test_precision_recall_op.py 6.0 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

Y
yangyaming 已提交
15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37
import unittest
import numpy as np
from op_test import OpTest


def calc_precision(tp_count, fp_count):
    if tp_count > 0.0 or fp_count > 0.0:
        return tp_count / (tp_count + fp_count)
    return 1.0


def calc_recall(tp_count, fn_count):
    if tp_count > 0.0 or fn_count > 0.0:
        return tp_count / (tp_count + fn_count)
    return 1.0


def calc_f1_score(precision, recall):
    if precision > 0.0 or recall > 0.0:
        return 2 * precision * recall / (precision + recall)
    return 0.0


Y
yangyaming 已提交
38 39
def get_states(idxs, labels, cls_num, weights=None):
    ins_num = idxs.shape[0]
Y
yangyaming 已提交
40
    # TP FP TN FN
Y
yangyaming 已提交
41
    states = np.zeros((cls_num, 4)).astype('float32')
Y
yangyaming 已提交
42 43
    for i in xrange(ins_num):
        w = weights[i] if weights is not None else 1.0
Y
yangyaming 已提交
44 45 46 47 48
        idx = idxs[i][0]
        label = labels[i][0]
        if idx == label:
            states[idx][0] += w
            for j in xrange(cls_num):
Y
yangyaming 已提交
49
                states[j][2] += w
Y
yangyaming 已提交
50
            states[idx][2] -= w
Y
yangyaming 已提交
51
        else:
Y
yangyaming 已提交
52 53 54
            states[label][3] += w
            states[idx][1] += w
            for j in xrange(cls_num):
Y
yangyaming 已提交
55
                states[j][2] += w
Y
yangyaming 已提交
56 57
            states[label][2] -= w
            states[idx][2] -= w
Y
yangyaming 已提交
58 59 60
    return states


Y
yangyaming 已提交
61
def compute_metrics(states, cls_num):
Y
yangyaming 已提交
62 63 64 65 66
    total_tp_count = 0.0
    total_fp_count = 0.0
    total_fn_count = 0.0
    macro_avg_precision = 0.0
    macro_avg_recall = 0.0
Y
yangyaming 已提交
67
    for i in xrange(cls_num):
Y
yangyaming 已提交
68 69 70 71 72 73
        total_tp_count += states[i][0]
        total_fp_count += states[i][1]
        total_fn_count += states[i][3]
        macro_avg_precision += calc_precision(states[i][0], states[i][1])
        macro_avg_recall += calc_recall(states[i][0], states[i][3])
    metrics = []
Y
yangyaming 已提交
74 75
    macro_avg_precision /= cls_num
    macro_avg_recall /= cls_num
Y
yangyaming 已提交
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90
    metrics.append(macro_avg_precision)
    metrics.append(macro_avg_recall)
    metrics.append(calc_f1_score(macro_avg_precision, macro_avg_recall))
    micro_avg_precision = calc_precision(total_tp_count, total_fp_count)
    metrics.append(micro_avg_precision)
    micro_avg_recall = calc_recall(total_tp_count, total_fn_count)
    metrics.append(micro_avg_recall)
    metrics.append(calc_f1_score(micro_avg_precision, micro_avg_recall))
    return np.array(metrics).astype('float32')


class TestPrecisionRecallOp_0(OpTest):
    def setUp(self):
        self.op_type = "precision_recall"
        ins_num = 64
Y
yangyaming 已提交
91 92 93
        cls_num = 10
        max_probs = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32')
        idxs = np.random.choice(xrange(cls_num), ins_num).reshape(
Y
yangyaming 已提交
94
            (ins_num, 1)).astype('int32')
Y
yangyaming 已提交
95 96 97 98 99 100
        labels = np.random.choice(xrange(cls_num), ins_num).reshape(
            (ins_num, 1)).astype('int32')
        states = get_states(idxs, labels, cls_num)
        metrics = compute_metrics(states, cls_num)

        self.attrs = {'class_number': cls_num}
Y
yangyaming 已提交
101

Y
yangyaming 已提交
102
        self.inputs = {'MaxProbs': max_probs, 'Indices': idxs, 'Labels': labels}
Y
yangyaming 已提交
103 104 105 106 107 108 109 110 111 112 113 114 115 116 117

        self.outputs = {
            'BatchMetrics': metrics,
            'AccumMetrics': metrics,
            'AccumStatesInfo': states
        }

    def test_check_output(self):
        self.check_output()


class TestPrecisionRecallOp_1(OpTest):
    def setUp(self):
        self.op_type = "precision_recall"
        ins_num = 64
Y
yangyaming 已提交
118 119 120 121
        cls_num = 10
        max_probs = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32')
        idxs = np.random.choice(xrange(cls_num), ins_num).reshape(
            (ins_num, 1)).astype('int32')
Y
yangyaming 已提交
122
        weights = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32')
Y
yangyaming 已提交
123
        labels = np.random.choice(xrange(cls_num), ins_num).reshape(
Y
yangyaming 已提交
124 125
            (ins_num, 1)).astype('int32')

Y
yangyaming 已提交
126 127 128 129 130
        states = get_states(idxs, labels, cls_num, weights)
        metrics = compute_metrics(states, cls_num)

        self.attrs = {'class_number': cls_num}

Y
yangyaming 已提交
131
        self.inputs = {
Y
yangyaming 已提交
132 133
            'MaxProbs': max_probs,
            'Indices': idxs,
Y
yangyaming 已提交
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151
            'Labels': labels,
            'Weights': weights
        }

        self.outputs = {
            'BatchMetrics': metrics,
            'AccumMetrics': metrics,
            'AccumStatesInfo': states
        }

    def test_check_output(self):
        self.check_output()


class TestPrecisionRecallOp_2(OpTest):
    def setUp(self):
        self.op_type = "precision_recall"
        ins_num = 64
Y
yangyaming 已提交
152 153 154 155
        cls_num = 10
        max_probs = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32')
        idxs = np.random.choice(xrange(cls_num), ins_num).reshape(
            (ins_num, 1)).astype('int32')
Y
yangyaming 已提交
156
        weights = np.random.uniform(0, 1.0, (ins_num, 1)).astype('float32')
Y
yangyaming 已提交
157
        labels = np.random.choice(xrange(cls_num), ins_num).reshape(
Y
yangyaming 已提交
158
            (ins_num, 1)).astype('int32')
Y
yangyaming 已提交
159
        states = np.random.randint(0, 30, (cls_num, 4)).astype('float32')
Y
yangyaming 已提交
160

Y
yangyaming 已提交
161 162
        accum_states = get_states(idxs, labels, cls_num, weights)
        batch_metrics = compute_metrics(accum_states, cls_num)
Y
yangyaming 已提交
163
        accum_states += states
Y
yangyaming 已提交
164 165 166
        accum_metrics = compute_metrics(accum_states, cls_num)

        self.attrs = {'class_number': cls_num}
Y
yangyaming 已提交
167 168

        self.inputs = {
Y
yangyaming 已提交
169 170
            'MaxProbs': max_probs,
            'Indices': idxs,
Y
yangyaming 已提交
171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
            'Labels': labels,
            'Weights': weights,
            'StatesInfo': states
        }

        self.outputs = {
            'BatchMetrics': batch_metrics,
            'AccumMetrics': accum_metrics,
            'AccumStatesInfo': accum_states
        }

    def test_check_output(self):
        self.check_output()


if __name__ == '__main__':
    unittest.main()