Distribution padding

This is the official implementation of the ICIP paper "DISTRIBUTION PADDING IN CONVOLUTIONAL NEURAL NETWORKS".

Environment

Recently, I have experienced some NaN error caused by my custom frac_bilinear_upsampling. Therefore, the following environment is highly recommended.

Python == 3.6

Theano == 1.0.3

neuralnet

CUDA == 9.0

CuDNN == 7.5

NVDIA GPU TiTAN X 12GB

If you encounter NaN, one workaround is to use the Theano's frac_bilinear_upsampling. However, it will require some modification as the op does not accept symbolic variables for the resize argument.

Training and evaluation

To train the model on CIFAR10, simply run

python train.py path/to/CIFAR10

If the dataset is not in path/to/CIFAR10, it will be automatically downloaded. By default, the model architecture is ResNet34 with mean interpolation padding. Training will run for 100 epochs and evaluates on the test set every 1000 iterations. Type python train.py -h to see more customizations.

Results

Citation

If you find this code helpful for your research, please consider citing the work

@INPROCEEDINGS{DistPaddingNguyen2019,
author={A. {Nguyen} and S. {Choi} and W. {Kim} and S. {Ahn} and J. {Kim} and S. {Lee}},
booktitle={2019 IEEE International Conference on Image Processing (ICIP)},
title={Distribution Padding in Convolutional Neural Networks},
year={2019},
volume={},
number={},
pages={4275-4279},
keywords={Deep learning;convolutional neural network;image padding},
doi={10.1109/ICIP.2019.8803537},
ISSN={},
month={Sep.},}

References

The partial convolution is adapted and simplified from the official repo.