pytorch STN

This is a third party implementation of STN in pytorch which written in paper

Requirements

• opencv(for load data faster)
• python3
• Pytorch >= 0.4.1
• torchvision
• numpy

TODO

• Network building
• Multi STN
• Train

Results

Network	Test Acc
Paper Best	84.1
Implement	85.71

Usage

For training, use following command.

$ python train.py

Currently only cuda available device support.

docker

See README.md

Visualization

single stn (learn [[a, 0, c], [0, e, f]])

multi stn (learn [[parallel, 0, c], [0, parallel, f]], parallel=[0.9, 0.7, 0.5, 0.3, 0.1]),

Seen from the results, parallel=[0.9, 0.7, 0.5] is enough.

References