New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[API/OP]Add a new API paddle.diagonal #33586

Merged

zhiqiu merged 19 commits into PaddlePaddle:develop from zhangbo9674:dev/diagonal

Jun 22, 2021

Contributor

zhangbo9674 commented Jun 16, 2021 •

edited

PR types

New features

PR changes

APIs

Describe

This OP computes the diagonals of the input tensor x.

If x is 2D, returns the diagonal.
If x has larger dimensions, diagonals be taken from the 2D planes specified by axis1 and axis2.
By default, the 2D planes formed by the first and second axis of the input tensor x.

The argument offset determines where diagonals are taken from input tensor x:
- If offset = 0, it is the main diagonal.
- If offset > 0, it is above the main diagonal.
- If offset < 0, it is below the main diagonal.

Args:

x(Tensor): The input tensor x. Must be at least 2-dimensional. The input data type should be int32, int64, float32, float64.
offset(int, optional): Which diagonals in input tensor x will be taken. Default: 0 (main diagonals).
axis1(int, optional): The first axis with respect to take diagonal. Default: 0.
axis2(int, optional): The second axis with respect to take diagonal. Default: 1.
name (str, optional): Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Default: None.

if attr axis1 or axis2 is over the dims if imput Tensor, it will return a OutOfRange error.
if attr offset is over the range of [axis1, axis2], it will set the new dim as 0.

Returns:
Tensor: the output data type is the same as input data type.

Examples:

import paddle

x = paddle.rand([2,2,3],'float32')
print(x)
# Tensor(shape=[2, 2, 3], dtype=float32, place=CUDAPlace(0), stop_gradient=True,
#        [[[0.45661032, 0.03751532, 0.90191704],
#          [0.43760979, 0.86177313, 0.65221709]],

#         [[0.17020577, 0.00259554, 0.28954273],
#          [0.51795638, 0.27325270, 0.18117726]]])

out1 = paddle.diagonal(x)
print(out1)
#Tensor(shape=[3, 2], dtype=float32, place=CUDAPlace(0), stop_gradient=True,
#       [[0.45661032, 0.51795638],
#        [0.03751532, 0.27325270],
#        [0.90191704, 0.18117726]])

out2 = paddle.diagonal(x, offset=0, axis1=2, axis2=1)
print(out2)
#Tensor(shape=[2, 2], dtype=float32, place=CUDAPlace(0), stop_gradient=True,
#       [[0.45661032, 0.86177313],
#        [0.17020577, 0.27325270]])

out3 = paddle.diagonal(x, offset=1, axis1=0, axis2=1)
print(out3)
#Tensor(shape=[3, 1], dtype=float32, place=CUDAPlace(0), stop_gradient=True,
#       [[0.43760979],
#        [0.86177313],
#        [0.65221709]])

out4 = paddle.diagonal(x, offset=0, axis1=1, axis2=2)
print(out4)
#Tensor(shape=[2, 2], dtype=float32, place=CUDAPlace(0), stop_gradient=True,
#       [[0.45661032, 0.86177313],
#        [0.17020577, 0.27325270]])

API doc：

zhangbo9674 added 5 commits

June 15, 2021 04:44


          new api diagonal, test=develop

af143e2


          add new api diagonal, test=develop

d046b6b


          Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

e14531a

… dev/diagonal


          new api diagonal, test=develop

b5951da


          add new api paddle.diagonal, test=develop

6706c76

paddle-bot-old bot commented Jun 16, 2021

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

zhiqiu reviewed

View reviewed changes

paddle/fluid/operators/diagonal_op.cc Outdated

+                using framework::OperatorWithKernel::OperatorWithKernel;
+                void InferShape(framework::InferShapeContext *ctx) const override {
+                  PADDLE_ENFORCE_EQ(

Contributor

zhiqiu Jun 16, 2021

Better use OP_INOUT_CHECK

Contributor Author

zhangbo9674 Jun 16, 2021

Done, thanks!

paddle/fluid/operators/diagonal_op.cc Outdated

+                      platform::errors::NotFound("Output of DiagonalOp is not found."));
+                  int offset_ = ctx->Attrs().Get<int>("offset");
+                  int dim1 = ctx->Attrs().Get<int>("axis1");

Contributor

zhiqiu Jun 16, 2021

Maybe dim1->axis1 is better, to be consistent with op's attr.

