tensorflow-wrapped layers from Caffe
available layers (GPU implementation only):
coordinate_roiwarp - roi warping with backprop on bbox coords only
feature_maskpool - mask pooling with backprop on features only
feature_roiwarp - roi warping with backprop on features only
girshick_roipool - roi pooling from base faster rcnn
mask_resize - mnc mask resizing
ps_roipool - position sensitive roi pooling for rfcn
roiwarp - mnc roi warping with backprop on both bbox coords and features
original caffe layers for girshick_roipool are from [py-faster-rcnn] (https://github.com/rbgirshick/py-faster-rcnn), by rbgirshick
for roiwarp, and mask pooling / resizing, original caffe layers are from [mnc] (https://github.com/daijifeng001/MNC), by daijifeng001
ps_roipool is from [py-rfcn] (https://github.com/Orpine/py-R-FCN), by Orpine
created by A. Labao and P. Naval, CVMIG lab, Univ of the Philippines
to install, copy desired folder (i.e. roiwarp) to /tensorflow/tensorflow/core/user_ops/, and build .so file through bazel
bazel build -c opt --config=cuda //tensorflow/core/user_ops/<folder_name>:<so_name>
the folder user_ops_tests contains python files for testing gradients and registering gradients in tensorflow
the layers are tested with the ff specs:
tensorflow v0.10 and up
python 2.7
python numpy
CuDNN v5
CUDA v8.0