forked from WenRichard/KBQA-BERT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
tf_metrics.py
215 lines (191 loc) · 8 KB
/
tf_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
"""
Multiclass
from:
https://github.com/guillaumegenthial/tf_metrics/blob/master/tf_metrics/__init__.py
"""
__author__ = "Guillaume Genthial"
import numpy as np
import tensorflow as tf
from tensorflow.python.ops.metrics_impl import _streaming_confusion_matrix
def precision(labels, predictions, num_classes, pos_indices=None,
weights=None, average='micro'):
"""Multi-class precision metric for Tensorflow
Parameters
----------
labels : Tensor of tf.int32 or tf.int64
The true labels
predictions : Tensor of tf.int32 or tf.int64
The predictions, same shape as labels
num_classes : int
The number of classes
pos_indices : list of int, optional
The indices of the positive classes, default is all
weights : Tensor of tf.int32, optional
Mask, must be of compatible shape with labels
average : str, optional
'micro': counts the total number of true positives, false
positives, and false negatives for the classes in
`pos_indices` and infer the metric from it.
'macro': will compute the metric separately for each class in
`pos_indices` and average. Will not account for class
imbalance.
'weighted': will compute the metric separately for each class in
`pos_indices` and perform a weighted average by the total
number of true labels for each class.
Returns
-------
tuple of (scalar float Tensor, update_op)
"""
cm, op = _streaming_confusion_matrix(
labels, predictions, num_classes, weights)
pr, _, _ = metrics_from_confusion_matrix(
cm, pos_indices, average=average)
op, _, _ = metrics_from_confusion_matrix(
op, pos_indices, average=average)
return (pr, op)
def recall(labels, predictions, num_classes, pos_indices=None, weights=None,
average='micro'):
"""Multi-class recall metric for Tensorflow
Parameters
----------
labels : Tensor of tf.int32 or tf.int64
The true labels
predictions : Tensor of tf.int32 or tf.int64
The predictions, same shape as labels
num_classes : int
The number of classes
pos_indices : list of int, optional
The indices of the positive classes, default is all
weights : Tensor of tf.int32, optional
Mask, must be of compatible shape with labels
average : str, optional
'micro': counts the total number of true positives, false
positives, and false negatives for the classes in
`pos_indices` and infer the metric from it.
'macro': will compute the metric separately for each class in
`pos_indices` and average. Will not account for class
imbalance.
'weighted': will compute the metric separately for each class in
`pos_indices` and perform a weighted average by the total
number of true labels for each class.
Returns
-------
tuple of (scalar float Tensor, update_op)
"""
cm, op = _streaming_confusion_matrix(
labels, predictions, num_classes, weights)
_, re, _ = metrics_from_confusion_matrix(
cm, pos_indices, average=average)
_, op, _ = metrics_from_confusion_matrix(
op, pos_indices, average=average)
return (re, op)
def f1(labels, predictions, num_classes, pos_indices=None, weights=None,
average='micro'):
return fbeta(labels, predictions, num_classes, pos_indices, weights,
average)
def fbeta(labels, predictions, num_classes, pos_indices=None, weights=None,
average='micro', beta=1):
"""Multi-class fbeta metric for Tensorflow
Parameters
----------
labels : Tensor of tf.int32 or tf.int64
The true labels
predictions : Tensor of tf.int32 or tf.int64
The predictions, same shape as labels
num_classes : int
The number of classes
pos_indices : list of int, optional
The indices of the positive classes, default is all
weights : Tensor of tf.int32, optional
Mask, must be of compatible shape with labels
average : str, optional
'micro': counts the total number of true positives, false
positives, and false negatives for the classes in
`pos_indices` and infer the metric from it.
'macro': will compute the metric separately for each class in
`pos_indices` and average. Will not account for class
imbalance.
'weighted': will compute the metric separately for each class in
`pos_indices` and perform a weighted average by the total
number of true labels for each class.
beta : int, optional
Weight of precision in harmonic mean
Returns
-------
tuple of (scalar float Tensor, update_op)
"""
cm, op = _streaming_confusion_matrix(
labels, predictions, num_classes, weights)
_, _, fbeta = metrics_from_confusion_matrix(
cm, pos_indices, average=average, beta=beta)
_, _, op = metrics_from_confusion_matrix(
op, pos_indices, average=average, beta=beta)
return (fbeta, op)
def safe_div(numerator, denominator):
"""Safe division, return 0 if denominator is 0"""
numerator, denominator = tf.to_float(numerator), tf.to_float(denominator)
zeros = tf.zeros_like(numerator, dtype=numerator.dtype)
denominator_is_zero = tf.equal(denominator, zeros)
return tf.where(denominator_is_zero, zeros, numerator / denominator)
def pr_re_fbeta(cm, pos_indices, beta=1):
"""Uses a confusion matrix to compute precision, recall and fbeta"""
num_classes = cm.shape[0]
neg_indices = [i for i in range(num_classes) if i not in pos_indices]
cm_mask = np.ones([num_classes, num_classes])
cm_mask[neg_indices, neg_indices] = 0
diag_sum = tf.reduce_sum(tf.diag_part(cm * cm_mask))
cm_mask = np.ones([num_classes, num_classes])
cm_mask[:, neg_indices] = 0
tot_pred = tf.reduce_sum(cm * cm_mask)
cm_mask = np.ones([num_classes, num_classes])
cm_mask[neg_indices, :] = 0
tot_gold = tf.reduce_sum(cm * cm_mask)
pr = safe_div(diag_sum, tot_pred)
re = safe_div(diag_sum, tot_gold)
fbeta = safe_div((1. + beta**2) * pr * re, beta**2 * pr + re)
return pr, re, fbeta
def metrics_from_confusion_matrix(cm, pos_indices=None, average='micro',
beta=1):
"""Precision, Recall and F1 from the confusion matrix
Parameters
----------
cm : tf.Tensor of type tf.int32, of shape (num_classes, num_classes)
The streaming confusion matrix.
pos_indices : list of int, optional
The indices of the positive classes
beta : int, optional
Weight of precision in harmonic mean
average : str, optional
'micro', 'macro' or 'weighted'
"""
num_classes = cm.shape[0]
if pos_indices is None:
pos_indices = [i for i in range(num_classes)]
if average == 'micro':
return pr_re_fbeta(cm, pos_indices, beta)
elif average in {'macro', 'weighted'}:
precisions, recalls, fbetas, n_golds = [], [], [], []
for idx in pos_indices:
pr, re, fbeta = pr_re_fbeta(cm, [idx], beta)
precisions.append(pr)
recalls.append(re)
fbetas.append(fbeta)
cm_mask = np.zeros([num_classes, num_classes])
cm_mask[idx, :] = 1
n_golds.append(tf.to_float(tf.reduce_sum(cm * cm_mask)))
if average == 'macro':
pr = tf.reduce_mean(precisions)
re = tf.reduce_mean(recalls)
fbeta = tf.reduce_mean(fbetas)
return pr, re, fbeta
if average == 'weighted':
n_gold = tf.reduce_sum(n_golds)
pr_sum = sum(p * n for p, n in zip(precisions, n_golds))
pr = safe_div(pr_sum, n_gold)
re_sum = sum(r * n for r, n in zip(recalls, n_golds))
re = safe_div(re_sum, n_gold)
fbeta_sum = sum(f * n for f, n in zip(fbetas, n_golds))
fbeta = safe_div(fbeta_sum, n_gold)
return pr, re, fbeta
else:
raise NotImplementedError()