What does it do?

It works like torch grid_sample with bilinear interpolation and zero padding, but takes in uint8 datatype with shape (B, H, W), where each bit of a pixel corresponds to a channel. Encoding can be done like

sum(x[:, i] * 2**i for i in range(x.shape[1]))

Gradient flow through the input grid is supported.

Why?

To safe GPU memory. Consider the case where you want to sample small regions from large images, which only consist of binary masks. E.g. CMYK printing motives. Torch grid_sample only works for floating point types, so in case of 4 channels, memory consumption is reduced by sizeof(float) * channels = 4 * 4 = 16.

Installation

Edit the arguments BLOCKSIZE in setup.py, accordingly to your requirements. Then execute setup.py as needed, e.g.

python setup.py develop