Contributor Author

zhangbo9674 Jun 16, 2021

Done, thanks!

paddle/fluid/operators/diagonal_op.cc Outdated

+                  PADDLE_ENFORCE_NE(dim1_, dim2_,
+                                    platform::errors::InvalidArgument(
+                                        "The dimensions should not be identical "
+                                        "%ld vs %ld.",

Contributor

zhiqiu Jun 16, 2021

%d for int

Contributor Author

zhangbo9674 Jun 16, 2021

Done, thanks!

paddle/fluid/operators/diagonal_op.cc Outdated

Comment on lines 65 to 68

+                  auto dim1_size = out_dims[dim1_];
+                  auto dim2_size = out_dims[dim2_];
+                  out_dims.erase(out_dims.begin() + std::max(dim1_, dim2_));
+                  out_dims.erase(out_dims.begin() + std::min(dim1_, dim2_));

Contributor

zhiqiu Jun 16, 2021

add some comments

Contributor Author

zhangbo9674 Jun 16, 2021

Done, thanks!

paddle/fluid/operators/diagonal_op.cc Outdated

+                      "(Tensor) The partial view of input with the its diagonal elements.");
+                  AddAttr<int>(
+                      "offset",
+                      R"DOC((int, default 0), offset of the diagonal from the main diagonal. Can be both positive and negative. Defaults: 0.

Contributor

zhiqiu Jun 16, 2021

Defaults -> Default

Contributor Author

zhangbo9674 Jun 16, 2021

Done, thanks!

paddle/fluid/operators/diagonal_op.cu Outdated

		@@ -0,0 +1,362 @@
		/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.

Contributor

zhiqiu Jun 16, 2021

2019 -> 2021

Contributor Author

zhangbo9674 Jun 16, 2021

Done, thanks!

paddle/fluid/operators/diagonal_op.cu Outdated

Comment on lines 92 to 96

+                  cudaMalloc(reinterpret_cast<void**>(&input_stride),
+                             input_stride_size * sizeof(int64_t));
+                  cudaMemcpy(reinterpret_cast<void*>(input_stride),
+                             reinterpret_cast<void*>(host_input_stride),
+                             input_stride_size * sizeof(int64_t), cudaMemcpyHostToDevice);

Contributor

zhiqiu Jun 16, 2021

Do not use cudaMemcpy/cudaMalloc in op kernel directly. Firstly, try to use less memory as possible, i.e., do not malloc extra memory if can be avoid by refine the algorithm. Secondly, If temporary memory is exactly needed, use temp Tensor, and call tensor.mutable_data.

Contributor Author

zhangbo9674 Jun 17, 2021

Replace cudaMalloc/cudaMemcpy by TensorFormVector in cudaKernel and cudaGradKernel.

zhiqiu reviewed

View reviewed changes

paddle/fluid/operators/diagonal_op.cu Outdated Show resolved Hide resolved

zhangbo9674 added 5 commits

June 16, 2021 10:48


          use framework::stride replace ComputeDimStride


          replace cudaMalloc/cudaMemcpy by TensorFormVector in cudaKernel and c…

5a8b4c9

…udaGradKernel


          Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

9c1355b

… dev/diagonal


          perfect funciton: when attr(offset) is exceed attr(axis1) or attr(axi…

89a794a

…s2), set the diagonal dim is 0


          fix RP-Mac-CI bug: replace framework::stride() by ComputDimStride.

