Ali-Sahili / STN-CoordConv

Pytorch implementation of spatial transformer networks (STN), CoordConv for ConvLayers and some experiments.

STN-CoordConv

Pytorch implementation of spatial transformer networks (STN) and CoordConv for ConvLayers with some experiments on toy datasets.

Implemented Methods/Papers

• Spatial Transformer Networks
• An Intriguing Failing of Convolutional Neural Networks and the CoordConv Solution
• STN-OCR: A single Neural Network for Text Detection and Text Recognition

Requirements

The experiments were performed using Python 3.8.5 with the following Python packages:

• numpy == 1.18.5
• torch == 1.5.1
• torchvision = 0.6.1
• matplotlib == 3.3.3

Usage

To play with my implementation, you can simply put the following command into your terminal after adjusting the necessary parameters:

python3 main.py [--seed SEED] [--use_cuda USE_CUDA]
                [--batch_size BATCH_SIZE] [--lr LR]
                [--num_workers NUM_WORKERS] 
                [--num_epochs NUM_EPOCHS] 
                [--optimizer OPTIMIZER] [--beta1 BETA1] [--beta2 BETA2]

Acknowledgements

1. implementation of STN by Ankit.
2. implementation of CoordConv by Chao Wen.