RoIAlign for PyTorch

This is a PyTorch version of RoIAlign. This implementation is based on crop_and_resize and supports both forward and backward on CPU and GPU.

Introduction

The crop_and_resize function is ported from tensorflow, and has the same interface with tensorflow version, except the input feature map should be in NCHW order in PyTorch. They also have the same output value (error < 1e-5) for both forward and backward as we expected, see the comparision in test.py.

Note: Document of crop_and_resize can be found here. And RoIAlign is a wrap of crop_and_resize that uses boxes with unnormalized (x1, y1, x2, y2) as input (while crop_and_resize use normalized (y1, x1, y2, x2) as input). See more details about the difference of RoIAlign and crop_and_resize in tensorpack.

Usage

IMPORTANT: check your GPU's capability and change the -arch flag in line 7 of make.sh accordingly to your specs.

    $CUDA_PATH/bin/nvcc -c -o crop_and_resize_kernel.cu.o crop_and_resize_kernel.cu -x cu -Xcompiler -fPIC -arch=sm_52

E.g. On my laptop I have a Geforce GTX 960M which has capability 5.0 so I changed sm_52 with sm_50 and it worked.

• Build and test

  chmod a+x make.sh
  ./make.sh
  python test.py

• Use RoIAlign or crop_and_resize

  from roi_align.roi_align import RoIAlign      # RoIAlign module
  from roi_align.roi_align import CropAndResize # crop_and_resize module
  
  # input data
  image = to_varabile(image_data, requires_grad=True, is_cuda=is_cuda)
  boxes = to_varabile(boxes_data, requires_grad=False, is_cuda=is_cuda)
  box_index = to_varabile(box_index_data, requires_grad=False, is_cuda=is_cuda)
  
  # RoIAlign layer
  roi_align = RoIAlign(crop_height, crop_width)
  crops = roi_align(image, boxes, box_index)

• Issue1: gradcheck and difference of RoIAlign and crop_and_resize.

Example

Check crop_and_resize_example.py to see how crop_and_resize works