-
Notifications
You must be signed in to change notification settings - Fork 0
/
metrics.py
239 lines (190 loc) · 9.93 KB
/
metrics.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
from __future__ import annotations
import functools
import os
from typing import List, Optional, Tuple
import xgboost as xgb
import numpy as np
import pandas as pd
import torch
from sklearn.metrics import (accuracy_score, f1_score, precision_score,
recall_score, roc_auc_score)
from torch import Tensor
from model import find_idx
real_dir = os.path.dirname(__file__)
class Metrics:
def __init__(self, name: str):
self.name = name
self.metrics = ['acc', 'auc', 'f1', 'macro-f1', 'prec', 'recall']
self.src: List[Tensor] = []
self.dst: List[Optional[Tensor]] = []
self.emb: List[Tensor] = []
self.label: List[Tensor] = []
self.pred: List[Tensor] = []
self.acc: float = -1.0
self.auc: float = -1.0
self.f1: float = -1.0
self.marco_f1: float = -1.0
self.prec: float = -1.0
self.recall: float = -1.0
def get_metrics(self):
if len(self.label) == 0 or len(self.pred) == 0:
# logging.warning('No label or pred found.')
self.acc = -1
self.auc = -1
self.f1 = -1
self.prec = -1
self.recall = -1
return [0.0] * len(self.metrics)
label = np.array(torch.cat(self.label, dim=0).numpy()).astype(np.int64)
pred = np.array(torch.cat(self.pred, dim=0).numpy()).astype(np.float32)
pred_label = np.array(pred > 0.5).astype(np.int64)
marco_f1_func = functools.partial(f1_score, average='macro')
# print(f'{self.name:<5s}', 'pred ', (pred_label==1).sum(), pred_label.shape, pred_label[:10].tolist())
# print(f'{self.name:<5s}', 'label', (label==1).sum(), label.shape, label[:10].tolist())
self.acc = float(accuracy_score(label, pred_label))
self.auc = float(roc_auc_score(label, pred))
func = lambda f: float(f(label, pred_label, zero_division=0))
self.f1, self.marco_f1, self.prec, self.recall = map(func, [f1_score, marco_f1_func, precision_score, recall_score])
return [self.acc, self.auc, self.f1, self.marco_f1, self.prec, self.recall]
def get_pred(self):
src = torch.cat(self.src, dim=0).long()
dst: Optional[Tensor]
if self.dst[0] is not None:
dst = torch.cat(self.dst, dim=0).long() # type: ignore
assert len(src) == len(dst), f'{len(src)} != {len(dst)}' # type: ignore
else:
dst = None
emb = torch.cat(self.emb, dim=0)
label = torch.cat(self.label, dim=0).long()
pred = torch.cat(self.pred, dim=0).float()
return src, dst, emb, label, pred
def update(self, src: Tensor, dst: Optional[Tensor], emb: Tensor, label: Tensor, pred: Tensor):
assert label.shape == pred.shape, f'Label and pred shape mismatch: {label.shape} != {pred.shape}'
assert len(src) == len(emb) == len(label) == len(pred), f'{len(src)}, {len(emb)}, {len(label)}, {len(pred)}'
idx = ~torch.isnan(pred)
src, emb, label, pred = src[idx], emb[idx], label[idx], pred[idx]
assert len(src) == len(label) == len(pred), f'{len(src)}, {len(label)}, {len(pred)}'
self.src.append(src.detach().cpu())
if dst is not None:
dst = dst[idx]
self.dst.append(dst.detach().cpu())
else:
self.dst.append(None)
self.emb.append(emb.detach().cpu())
self.label.append(label.detach().cpu())
self.pred.append(pred.detach().cpu())
class MetricsAll():
"""Metrics for train, val, and test."""
def __init__(self, name: str):
self.name = name
self.loss: List[float] = []
self.train = Metrics('train')
self.val = Metrics('val')
self.test = Metrics('test')
def update(self, src: Tensor, dst: Optional[Tensor], emb: Tensor, loss: float, label: Tensor, pred: Tensor, mask: Tensor):
assert len(src) == len(emb) == len(label) == len(pred) == len(mask), f'{len(src)} != {len(emb)} != {len(label)} != {len(pred)} != {len(mask)}'
assert src.shape == label.shape == pred.shape == mask.shape, f'{src.shape} != {label.shape} != {pred.shape} != {mask.shape}'
loss = float(loss)
if str(loss) != 'nan':
self.loss.append(loss)
def update_metrics(metrics: Metrics, mask: Tensor):
if dst is not None:
assert (len(src) == len(dst)), f'{len(src)} != {len(dst)}'
_dst = dst[mask]
else:
_dst = None
metrics.update(src[mask], _dst, emb[mask], label[mask], pred[mask])
update_metrics(self.train, mask == 0)
update_metrics(self.val, mask == 1)
update_metrics(self.test, mask == 2)
def __str__(self):
res = f'{self.name} loss:{np.mean(self.loss):.5f}'
zipped = zip(self.train.metrics, self.train.get_metrics(),
self.val.get_metrics(), self.test.get_metrics())
for name, train_metric, val_metric, test_metric in zipped:
if name in ['auc', 'macro-f1']:
res += f' {name}:{train_metric:.4f}/{val_metric:.4f}/{test_metric:.4f}'
return res
def get_label(self) -> Tuple[Tensor, Optional[Tensor], Tensor, Tensor, Tensor, Tensor]:
train_src, train_dst, train_emb, train_label, train_pred = self.train.get_pred()
val_src, val_dst, val_emb, val_label, val_pred = self.val.get_pred()
test_src, test_dst, test_emb, test_label, test_pred = self.test.get_pred()
src = torch.cat([train_src, val_src, test_src], dim=0)
dst = torch.cat([train_dst, val_dst, test_dst], dim=0) if train_dst is not None else None # type: ignore
emb = torch.cat([train_emb, val_emb, test_emb], dim=0)
label = torch.cat([train_label, val_label, test_label], dim=0)
pred = torch.cat([train_pred, val_pred, test_pred], dim=0)
mask = torch.cat([torch.zeros_like(train_src), torch.ones_like(val_src), torch.ones_like(test_src) * 2], dim=0).long()
sort = torch.argsort(src)
src, emb, label, pred, mask = map(lambda x : x[sort], [src, emb, label, pred, mask])
return src, dst, emb, label, pred, mask
def refresh(self):
self.train.get_metrics()
self.val.get_metrics()
self.test.get_metrics()
def __lt__(self, other: MetricsAll):
return self.val.f1 < other.val.f1
def __gt__(self, other: MetricsAll):
return self.val.f1 > other.val.f1
def __eq__(self, other: MetricsAll):
return self.val.f1 == other.val.f1
def align_metrics(metrics1: MetricsAll, metrics2: MetricsAll):
"""Align two metrics by src."""
