb')
if args.use_xgb else None)
print(f'saving to {lgb_save_path}, {xgb_save_path}')
if lgb_save_path:
lgb_model.save_model(
lgb_save_path, num_iteration=lgb_model.best_iteration)
if xgb_save_path:
with open(xgb_save_path, 'wb') as f:
pickle.dump(xgb_model, f)
print('prediction')
predictions = []
if args.use_lgb:
predictions.append(lgb_model.predict(
valid_features, num_iteration=lgb_model.best_iteration))
if args.use_xgb:
predictions.append(xgb_model.predict(
xgb_valid_data, ntree_limit=xgb_model.best_ntree_limit))
valid_df['y_pred'] = np.mean(predictions, axis=0)
if args.seg_fp_adjust:
valid_df.loc[valid_df['candidate_cls'] == -1, 'y_pred'] += \
args.seg_fp_adjust
y_preds.append(valid_df['y_pred'].values)
max_by_item = get_max_by_item(valid_df)
print('scoring')
detailed['pred'] = \
max_by_item['candidate_cls'].apply(cls_by_idx.__getitem__)
print(f'SEG_FP ratio: {(detailed["pred"] == SEG_FP).mean():.5f}')
predictions_by_image_id = get_predictions_by_image_id(detailed)
if not args.output:
metrics = {
'accuracy': (detailed["pred"] == detailed["true"]).mean(),
}
metrics.update(
score_predictions_by_image_id(predictions_by_image_id))
print_metrics(metrics)
all_metrics.append(metrics)
if args.output:
valid_df['y_pred'] = np.mean(y_preds, axis=0)
max_by_item = get_max_by_item(valid_df)
detailed['pred'] = \
max_by_item['candidate_cls'].apply(cls_by_idx.__getitem__)
predictions_by_image_id = get_predictions_by_image_id(detailed)
submission = submission_from_predictions_by_image_id(
predictions_by_image_id)
submission.to_csv(args.output, index=False)
else:
print('\nAll folds:')
print_metrics(get_metrics(all_metrics))
def train_lgb(train_features, train_y, valid_features, valid_y, *,
lr, num_boost_round):
train_data = lgb.Dataset(train_features, train_y)
valid_data = lgb.Dataset(valid_features, valid_y, reference=train_data)
params = {
'objective': 'binary',
'metric': 'binary_logloss',
'learning_rate': lr,
'bagging_fraction': 0.8,
'bagging_freq': 5,
'feature_fraction': 0.9,
'min_data_in_leaf': 20,
'num_leaves': 41,
'scale_pos_weight': 1.2,
'lambda_l2': 1,
}
print(params)
return lgb.train(
params=params,
train_set=train_data,
num_boost_round=num_boost_round,
early_stopping_rounds=20,
valid_sets=[valid_data],
verbose_eval=10,
)
def train_xgb(train_features, train_y, valid_features, valid_y, *,
eta, num_boost_round):
train_data = xgb.DMatrix(train_features, label=train_y)
valid_data = xgb.DMatrix(valid_features, label=valid_y)
params = {
'eta': eta,
'objective': 'binary:logistic',
'gamma': 0.01,
'max_depth': 8,
}
print(params)
eval_list = [(valid_data, 'eval')]
return xgb.train(
params, train_data, num_boost_round, eval_list,
early_stopping_rounds=20,
verbose_eval=10,
)
def get_max_by_item(df):
return (df.iloc[df.groupby('item')['y_pred'].idxmax()]
.reset_index(drop=True))
def get_predictions_by_image_id(detailed):
predictions_by_image_id = defaultdict(list)
for item in detailed.itertuples():
if item.pred != SEG_FP:
predictions_by_image_id[item.image_id].append({
'cls': item.pred,
'center': (item.x + item.w / 2, item.y + item.h / 2),
})
return predictions_by_image_id
if __name__ == '__main__':
main()