# 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. """ 动态图自动微分单元测试 / Dynamic Graph Autograd Unit Tests 测试目标 / Test Target: paddle动态图自动微分功能 (paddle.grad, paddle.jacobian, paddle.hessian) 覆盖的模块 / Covered Modules: - paddle.grad: 梯度计算 - paddle.jacobian: Jacobian矩阵计算 - paddle.hessian: Hessian矩阵计算 - Tensor.backward: 反向传播 - Tensor.gradient: 获取梯度 作用 / Purpose: 覆盖自动微分机制的各种代码路径，包括高阶导数、梯度累积等功能。 """ import unittest import numpy as np import paddle paddle.disable_static() class TestGradBasic(unittest.TestCase): """测试基本梯度计算 / Test basic gradient computation""" def test_grad_simple(self): """测试简单梯度计算 / Test simple gradient computation""" x = paddle.to_tensor([1.0, 2.0, 3.0], stop_gradient=False) y = x * x grad = paddle.grad(y, x, grad_outputs=paddle.ones_like(y)) expected = np.array([2.0, 4.0, 6.0]) np.testing.assert_allclose(grad[0].numpy(), expected, rtol=1e-5) def test_grad_with_create_graph(self): """测试创建计算图的梯度 / Test gradient with create_graph""" x = paddle.to_tensor([1.0, 2.0, 3.0], stop_gradient=False) y = x**3 grad = paddle.grad( y, x, grad_outputs=paddle.ones_like(y), create_graph=True ) # dy/dx = 3x^2 = [3, 12, 27] expected = np.array([3.0, 12.0, 27.0]) np.testing.assert_allclose(grad[0].numpy(), expected, rtol=1e-5) def test_grad_sum(self): """测试求和函数的梯度 / Test gradient of sum""" x = paddle.randn([3, 4]) x.stop_gradient = False y = x.sum() y.backward() np.testing.assert_allclose( x.grad.numpy(), np.ones((3, 4), dtype='float32'), rtol=1e-5 ) def test_grad_chain_rule(self): """测试链式法则 / Test chain rule""" x = paddle.to_tensor([2.0], stop_gradient=False) y = x * x # y = x^2 z = y * y # z = x^4 dz_dx = paddle.grad(z, x) # dz/dx = 4x^3 = 4*8 = 32 self.assertAlmostEqual(float(dz_dx[0].item()), 32.0, places=4) def test_retain_graph(self): """测试保留计算图 / Test retain_graph""" x = paddle.to_tensor([2.0], stop_gradient=False) y = x * x grad1 = paddle.grad(y, x, retain_graph=True) grad2 = paddle.grad(y, x, retain_graph=False) np.testing.assert_allclose(grad1[0].numpy(), grad2[0].numpy()) class TestGradMultiOutput(unittest.TestCase): """测试多输出梯度 / Test multi-output gradient""" def test_grad_multiple_outputs(self): """测试多输出梯度计算 / Test gradient with multiple outputs""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) y1 = x * x y2 = x * 2 grad = paddle.grad( [y1, y2], x, grad_outputs=[paddle.ones([2]), paddle.ones([2])] ) # dy1/dx = 2x, dy2/dx = 2 expected = np.array([4.0, 6.0]) np.testing.assert_allclose(grad[0].numpy(), expected, rtol=1e-5) def test_grad_multiple_inputs(self): """测试多输入梯度计算 / Test gradient with multiple inputs""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) w = paddle.to_tensor([3.0, 4.0], stop_gradient=False) y = (x * w).sum() grad_x, grad_w = paddle.grad(y, [x, w]) np.testing.assert_allclose(grad_x.numpy(), w.numpy(), rtol=1e-5) np.testing.assert_allclose(grad_w.numpy(), x.numpy(), rtol=1e-5) class TestBackwardBasic(unittest.TestCase): """测试backward方法 / Test backward method""" def test_backward_simple(self): """测试简单backward / Test simple backward""" x = paddle.to_tensor([1.0, 2.0, 3.0], stop_gradient=False) y = (x * x).sum() y.backward() expected = np.array([2.0, 4.0, 6.0]) np.testing.assert_allclose(x.grad.numpy(), expected, rtol=1e-5) def test_backward_accumulate(self): """测试梯度累积 / Test gradient accumulation""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) y = (x * x).sum() y.backward() first_grad = x.grad.numpy().copy() y = (x * x).sum() y.backward() # Gradient should be accumulated second_grad = x.grad.numpy() np.testing.assert_allclose(second_grad, first_grad * 2, rtol=1e-5) def test_clear_grad(self): """测试清除梯度 / Test clear gradient""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) y = (x * x).sum() y.backward() x.clear_gradient() self.assertTrue(x.grad is None or np.all(x.grad.numpy() == 0)) def test_stop_gradient(self): """测试停止梯度 / Test stop_gradient""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=True) y = x * 2 self.assertTrue(y.stop_gradient) def test_no_grad_decorator(self): """测试no_grad装饰器 / Test no_grad decorator""" @paddle.no_grad() def func(x): return x * x x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) y = func(x) self.assertTrue(y.stop_gradient) class TestJacobian(unittest.TestCase): """测试Jacobian矩阵计算 / Test Jacobian matrix computation""" def test_jacobian_via_grad(self): """通过grad计算Jacobian / Compute Jacobian via grad""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) # f(x) = x^2 element-wise, df/dx = diag(2x) y = x**2 # Compute row by row jac_rows = [] for i in range(y.shape[0]): if x.grad is not None: x.clear_gradient() grad = paddle.grad(y[i], x, retain_graph=True) jac_rows.append(grad[0].numpy()) # Diagonal should be [2*1, 2*2] = [2, 4] self.assertAlmostEqual(jac_rows[0][0], 2.0, places=4) self.assertAlmostEqual(jac_rows[1][1], 4.0, places=4) def test_jacobian_single_output(self): """测试单输出的梯度 / Test gradient with single output""" x = paddle.randn([4]) x.stop_gradient = False y = x.sum() y.backward() np.testing.assert_allclose(x.grad.numpy(), np.ones(4, dtype='float32')) class TestHessian(unittest.TestCase): """测试二阶导数计算 / Test second-order derivative computation""" def test_second_order_gradient(self): """测试二阶梯度计算 / Test second-order gradient computation""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) # f(x) = x^3, f'(x) = 3x^2, f''(x) = 6x y = (x**3).sum() first_grad = paddle.grad(y, x, create_graph=True) second_grad = paddle.grad(first_grad[0].sum(), x) # f''(x) = 6x = [6, 12] expected = np.array([6.0, 12.0]) np.testing.assert_allclose(second_grad[0].numpy(), expected, rtol=1e-4) class TestGradientContext(unittest.TestCase): """测试梯度上下文管理 / Test gradient context management""" def test_no_grad_context(self): """测试no_grad上下文 / Test no_grad context""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) with paddle.no_grad(): y = x * x self.assertTrue(y.stop_gradient) def test_enable_grad_context(self): """测试enable_grad上下文 / Test enable_grad context""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) with paddle.no_grad(), paddle.enable_grad(): y = x * x self.assertFalse(y.stop_gradient) def test_set_grad_enabled(self): """测试set_grad_enabled / Test set_grad_enabled""" x = paddle.to_tensor([1.0, 2.0], stop_gradient=False) with paddle.set_grad_enabled(False): y = x * x self.assertTrue(y.stop_gradient) with paddle.set_grad_enabled(True): z = x * x self.assertFalse(z.stop_gradient) if __name__ == '__main__': unittest.main()