Chainer implementations of Spatial Transformer Networks https://arxiv.org/abs/1506.02025.
This implementation tries to reproduce the distorted MNIST dataset described in the original paper.
An animation of the transformation grids from iteration 0 to 200 using a batch size of 128.
Loss and accuracy plots of the ST-CNN model, compared to a CNN without the spatial transformer (ST) layer, CNN(Pooling). Since the first layer becomes an average pooling layer, we also plot the training curves of a CNN without both the ST layer and the average pooling operation, CNN.
python train.py --max-iter 1000 --out result --gpu 0