modulated-deform-conv

该项目是一个 Pytorch C++ and CUDA Extension,采用C++和Cuda实现了deformable-conv2d,modulated-deformable-conv2d,deformable-conv3d,modulated-deformable-conv3d的forward function和backward function,并在Python中对其进行了包装。
This Project is a Pytorch C++ and CUDA Extension, which implements the forward function and backward function for deformable-conv2d, modulated-deformable-conv2d, deformable-conv3d, modulated-deformable-conv3d, then encapsulates C++ and CUDA code into Python Package.

安装 Install

run pip install modulated-deform-conv
or git clone https://github.com/CHONSPQX/modulated-deform-conv.git,then cd modulated-deform-conv and run python setup.py install

要求 Requires

Python 3
Pytorch>=1.3
Linux, gcc版本>=4.9(For Linux, gcc version>=4.9)
Windows,CUDA版本需要VS版本兼容(For Windows, CUDA version must be compatiable with Visual Studio version)

由于资源有限，目前测试过的环境有(Because of limited resources, only the following environment are tested)

Ubuntu18.04 , gcc 7.4 , CUDA 10.2 ,Python3.7.4, Pytorch 1.3.1
Ubuntu18.04 , gcc 7.4 , CUDA 10.2 ,Python3.7.4, Pytorch 1.4.0
Ubuntu18.04 , gcc 7.5 , CUDA 11.1 ,Python3.6.12, Pytorch 1.7.0
Windows10 , Visual Studio 2017 , CUDA 10.1 ,Python3.7.6, Pytorch 1.4.0
Windows10 , Visual Studio 2019 , CUDA 11.1 ,Python3.6.12, Pytorch 1.7.0

速度优化 Speed Optimization

pip download modulated-deform-conv 解压得到的压缩文件，进入modulated-deform-conv，打开src/config.h，用户可根据自身显卡情况，设置以下两个变量，获得更快运行速度，然后运行 python setup.py install
Unzip the downloaded compressed file, cd modulated-deform-conv, then open src/config.h,users are recommended to set the following VARIABLES to optimize run speed according to their NVIDIA GPU condition, then run python setup.py install
- const int CUDA_NUM_THREADS
- const int MAX_GRID_NUM
运行时可以通过传递in_step参数来优化速度，该变量控制每次并行处理的batch 大小。
Or users can set different in_step value in run time, which controls the batch size of each parallel processing .

使用 Use

直接使用C++函数，请import MDCONV_CUDA 使用封装后的python类，请import modulated_deform_conv Using C++ functions directly, please import MDCONV_CUDA Using the packaged function by Python, please import modulated_deform_conv

文档 Documents

1.C++ and CUDA Code

文件 Files

Filename	Content
`config.h`	macro&gloabl variables&inline functions
`deformable_conv.cu`	MDCONV_CUDA.deform_conv2d_forward_cuda MDCONV_CUDA.deform_conv2d_backward_cuda
`mdeformable_conv.cu`	MDCONV_CUDA.modulated_deform_conv2d_forward_cuda MDCONV_CUDA.modulated_deform_conv2d_backward_cuda
`deformable_conv3d.cu`	MDCONV_CUDA.deform_conv3d_forward_cuda MDCONV_CUDA.deform_conv3d_backward_cuda
`mdeformable_conv3d.cu`	MDCONV_CUDA.modulated_deform_conv3d_forward_cuda MDCONV_CUDA.modulated_deform_conv2d_backward_cuda
`utils.cu`	some code for display debug outputs
`warp.cpp`	glue code between C++ and Python