726ab58

zhangbo9674 changed the title ~~Add a new API paddle.diagonal~~ [API/OP]Add a new API paddle.diagonal

TCChenlong reviewed

View reviewed changes

python/paddle/tensor/__init__.py

@@ @@ -355,6 +356,8 @@ @@
                          'shape',
                          'real',
                          'imag',
+                         'digamma',

Contributor

TCChenlong Jun 18, 2021

这行要删除，和362重复

Contributor Author

zhangbo9674 Jun 18, 2021

Done, thanks!

python/paddle/tensor/math.py Outdated

+                          #         [[0.50427347, 0.78351408, 0.00833563],
+                          #          [0.36932808, 0.83527362, 0.49408615]]])
+                          out = paddle.diagonal(x)

Contributor

TCChenlong Jun 18, 2021

需要补充以下几个示例：

axis1 = 1的示例
offset != 0，axis1 = 1 的示例
offset != 0、axis1 = 1、axis2 != 1 的示例

Contributor Author

zhangbo9674 Jun 18, 2021

Done, thanks!

python/paddle/tensor/math.py Outdated

+                      name (str, optional): Normally there is no need for user to set this property. For more information, please refer to :ref:`api_guide_Name`. Default: None.
+                  Returns:
+                      Tensor: the output data type is the same as input data type.

Contributor

TCChenlong Jun 18, 2021

少了，输入 Tensor 在指定二维平面的局部视图这一句

Contributor Author

zhangbo9674 Jun 18, 2021

Done, thanks!

zhangbo9674 added 2 commits

June 18, 2021 11:24


          perfect code-block

219a664


          Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

f426951

… dev/diagonal

lanxianghit reviewed

View reviewed changes

python/paddle/tensor/math.py Outdated

+                  if in_dygraph_mode():
+                      return core.ops.diagonal(x, 'offset', offset, 'axis1', axis1, 'axis2', axis2)
+                  if not in_dygraph_mode():

Contributor

lanxianghit Jun 21, 2021

这行代码是否多余？

Contributor Author

zhangbo9674 Jun 21, 2021

已删除，谢谢！

python/paddle/tensor/math.py Outdated

+                  """
+                  inputs = {'Input': [x]}
+                  attrs = {'offset': offset, 'axis1': axis1, 'axis2': axis2}

Contributor

lanxianghit Jun 21, 2021

这两个操作为什么放在动态图调用之前？

Contributor Author

zhangbo9674 Jun 21, 2021

该段代码没有什么意义，已删除，谢谢！

python/paddle/tensor/math.py Outdated

+                  - If offset < 0, it is below the main diagonal.
+                  Args:
+                      x(Tensor): The input tensor x. Must be at least 2-dimensional. The input data type should be int32, int64, float32, float64.

Contributor

lanxianghit Jun 21, 2021

应该考虑支持bool、fp16等更多类型

Contributor Author

zhangbo9674 Jun 21, 2021

已添加支持，谢谢！

zhangbo9674 added 5 commits

June 21, 2021 11:59


          perfect code of python API diagonal

fa981e1


          api supports dtype of float16 and bool

50f7ef4


          api supports dtype of float16 and bool

14a3b88


          modify unittest code

1e123ba


          modify unittest code

ae54477

lanxianghit reviewed

View reviewed changes

python/paddle/tensor/math.py Outdated

+                  - If offset < 0, it is below the main diagonal.
+                  Args:
+                      x(Tensor): The input tensor x. Must be at least 2-dimensional. The input data type should be int32, int64, float32, float64, float16, bool.

Contributor

lanxianghit Jun 22, 2021

Suggested change

      
                    x(Tensor): The input tensor x. Must be at least 2-dimensional. The input data type should be int32, int64, float32, float64, float16, bool.
          
                    x(Tensor): The input tensor x. Must be at least 2-dimensional. The input data type should be bool, int32, int64, float16, float32, float64.

Contributor Author

zhangbo9674 Jun 22, 2021

Done, thanks!


          perfect dtype describe

16ada2a

lanxianghit previously approved these changes

View reviewed changes

Contributor

lanxianghit left a comment

LGTM


          perfect code-block

f3932a1

zhangbo9674 dismissed lanxianghit’s stale review via

f3932a1

June 22, 2021 08:04

TCChenlong approved these changes

View reviewed changes

Contributor

TCChenlong left a comment

LGTM

lanxianghit approved these changes

View reviewed changes

zhiqiu approved these changes

View reviewed changes

Contributor

zhiqiu left a comment

LGTM

zhiqiu merged commit ad10629 into PaddlePaddle:develop

zhangbo9674 deleted the dev/diagonal branch

September 14, 2022 02:22

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment