This is a non-official PyTorch implementation of PSANet: Point-wise Spatial Attention Network for Scene Parsing. The official implementation can be found in here. This code has been tested on Python 2.7, PyTorch 0.3.1.
Change the corresponding architecture of your CUDA device in ./PSACUDA/make.sh with -arch=sm_xx.
The definition of architectures can be found in here.
cd ./PSACUDA
./make.sh
This will build the CUDA kernel code by NVCC and create a python extension.
from PSACUDA.moudles.PSANet import PSANetModule
net=PSANetModule(in_channels,reduced_channels,fea_h,fea_w,keep_channel_size = False)
- in_channels - the number of input feature channel
- reduced_channels - the number of channels after first reduction, i.e. the C2 in the paper of PSANet
- fea_h - the hight of input
- fea_w - the width of input
- keep_channel_size(optional) - if keep_channel_size is True, the output of this module would be the same size as the input.