变量 Variables

Variable Name	Type	Introduction
`kernel_h`	`const int`	first dimension size of the convolution kernel
`kernel_w`	`const int`	second dimension size of the convolution kernel
`kernel_l`	`const int`	third dimension size of the convolution kernel
`stride_h`	`const int`	stride for first dimension
`stride_w`	`const int`	stride for second dimension
`stride_l`	`const int`	stride for third dimension
`pad_h`	`const int`	zero padding for first dimension
`pad_w`	`const int`	zero padding for second dimension
`pad_l`	`const int`	zero padding for third dimension
`dilation_h`	`const int`	dilation rate for first dimension
`dilation_w`	`const int`	dilation rate for second dimension
`dilation_l`	`const int`	dilation rate for third dimension
`group`	`const int`	group of convolution
`deformable_group`	`const int`	group of offset and mask
`in_step`	`const int`	batch size of each parallel processing
`with_bias`	`const bool`	if have bias
`input`	`at::Tensor`	`B,I,H,W[,L]`,`I` must be divisible by`group` and `deformable_group`
`grad_input`	`at::Tensor`	`grad_input` must be size like `input`
`weight`	`at::Tensor`	`O,I/group,H,W[,L]`，`O`must be divisible by`group`
`grad_weight`	`at::Tensor`	`grad_weight` must be size like `weight`
`bias`	`at::Tensor`	`[O]`, if `with_bias=true`, `bias` must be non-null
`grad_bias`	`at::Tensor`	`grad_bias` must be size like `bias`
`offset`	`at::Tensor`	`B,deformable_group2kernel_hkernel_w,H,W` `B,deformable_group3kernel_hkernel_w*kernel_l,H,W,L`
`grad_offset`	`at::Tensor`	`grad_offset` must be size like `offset`
`mask`	`at::Tensor`	`B,deformable_groupkernel_hkernel_w,H,W` `B,deformable_groupkernel_hkernel_w*kernel_l,H,W,L`
`grad_mask`	`at::Tensor`	`grad_mask` must be size like `mask`
`output`	`at::Tensor`	`B,O,OH,OW[,OL]`
`grad_output`	`at::Tensor`	`grad_output` must be size like `output`

2.Python Code

Class Name	Type
`class DeformConv2dFunction`	`torch.autograd.Function`
`class ModulatedDeformConv2dFunction`	`torch.autograd.Function`
`class DeformConv3dFunction`	`torch.autograd.Function`
`class ModulatedDeformConv3dFunction`	`torch.autograd.Function`
`class DeformConv2d`	`torch.nn.Module`
`class ModulatedDeformConv2d`	`torch.nn.Module`
`class DeformConv3d`	`torch.nn.Module`
`class ModulatedDeformConv3d`	`torch.nn.Module`

Author

Xin Qiao [email protected]

github/chonspqx

License

TypeError: a bytes-like object is required, not 'str'

(pt) D:\detect\modulated-deform-conv> python setup.py install
running install
running bdist_egg
running egg_info
writing modulated_deform_conv.egg-info\PKG-INFO
Traceback (most recent call last):
File "setup.py", line 66, in
cmdclass={'build_ext': BuildExtension}, zip_safe=False)
File "C:\Users\84791.conda\envs\pt\lib\site-packages\setuptools_init_.py", line 144, in setup
return distutils.core.setup(**attrs)
File "C:\Users\84791.conda\envs\pt\lib\distutils\core.py", line 148, in setup
dist.run_commands()
File "C:\Users\84791.conda\envs\pt\lib\distutils\dist.py", line 955, in run_commands
self.run_command(cmd)
File "C:\Users\84791.conda\envs\pt\lib\distutils\dist.py", line 974, in run_command
cmd_obj.run()
File "C:\Users\84791.conda\envs\pt\lib\site-packages\setuptools\command\install.py", line 67, in run
self.do_egg_install()
File "C:\Users\84791.conda\envs\pt\lib\site-packages\setuptools\command\install.py", line 109, in do_egg_install
self.run_command('bdist_egg')
File "C:\Users\84791.conda\envs\pt\lib\distutils\cmd.py", line 313, in run_command
self.distribution.run_command(command)
File "C:\Users\84791.conda\envs\pt\lib\distutils\dist.py", line 974, in run_command
cmd_obj.run()
File "C:\Users\84791.conda\envs\pt\lib\site-packages\setuptools\command\bdist_egg.py", line 163, in run
self.run_command("egg_info")
File "C:\Users\84791.conda\envs\pt\lib\distutils\cmd.py", line 313, in run_command
self.distribution.run_command(command)
File "C:\Users\84791.conda\envs\pt\lib\distutils\dist.py", line 974, in run_command
cmd_obj.run()
File "C:\Users\84791.conda\envs\pt\lib\site-packages\setuptools\command\egg_info.py", line 290, in run
writer(self, ep.name, os.path.join(self.egg_info, ep.name))
File "C:\Users\84791.conda\envs\pt\lib\site-packages\setuptools\command\egg_info.py", line 622, in write_pkg_info
metadata.write_pkg_info(cmd.egg_info)
File "C:\Users\84791.conda\envs\pt\lib\distutils\dist.py", line 1106, in write_pkg_info
self.write_pkg_file(pkg_info)
File "C:\Users\84791.conda\envs\pt\lib\site-packages\setuptools\dist.py", line 167, in write_pkg_file
long_desc = rfc822_escape(self.get_long_description())
File "C:\Users\84791.conda\envs\pt\lib\distutils\util.py", line 474, in rfc822_escape
lines = header.split('\n')
TypeError: a bytes-like object is required, not 'str'

chonspqx / modulated-deform-conv Goto Github PK