test_sigmoid_focal_loss.py 5.9 KB
Newer Older
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
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# 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.

import paddle
import paddle.fluid as fluid
import numpy as np
import unittest
from op_test import OpTest
from test_sigmoid_focal_loss_op import sigmoid_focal_loss_forward


def call_sfl_functional(logit,
                        label,
                        normalizer,
                        alpha=0.25,
                        gamma=2.0,
                        reduction='sum'):
    res = paddle.nn.functional.sigmoid_focal_loss(
        logit, label, normalizer, alpha=alpha, gamma=gamma, reduction=reduction)
    return res


def test_static(place,
                logit_np,
                label_np,
                normalizer_np,
                alpha=0.25,
                gamma=2.0,
                reduction='sum'):
    paddle.enable_static()
    prog = paddle.static.Program()
    startup_prog = paddle.static.Program()
    with paddle.static.program_guard(prog, startup_prog):
45 46
        logit = paddle.fluid.data(name='logit', shape=logit_np.shape, dtype='float64')
        label = paddle.fluid.data(name='label', shape=label_np.shape, dtype='float64')
47 48 49 50
        feed_dict = {"logit": logit_np, "label": label_np}

        normalizer = None
        if normalizer_np is not None:
51
            normalizer = paddle.fluid.data(
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
                name='normalizer', shape=normalizer_np.shape, dtype='float64')
            feed_dict["normalizer"] = normalizer_np

        res = call_sfl_functional(logit, label, normalizer, alpha, gamma,
                                  reduction)
        exe = paddle.static.Executor(place)
        static_result = exe.run(prog, feed=feed_dict, fetch_list=[res])
    return static_result


def test_dygraph(place,
                 logit_np,
                 label_np,
                 normalizer_np,
                 alpha=0.25,
                 gamma=2.0,
                 reduction='sum'):
    paddle.disable_static()
    logit = paddle.to_tensor(logit_np)
    label = paddle.to_tensor(label_np)
    normalizer = None
    if normalizer_np is not None:
        normalizer = paddle.to_tensor(normalizer_np)
    dy_res = call_sfl_functional(logit, label, normalizer, alpha, gamma,
                                 reduction)
    dy_result = dy_res.numpy()
    paddle.enable_static()
    return dy_result


def calc_sigmoid_focal_loss(logit_np,
                            label_np,
                            normalizer_np,
                            alpha=0.25,
                            gamma=2.0,
                            reduction='sum'):

    loss = np.maximum(
        logit_np,
        0) - logit_np * label_np + np.log(1 + np.exp(-np.abs(logit_np)))

    pred = 1 / (1 + np.exp(-logit_np))
    p_t = pred * label_np + (1 - pred) * (1 - label_np)

    if alpha is not None:
        alpha_t = alpha * label_np + (1 - alpha) * (1 - label_np)
        loss = alpha_t * loss

    if gamma is not None:
        loss = loss * ((1 - p_t)**gamma)

    if normalizer_np is not None:
        loss = loss / normalizer_np

    if reduction == 'mean':
        loss = np.mean(loss)
    elif reduction == 'sum':
        loss = np.sum(loss)

    return loss


class TestSigmoidFocalLoss(unittest.TestCase):
    def test_SigmoidFocalLoss(self):
        logit_np = np.random.uniform(
            0.1, 0.8, size=(2, 3, 4, 10)).astype(np.float64)
        label_np = np.random.randint(
            0, 2, size=(2, 3, 4, 10)).astype(np.float64)
        normalizer_nps = [
            np.asarray(
                [np.sum(label_np > 0)], dtype=label_np.dtype), None
        ]
        places = [fluid.CPUPlace()]
        if fluid.core.is_compiled_with_cuda():
            places.append(fluid.CUDAPlace(0))
        reductions = ['sum', 'mean', 'none']
        alphas = [0.25, 0.5]
        gammas = [3, 0.]
        for place in places:
            for reduction in reductions:
                for alpha in alphas:
                    for gamma in gammas:
                        for normalizer_np in normalizer_nps:
                            static_result = test_static(place, logit_np,
                                                        label_np, normalizer_np,
                                                        alpha, gamma, reduction)
                            dy_result = test_dygraph(place, logit_np, label_np,
                                                     normalizer_np, alpha,
                                                     gamma, reduction)
                            expected = calc_sigmoid_focal_loss(
                                logit_np, label_np, normalizer_np, alpha, gamma,
                                reduction)
                            self.assertTrue(
                                np.allclose(static_result, expected))
                            self.assertTrue(
                                np.allclose(static_result, dy_result))
                            self.assertTrue(np.allclose(dy_result, expected))

    def test_SigmoidFocalLoss_error(self):
        paddle.disable_static()
        logit = paddle.to_tensor([[0.97], [0.91], [0.03]], dtype='float32')
        label = paddle.to_tensor([[1.0], [1.0], [0.0]], dtype='float32')
        self.assertRaises(
            ValueError,
            paddle.nn.functional.sigmoid_focal_loss,
            logit=logit,
            label=label,
            normalizer=None,
            reduction="unsupport reduction")
        paddle.enable_static()


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