|
| 1 | +# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, |
| 10 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 11 | +# See the License for the specific language governing permissions and |
| 12 | +# limitations under the License. |
| 13 | + |
| 14 | +import sys |
| 15 | + |
| 16 | +sys.path.append("..") |
| 17 | +import paddle |
| 18 | +import unittest |
| 19 | +import numpy as np |
| 20 | +from paddle.nn.functional import kl_div |
| 21 | +from op_test_xpu import XPUOpTest |
| 22 | +from xpu.get_test_cover_info import ( |
| 23 | + create_test_class, |
| 24 | + get_xpu_op_support_types, |
| 25 | + XPUOpTestWrapper, |
| 26 | +) |
| 27 | + |
| 28 | +paddle.enable_static() |
| 29 | + |
| 30 | + |
| 31 | +def kldiv_loss(x, target, reduction): |
| 32 | + output = target * (np.log(target) - x) |
| 33 | + loss = np.where(target >= 0, output, np.zeros_like(x)) |
| 34 | + |
| 35 | + if reduction == "batchmean": |
| 36 | + if len(x.shape) > 0: |
| 37 | + return loss.sum() / x.shape[0] |
| 38 | + else: |
| 39 | + return loss.sum() |
| 40 | + if reduction == "mean": |
| 41 | + return loss.mean() |
| 42 | + if reduction == "sum": |
| 43 | + return loss.sum() |
| 44 | + |
| 45 | + return loss |
| 46 | + |
| 47 | + |
| 48 | +class XPUTestKLDivLossOp(XPUOpTestWrapper): |
| 49 | + def __init__(self): |
| 50 | + self.op_name = 'kldiv_loss' |
| 51 | + self.use_dynamic_create_class = False |
| 52 | + |
| 53 | + class TestKLDivLossOp(XPUOpTest): |
| 54 | + def setUp(self): |
| 55 | + self.initTestCase() |
| 56 | + self.op_type = 'kldiv_loss' |
| 57 | + self.dtype = np.float32 |
| 58 | + self.__class__.use_xpu = True |
| 59 | + self.python_api = kl_div |
| 60 | + x = np.random.uniform(-10, 10, self.x_shape).astype('float32') |
| 61 | + target = np.random.uniform(-10, 10, self.x_shape).astype('float32') |
| 62 | + |
| 63 | + self.attrs = {"reduction": self.reduction} |
| 64 | + |
| 65 | + self.inputs = { |
| 66 | + 'X': x, |
| 67 | + 'Target': target, |
| 68 | + } |
| 69 | + loss = kldiv_loss(x, target, self.reduction) |
| 70 | + self.outputs = {'Loss': loss.astype('float32')} |
| 71 | + |
| 72 | + def test_check_output(self): |
| 73 | + self.check_output(check_eager=True) |
| 74 | + |
| 75 | + def test_check_grad(self): |
| 76 | + self.check_grad_with_place( |
| 77 | + paddle.XPUPlace(0), |
| 78 | + ['X'], |
| 79 | + 'Loss', |
| 80 | + no_grad_set=set(["Target"]), |
| 81 | + check_eager=True, |
| 82 | + ) |
| 83 | + |
| 84 | + def initTestCase(self): |
| 85 | + self.x_shape = (4, 5, 5) |
| 86 | + self.reduction = 'none' |
| 87 | + |
| 88 | + class TestKLDivLossOp2(TestKLDivLossOp): |
| 89 | + def initTestCase(self): |
| 90 | + self.x_shape = (3, 2, 7, 7) |
| 91 | + self.reduction = 'none' |
| 92 | + |
| 93 | + class TestKLDivLossOp3(TestKLDivLossOp): |
| 94 | + def initTestCase(self): |
| 95 | + self.x_shape = (2, 3, 5, 7, 9) |
| 96 | + self.reduction = 'none' |
| 97 | + |
| 98 | + class TestKLDivLossOp4(TestKLDivLossOp): |
| 99 | + def initTestCase(self): |
| 100 | + self.x_shape = (5, 20) |
| 101 | + self.reduction = 'none' |
| 102 | + |
| 103 | + class TestKLDivLossDygraph(unittest.TestCase): |
| 104 | + def run_kl_loss(self, reduction, shape=(5, 20)): |
| 105 | + x = np.random.uniform(-10, 10, shape).astype('float32') |
| 106 | + target = np.random.uniform(-10, 10, shape).astype('float32') |
| 107 | + gt_loss = kldiv_loss(x, target, reduction) |
| 108 | + |
| 109 | + with paddle.fluid.dygraph.guard(): |
| 110 | + kldiv_criterion = paddle.nn.KLDivLoss(reduction) |
| 111 | + pred_loss = kldiv_criterion( |
| 112 | + paddle.to_tensor(x), paddle.to_tensor(target) |
| 113 | + ) |
| 114 | + np.testing.assert_allclose( |
| 115 | + pred_loss.numpy(), gt_loss, rtol=1e-05 |
| 116 | + ) |
| 117 | + |
| 118 | + def test_kl_loss_none(self): |
| 119 | + self.run_kl_loss('none') |
| 120 | + |
| 121 | + def test_kl_loss_static_api(self): |
| 122 | + input = paddle.fluid.data(name='input', shape=[5, 20]) |
| 123 | + label = paddle.fluid.data(name='label', shape=[5, 20]) |
| 124 | + |
| 125 | + paddle.nn.functional.kl_div(input, label) |
| 126 | + |
| 127 | + class TestKLDivLossTypePromotion(unittest.TestCase): |
| 128 | + def test_kl_div_promotion(self): |
| 129 | + |
| 130 | + with paddle.fluid.dygraph.guard(): |
| 131 | + x1 = paddle.rand([5, 20], dtype='float32') |
| 132 | + target1 = paddle.rand([5, 20], dtype='float32') |
| 133 | + |
| 134 | + kldiv_criterion = paddle.nn.KLDivLoss() |
| 135 | + pred_loss1 = kldiv_criterion(x1, target1) |
| 136 | + |
| 137 | + x2 = paddle.rand([5, 20], dtype='float32') |
| 138 | + target2 = paddle.rand([5, 20], dtype='float32') |
| 139 | + pred_loss2 = paddle.nn.functional.kl_div(x2, target2) |
| 140 | + |
| 141 | + |
| 142 | +support_types = get_xpu_op_support_types('kldiv_loss') |
| 143 | +for stype in support_types: |
| 144 | + create_test_class(globals(), XPUTestKLDivLossOp, stype) |
| 145 | + |
| 146 | +if __name__ == "__main__": |
| 147 | + unittest.main() |
0 commit comments