Awesome

PyTorch version of spatial transformer network

Ported from https://github.com/qassemoquab/stnbhwd according to pytorch tutorial. Now support CPU and GPU. To use the ffi you need to install the cffi package from pip.

Build and test

cd script
./make.sh #build cuda code, don't forget to modify -arch argument for your GPU computational capacity version
python build.py
python test.py

There is a demo in test_stn.ipynb

Modules

STN is the spatial transformer module, it takes a B*H*W*D tensor and a B*H*W*2 grid normalized to [-1,1] as an input and do bilinear sampling.

AffineGridGen takes a B*2*3 matrix and generate an affine transformation grid.

CylinderGridGen takes a B*1 theta vector and generate a transformation grid to remap equirectangular images along x axis.

DenseAffineGridGen takes a B*H*W*6 tensor and do affine transformation for each pixel. Example of convolutional spatial transformer can be found in test_conv_stn.ipynb.

An example of the landscape of the loss function of a simple STN with L1 Loss can be found in the demo.

Train hacks

• set a learning rate multiplier, 1e-3 or 1e-4 would work fine.
• add an auxiliary loss to regularized the difference of the affine transformation from identity mapping, to aviod sampling outside the original image.

Complex grid demo

STN is able to handle a complex grid, however, how to parameterize the grid is a problem.