Add CPU and GPU eigh op implementation #34990
Conversation
|
Thanks for your contribution! |
| @@ -946,53 +946,42 @@ def __check_input(x, vec): | |||
| def matrix_power(x, n, name=None): | |||
| r""" | |||
| Computes the n-th power of a square matrix or a batch of square matrices. | |||
|
|
|||
paddle/fluid/operators/eigh_helper.h
Outdated
| op_inputs[name] = name_vector; | ||
| } | ||
| auto op = | ||
| framework::OpRegistry::CreateOp(type, op_inputs, op_outputs, attrs); |
There was a problem hiding this comment.
这是需要在eigh算子里面创建别的op？不建议使用这种方式。
| self.x_np = np.random.random(self.x_shape).astype(self.x_type) | ||
|
|
||
| def test_check_output(self): | ||
| self.check_output(no_check_set=['Eigenvectors']) |
There was a problem hiding this comment.
这里对于去掉'Eigenvectors'解释：
1、特征值是唯一的，但是特征向量不唯一；
2、numpy采用Lapack库，paddle目前采用eigen库
3、eigen计算出来的特征向量与lapack会有一个符号上有区别，比如：
eigen计算出来特征向量 [[2, 3],[4, 5]]， Lapack计算出来的为[[-2, 3],[-4, 5]]，第一列符号不同，但是结果都是争取的。
There was a problem hiding this comment.
需要解释一下为什么会出现符号上的差别。以及说明下在其他单测里面会检查Eigenvectors。
| std::vector<int64_t> v_dim = {input_dim[1]}; | ||
| if (rank > 2) { | ||
| v_dim = {batch_size, input_dim[1]}; | ||
| } |
| OP_INOUT_CHECK(ctx->HasInput("Eigenvectors"), "Input", "Eigenvectors", | ||
| "EighGrad"); | ||
| OP_INOUT_CHECK(ctx->HasInputs(framework::GradVarName("Eigenvalues")), | ||
| "Input", "Eigenvalues@GRAD", "EighGrad"); |
There was a problem hiding this comment.
那这里应该使用ctx->HasInput而不是ctx->HasInputs，没有s？
| template <typename ValueType, typename T> | ||
| class EighGPUKernel : public framework::OpKernel<T> { | ||
| public: | ||
| void Compute(const framework::ExecutionContext &ctx) const override { |
There was a problem hiding this comment.
实现和EighKernel一样，GPU也可以直接使用EighKernel来注册，没有必要实现EighGPUKernel。
| void Compute(const framework::ExecutionContext& ctx) const override { | ||
| auto& x_grad = *ctx.Output<framework::Tensor>(framework::GradVarName("X")); | ||
| x_grad.mutable_data<T>(ctx.GetPlace()); | ||
| auto& output_w_var = *ctx.Input<Tensor>("Eigenvalues"); |
| using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>; | ||
| template <typename T, int MajorType = Eigen::RowMajor, | ||
| typename IndexType = Eigen::DenseIndex> | ||
| using EigenVector = framework::EigenVector<T, MajorType, IndexType>; |
|
|
||
| } // namespace math | ||
| } // namespace operators | ||
| } // namespace paddle |
| // symmetric matrices, and uses the variable compute_vectors to | ||
| // control whether to return the eigenvectors. | ||
| template <typename DeviceContext, typename ValueType, typename T> | ||
| struct MatrixEighFunctorCPU { |
There was a problem hiding this comment.
CPU、GPU的Functor都叫MatrixEighFunctor，针对CPUDeviceContext、CUDADeviceContext特化实现。
| } | ||
|
|
||
| // Support x and y are different data types | ||
| Tensor Div_(const Tensor& x, const Tensor& y) { |
There was a problem hiding this comment.
函数名不应该以_区分功能，以及这个是不是可以直接实现在Div里面，判断一下x和y的dtype是不是相同？但其实x和t的dtype也是有要求的，一个是实数，一个是实数对应的复数？
| return out; | ||
| } | ||
|
|
||
| framework::Tensor Sub_(const framework::Tensor& x, |
There was a problem hiding this comment.
这个Sub为啥要重新实现？若是因为原来的Sub没有调用InverseFunctor，我也认为是Sub实现的bug，应该直接修改原Sub函数。另外，Sub的修改参考下Div，GPU是不需要InverseFunctor的。
| self.x_np = np.random.random(self.x_shape).astype(self.x_type) | ||
|
|
||
| def test_check_output(self): | ||
| self.check_output(no_check_set=['Eigenvectors']) |
There was a problem hiding this comment.
需要解释一下为什么会出现符号上的差别。以及说明下在其他单测里面会检查Eigenvectors。
PR types
New features
PR changes
OPs
Describe
添加 paddle.linalg.eigh() API， 支持CPU和GPU计算.