src1, dst1, emb1, label1, pred1, mask1 = metrics1.get_label()
src2, dst2, emb2, label2, pred2, mask2 = metrics2.get_label()
idx = find_idx(src2, src1)
valid = idx != -1
src1, emb1, label1, pred1, mask1 = map(lambda x : x[idx][valid], [src1, emb1, label1, pred1, mask1])
src2, emb2, label2, pred2, mask2 = map(lambda x : x[valid], [src2, emb2, label2, pred2, mask2])
assert (src1==src2).sum() == len(src1), f'{(src1!=src2).sum()}/{len(src1)} not equal.'
assert (label1==label2).sum() == len(label1), f'{(label1!=label2).sum()}/{len(label1)} not equal.'
assert (mask1==mask2).sum() == len(mask1), f'{(mask1!=mask2).sum()}/{len(mask1)} not equal.'
nid, label, mask = src1, label1, mask1
return nid, label, mask, emb1, emb2, pred1, pred2
def align_pair_metrics(metrics3: MetricsAll, metrics4: MetricsAll):
src3, dst3, emb3, label3, pred3, mask3 = metrics3.get_label()
src4, dst4, emb4, label4, pred4, mask4 = metrics4.get_label()
assert dst3 is not None and dst4 is not None, 'dst3 and dst4 should not be None.'
def get_df(src: Tensor, dst: Tensor, pred: Tensor, label: Tensor):
nid = torch.cat([src, dst], dim=0).long()
pred = torch.cat([pred, pred], dim=0).float()
label = torch.cat([label, label], dim=0).long()
return pd.DataFrame({'id': nid.numpy(), 'pred': pred.numpy(), 'label': label.numpy()})
df3 = get_df(src3, dst3, pred3, label3)
df4 = get_df(src4, dst4, pred4, label4)
def group_stat(df: pd.DataFrame):
df = df.groupby('id').agg({'pred': ['mean', 'max', 'min', 'std']})
df = df.rename(columns={'mean': 'pred_mean', 'max': 'pred_max', 'min': 'pred_min', 'std': 'pred_std'})
df.columns = df.columns.droplevel(0)
df = df.reset_index()
return df
pair_pred_df3 = group_stat(df3)
pair_pred_df4 = group_stat(df4)
pair_pred_df = pair_pred_df3.merge(pair_pred_df4, on='id', how='outer', suffixes=('_3', '_4')).fillna(0.0)
return pair_pred_df
def get_metrics(labels, preds):
auc = float(roc_auc_score(labels, preds))
mf1 = float(f1_score(labels, preds > 0.5, average='macro'))
return auc, mf1
def merge_metrics(metrics1: MetricsAll, metrics2: MetricsAll, metrics3: MetricsAll, metrics4: MetricsAll, verbose: bool):
"""Merge metrics1 and metrics2."""
# Align metrics
nid, label, mask, emb1, emb2, pred1, pred2 = align_metrics(metrics1, metrics2)
res_df = pd.DataFrame({'id': nid.numpy(), 'label': label.numpy(), 'mask': mask.numpy(), 'pred1': pred1.numpy(), 'pred2': pred2.numpy()})
pair_pred_df = align_pair_metrics(metrics3, metrics4)
res_df = pd.merge(res_df, pair_pred_df, on='id', how='left').fillna(0.5)
# Convert to tensor
x = res_df[['pred1', 'pred2', 'pred_mean_3', 'pred_max_3', 'pred_std_3', 'pred_min_3', 'pred_mean_4', 'pred_max_4', 'pred_std_4', 'pred_min_4']].to_numpy()
y = res_df['label'].to_numpy()
mask = res_df['mask'].to_numpy()
# Train by xgboost
def xgb_train(x, y):
dtrain = xgb.DMatrix(x[mask==0], label=y[mask==0])
dtest = xgb.DMatrix(x[mask==2], label=y[mask==2])
param = {'max_depth': 3, 'eta': 0.1, 'objective': 'binary:logistic', 'eval_metric': 'auc'}
bst = xgb.train(param, dtrain, 20)
pred = bst.predict(dtest)
auc, f1 = get_metrics(y[mask==2], pred)
return auc, f1
auc, f1 = xgb_train(x, y)
return auc, f1