# Copyright (c) 2024 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. # [AUTO-GENERATED] Unit test for paddle.nn.initializer.Orthogonal # 自动生成的单测，覆盖 paddle.nn.initializer.orthogonal 模块中未覆盖的代码 # Target: cover uncovered lines 157-271 in paddle/python/paddle/nn/initializer/orthogonal.py """ 测试模块：paddle.nn.initializer.Orthogonal Test Module: paddle.nn.initializer.Orthogonal 本测试覆盖以下功能： This test covers the following functions: 1. Orthogonal 初始化器在动态图模式下的使用 / Orthogonal initializer in dynamic graph mode - rows > cols 情况：列向量正交 / rows > cols case: columns are orthogonal - rows < cols 情况：行向量正交 / rows < cols case: rows are orthogonal - rows == cols 情况：行列均正交 / rows == cols case: both rows and columns are orthogonal - gain 参数测试 / gain parameter test 2. Orthogonal 初始化器在静态图模式下的使用 / Orthogonal initializer in static graph mode - 覆盖静态图代码路径 lines 157-271 / covers static graph code path 覆盖的未覆盖行：157, 161, 167, 181, 187, 193, 204, 210, 218-219, 226, 234, 240-241, 247, 254, 256, 264, 271 注意：paddle.nn.Linear(in, out) 的weight形状为 [in, out] Note: paddle.nn.Linear(in, out) weight shape is [in, out] 所以 Linear(10, 15) → weight [10, 15]，rows=10 < cols=15 """ import unittest import numpy as np import paddle class TestOrthogonalInitializerDynamic(unittest.TestCase): """测试动态图模式下Orthogonal初始化器 Test Orthogonal initializer in dynamic graph mode""" def setUp(self): """设置测试环境 / Set up test environment""" paddle.disable_static() def test_rows_greater_than_cols(self): """测试rows > cols的情况，weight [15, 10]，列向量应正交 Test rows > cols case, weight [15, 10], columns should be orthogonal Linear(10, 15) → weight shape [10, 15], rows=10 < cols=15 So we use Linear(15, 10) → weight shape [15, 10], rows=15 > cols=10""" weight_attr = paddle.ParamAttr( initializer=paddle.nn.initializer.Orthogonal() ) # Linear(out=10, in=15) → weight [15, 10], but Paddle Linear(in, out) # Linear(15, 10) → weight [15, 10], rows=15 > cols=10 linear = paddle.nn.Linear(15, 10, weight_attr=weight_attr) weight = linear.weight.numpy() self.assertEqual(weight.shape, (15, 10)) # W^T * W 应接近单位矩阵 / W^T * W should be close to identity (10x10) wtw = np.matmul(weight.T, weight) identity = np.eye(10) np.testing.assert_allclose(wtw, identity, atol=1e-5) def test_rows_less_than_cols(self): """测试rows < cols的情况，weight [10, 15]，行向量应正交 Test rows < cols case, weight [10, 15], rows should be orthogonal Linear(10, 15) → weight [10, 15], rows=10 < cols=15""" weight_attr = paddle.ParamAttr( initializer=paddle.nn.initializer.Orthogonal() ) linear = paddle.nn.Linear(10, 15, weight_attr=weight_attr) weight = linear.weight.numpy() self.assertEqual(weight.shape, (10, 15)) # W * W^T 应接近单位矩阵 / W * W^T should be close to identity (10x10) wwt = np.matmul(weight, weight.T) identity = np.eye(10) np.testing.assert_allclose(wwt, identity, atol=1e-5) def test_square_matrix(self): """测试rows == cols的情况（10x10），行列均正交 Test rows == cols case (10x10), both rows and columns are orthogonal""" weight_attr = paddle.ParamAttr( initializer=paddle.nn.initializer.Orthogonal() ) linear = paddle.nn.Linear(10, 10, weight_attr=weight_attr) weight = linear.weight.numpy() self.assertEqual(weight.shape, (10, 10)) # W * W^T 应接近单位矩阵 / W * W^T should be close to identity wwt = np.matmul(weight, weight.T) identity = np.eye(10) np.testing.assert_allclose(wwt, identity, atol=1e-5) def test_with_gain(self): """测试带gain参数的Orthogonal初始化 Test Orthogonal initialization with gain parameter""" gain = 2.0 weight_attr = paddle.ParamAttr( initializer=paddle.nn.initializer.Orthogonal(gain=gain) ) linear = paddle.nn.Linear(10, 10, weight_attr=weight_attr) weight = linear.weight.numpy() # W * W^T / gain^2 应接近单位矩阵 # W * W^T / gain^2 should be close to identity wwt = np.matmul(weight, weight.T) / (gain**2) identity = np.eye(10) np.testing.assert_allclose(wwt, identity, atol=1e-5) def test_conv2d_orthogonal(self): """测试在Conv2D上使用Orthogonal初始化（维度>2） Test Orthogonal initialization on Conv2D (dimension > 2)""" weight_attr = paddle.ParamAttr( initializer=paddle.nn.initializer.Orthogonal() ) conv = paddle.nn.Conv2D(3, 8, 3, weight_attr=weight_attr) weight = conv.weight.numpy() self.assertEqual(len(weight.shape), 4) # [out_c, in_c, kH, kW] class TestOrthogonalInitializerStatic(unittest.TestCase): """测试静态图模式下的Orthogonal初始化器，覆盖未覆盖的静态图代码路径 Test Orthogonal initializer in static graph mode, covering uncovered static paths""" def test_static_graph_rows_greater_cols(self): """测试静态图模式下rows > cols的Orthogonal初始化 Test Orthogonal init with rows > cols in static graph mode Linear(15, 8) → weight [15, 8], rows=15 > cols=8""" paddle.enable_static() try: main_prog = paddle.static.Program() startup_prog = paddle.static.Program() with paddle.static.program_guard(main_prog, startup_prog): weight_attr = paddle.ParamAttr( initializer=paddle.nn.initializer.Orthogonal() ) linear = paddle.nn.Linear(15, 8, weight_attr=weight_attr) exe = paddle.static.Executor(paddle.CPUPlace()) exe.run(startup_prog) result = exe.run(main_prog, fetch_list=[linear.weight]) weight = result[0] self.assertEqual(weight.shape, (15, 8)) # W^T * W 应接近单位矩阵 / W^T * W should be close to identity (8x8) wtw = np.matmul(weight.T, weight) identity = np.eye(8) np.testing.assert_allclose(wtw, identity, atol=1e-4) finally: paddle.disable_static() def test_static_graph_rows_less_cols(self): """测试静态图模式下rows < cols的Orthogonal初始化（触发transpose分支） Test Orthogonal init with rows < cols in static graph (triggers transpose branch) Linear(8, 15) → weight [8, 15], rows=8 < cols=15""" paddle.enable_static() try: main_prog = paddle.static.Program() startup_prog = paddle.static.Program() with paddle.static.program_guard(main_prog, startup_prog): weight_attr = paddle.ParamAttr( initializer=paddle.nn.initializer.Orthogonal() ) linear = paddle.nn.Linear(8, 15, weight_attr=weight_attr) exe = paddle.static.Executor(paddle.CPUPlace()) exe.run(startup_prog) result = exe.run(main_prog, fetch_list=[linear.weight]) weight = result[0] self.assertEqual(weight.shape, (8, 15)) # W * W^T 应接近单位矩阵 / W * W^T should be close to identity (8x8) wwt = np.matmul(weight, weight.T) identity = np.eye(8) np.testing.assert_allclose(wwt, identity, atol=1e-4) finally: paddle.disable_static() def test_static_graph_with_gain(self): """测试静态图模式下带gain参数的Orthogonal初始化 Test Orthogonal init with gain parameter in static graph mode""" paddle.enable_static() try: gain = 1.5 main_prog = paddle.static.Program() startup_prog = paddle.static.Program() with paddle.static.program_guard(main_prog, startup_prog): weight_attr = paddle.ParamAttr( initializer=paddle.nn.initializer.Orthogonal(gain=gain) ) linear = paddle.nn.Linear(6, 6, weight_attr=weight_attr) exe = paddle.static.Executor(paddle.CPUPlace()) exe.run(startup_prog) result = exe.run(main_prog, fetch_list=[linear.weight]) weight = result[0] # W * W^T / gain^2 应接近单位矩阵 # W * W^T / gain^2 should be close to identity wwt = np.matmul(weight, weight.T) / (gain**2) identity = np.eye(6) np.testing.assert_allclose(wwt, identity, atol=1e-4) finally: paddle.disable_static() if __name__ == '__main__': unittest.main()