BIGKnight / deformable_conv2d_v2_tensorflow

deformablt_conv2d layer implemented in tensorflow

Deformable_Conv2d_Tensorflow

deformable_conv2d layer implemented in Tensorflow.
besides, I also complete an example net, here and I'm very sorry that I did not implement the swapaxis methods, so the im2col_step parameter are only allowed using value one.

ENVIRONMENT CONFIGURATION

1. OS: ubuntu16.04
   
2. GPU: 1 gtx1080Ti
   
3. LANGUAGE: python3.5 & c++11 & cuda c
   
4. DL FRAMEWORK: tensorflow_gpu 1.12.0
   
5. ANCILLARY LIB: numpy: 1.15.4, etc
6. GPU API: NVIDIA CUDA 9.0 & cuDNN 7.0
7. COMPILE: nvcc & gcc 5.4.0

INSTALL PROCEDURE

1. cd "current project"
2. run make.sh
   tips: you need to modify the path parameters first. and all the -I and -L path in the nvcc and g++ orders need to be checked, make sure they are the correct path in your system
   
3. load the dynamic lib which created by the make.sh, and do the python wrapper(see the script:deformable_conv2d.py)