# Copyright (c) 2026 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. """ Dropout层单元测试 / Dropout Layer Unit Tests 测试目标 / Test Target: paddle.nn.Dropout系列层 (覆盖率较高但部分边界情况未测) 覆盖的模块 / Covered Modules: - paddle.nn.Dropout: 1D Dropout - paddle.nn.Dropout2D: 2D通道Dropout - paddle.nn.Dropout3D: 3D通道Dropout - paddle.nn.AlphaDropout: Alpha Dropout 作用 / Purpose: 覆盖Dropout层在训练和评估模式下的代码路径，测试各种参数设置和边界情况。 """ import unittest import numpy as np import paddle from paddle import nn paddle.disable_static() class TestDropout(unittest.TestCase): """测试Dropout / Test Dropout""" def test_dropout_eval_mode(self): """测试评估模式下Dropout（无丢弃）/ Test Dropout in eval mode (no dropout)""" dropout = nn.Dropout(p=0.5) dropout.eval() x = paddle.randn([100, 100]) y = dropout(x) np.testing.assert_allclose(x.numpy(), y.numpy()) def test_dropout_train_mode(self): """测试训练模式下Dropout / Test Dropout in train mode""" dropout = nn.Dropout(p=0.5) dropout.train() x = paddle.ones([1000]) y = dropout(x) # Some values should be zero n_zeros = int(paddle.sum(y == 0).numpy()) self.assertTrue(n_zeros > 0) def test_dropout_p0(self): """测试p=0的Dropout（无丢弃）/ Test Dropout with p=0 (no dropout)""" dropout = nn.Dropout(p=0.0) x = paddle.randn([4, 10]) y = dropout(x) np.testing.assert_allclose(x.numpy(), y.numpy()) def test_dropout_p1(self): """测试p=1的Dropout（全丢弃）/ Test Dropout with p=1 (all zero)""" dropout = nn.Dropout(p=1.0) x = paddle.ones([100]) y = dropout(x) # All values should be zero in train mode (default) self.assertAlmostEqual(float(y.sum().numpy()), 0.0, places=5) def test_dropout_mode_upscale(self): """测试upscale模式 / Test upscale mode""" dropout = nn.Dropout(p=0.5, mode='upscale_in_train') x = paddle.ones([1000]) y = dropout(x) # Non-zero values should be scaled by 1/(1-p) non_zero = y[y != 0] if len(non_zero) > 0: self.assertAlmostEqual(float(non_zero[0].numpy()), 2.0, places=5) def test_dropout_shape_preserved(self): """测试Dropout保留形状 / Test Dropout preserves shape""" dropout = nn.Dropout(p=0.3) x = paddle.randn([4, 6, 8]) y = dropout(x) self.assertEqual(y.shape, [4, 6, 8]) def test_dropout_gradient(self): """测试Dropout梯度 / Test Dropout gradient""" dropout = nn.Dropout(p=0.0) # p=0: identity x = paddle.randn([4, 10]) x.stop_gradient = False y = dropout(x) y.sum().backward() self.assertIsNotNone(x.grad) class TestDropout2D(unittest.TestCase): """测试Dropout2D / Test Dropout2D""" def test_dropout2d_basic(self): """测试基本Dropout2D / Test basic Dropout2D""" dropout2d = nn.Dropout2D(p=0.5) x = paddle.randn([4, 8, 16, 16]) y = dropout2d(x) self.assertEqual(y.shape, [4, 8, 16, 16]) def test_dropout2d_channel_drop(self): """测试Dropout2D通道丢弃 / Test Dropout2D channel dropping""" dropout2d = nn.Dropout2D(p=0.5) dropout2d.train() x = paddle.ones([4, 100, 8, 8]) y = dropout2d(x) # Some entire channels should be zero # Check that some channels are zero channel_sums = y.mean(axis=[0, 2, 3]) n_zero_channels = int(paddle.sum(channel_sums == 0).numpy()) self.assertTrue(n_zero_channels > 0) def test_dropout2d_eval_mode(self): """测试评估模式Dropout2D / Test Dropout2D eval