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.

Warning: Currently it only works using the default GPU (index 0)

Usage

• Install and test

  conda activate <CONDA_ENVIRONMENT>
  ./install.sh
  ./test.sh

• Use RoIAlign or crop_and_resize

  from roi_align.roi_align import RoIAlign      # RoIAlign modulefrom roi_align.roi_align import CropAndResize # crop_and_resize module# input dataimage = 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 layerroi_align = RoIAlign(crop_height, crop_width)crops = roi_align(image, boxes, box_index)

• Issue1: gradcheck and difference of RoIAlign and crop_and_resize.

• Changing -arch in make.sh for your GPU

  # Which CUDA capabilities do we want to pre-build for?
  # https://developer.nvidia.com/cuda-gpus
  # Compute/shader model   Cards
  # 6.1                    P4, P40, Titan Xp, GTX 1080 Ti, GTX 1080
  # 6.0                    P100
  # 5.2                    M40, Titan X, GTX 980
  # 3.7                    K80
  # 3.5                    K40, K20
  # 3.0                    K10, Grid K520 (AWS G2)