Introduction

• This repository contains the PyTorch implementation of deformabel covolution network, including v1 and v2 version. The origin papers can be found from and . And the official implementation is .
• Only deformable convolution is implemented. The deformable RoIPooling is not yet implemented.
• This repository supports PyTorch > 0.4.0, and the codes have been tested with PyTorch = 1.0.0. And you can get PyTorch < 0.4.0 version from .
• Any questions are welcomed.

Installation and usage

• The installation of Deformable-ConvNets is much easy:

$ git clone https://github.com/liuqk3/Deformable-ConvNets-PyTorch.git
$ unzip Deformable-ConvNets-PyTorch.zip
$ cd Deformable-ConvNets-PyTorch/src
$ python jit.py

• Note that there are two ways of building C++ extensions: using setuptools or just in time (JIT). Here I use the JIT compilation. You can check the for more information about the difference between these two ways.
• To use the Deformable-ConvNets, see the example in ./test.py.

To-do list

• Implement deformable RoIPooling.
• Support PyTorch > 0.4, including PyTorch 1.0.