mode""" dropout2d = nn.Dropout2D(p=0.5) dropout2d.eval() x = paddle.randn([4, 8, 16, 16]) y = dropout2d(x) np.testing.assert_allclose(x.numpy(), y.numpy()) def test_dropout2d_nchw(self): """测试NCHW格式的Dropout2D / Test NCHW format Dropout2D""" dropout2d = nn.Dropout2D(p=0.3, data_format='NCHW') x = paddle.randn([4, 8, 16, 16]) y = dropout2d(x) self.assertEqual(y.shape, [4, 8, 16, 16]) class TestDropout3D(unittest.TestCase): """测试Dropout3D / Test Dropout3D""" def test_dropout3d_basic(self): """测试基本Dropout3D / Test basic Dropout3D""" dropout3d = nn.Dropout3D(p=0.5) x = paddle.randn([2, 8, 4, 8, 8]) y = dropout3d(x) self.assertEqual(y.shape, [2, 8, 4, 8, 8]) def test_dropout3d_eval_mode(self): """测试评估模式Dropout3D / Test Dropout3D eval mode""" dropout3d = nn.Dropout3D(p=0.5) dropout3d.eval() x = paddle.randn([2, 8, 4, 8, 8]) y = dropout3d(x) np.testing.assert_allclose(x.numpy(), y.numpy()) def test_dropout3d_channel_wise(self): """测试Dropout3D通道维度 / Test Dropout3D channel-wise""" dropout3d = nn.Dropout3D(p=0.5) dropout3d.train() x = paddle.ones([2, 100, 4, 8, 8]) y = dropout3d(x) self.assertEqual(y.shape, [2, 100, 4, 8, 8]) class TestAlphaDropout(unittest.TestCase): """测试AlphaDropout / Test AlphaDropout""" def test_alpha_dropout_basic(self): """测试基本AlphaDropout / Test basic AlphaDropout""" alpha_dropout = nn.AlphaDropout(p=0.5) x = paddle.randn([100, 100]) y = alpha_dropout(x) self.assertEqual(y.shape, [100, 100]) def test_alpha_dropout_eval_mode(self): """测试评估模式AlphaDropout / Test AlphaDropout eval mode""" alpha_dropout = nn.AlphaDropout(p=0.5) alpha_dropout.eval() x = paddle.randn([100, 100]) y = alpha_dropout(x) np.testing.assert_allclose(x.numpy(), y.numpy()) def test_alpha_dropout_selu_combination(self): """测试AlphaDropout与SELU组合 / Test AlphaDropout with SELU combination""" # AlphaDropout is designed to work with SELU model = nn.Sequential( nn.Linear(10, 10), nn.SELU(), nn.AlphaDropout(p=0.1), nn.Linear(10, 5), ) model.train() x = paddle.randn([4, 10]) y = model(x) self.assertEqual(y.shape, [4, 5]) class TestDropoutInModel(unittest.TestCase): """测试Dropout在模型中的使用 / Test Dropout in model""" def test_dropout_in_sequential(self): """测试Dropout在Sequential中 / Test Dropout in Sequential""" model = nn.Sequential( nn.Linear(10, 20), nn.ReLU(), nn.Dropout(p=0.5), nn.Linear(20, 5) ) x = paddle.randn([4, 10]) model.train() y_train = model(x) model.eval() y_eval = model(x) # Both should have same shape self.assertEqual(y_train.shape, [4, 5]) self.assertEqual(y_eval.shape, [4, 5]) def test_dropout_train_eval_difference(self): """测试训练和评估模式的不同 / Test difference between train and eval""" dropout = nn.Dropout(p=0.5) x = paddle.ones([1000]) # Multiple evaluations with eval mode should give same result dropout.eval() y1 = dropout(x) y2 = dropout(x) np.testing.assert_allclose(y1.numpy(), y2.numpy()) # In train mode, multiple runs may differ dropout.train() y3 = dropout(x) # y3 should have some zeros due to dropout self.assertTrue( float(y3.sum().numpy()) <= 2000 ) # max is 2000 due to upscaling if __name__ == '__main__': unittest.main()