MrCAE: multiresolution convolutional autoencoders (with adaptive filters)

Table of contents

• Structure
• Introduction
• Usage
• Citation

Structure

MrCAE/
  |- README.md
  |- src/
    |- __init__.py
        |- utils.py
        |- torch_cae_multilevel_V4.py (*this is our MrCAE torch code*)
    |- torch_cae_skip_connection.py (classical benchmark)
  |- scripts/
    |- MrCAE_{xxxx}.ipynb (individual experiments)
    |- benchmark_{xxxx}.ipynb (benchmark experiments)
    |- training/ (contains scripts for batch training on clusters)
  |- data/ (contains simulation data, fluids data, SST data)
  |- model/(contains models at different stages)
  |- result/(contains training results)

Introduction

• This repo provides the code for the paper "Multiresolution Convolutional Autoencoders" by Yuying Liu, Colin Ponce, Steven L. Brunton and J. Nathan Kutz (in review).
• The network architecture is recursively built up to process data across different resolutions — architectures built for processing coarser data are later embedded into the next-level architectures to ensure knowledge transfer.

• Within each level, we perform one deepening operation and a sequence of widening operations. The deepening operation inserts a convolutional/deconvolutional layer between the current and previous level inputs/outputs. This is the transfer learning step. The widening operation expands the network capacity in order to capture the new, finer-grained features of the higher resolution data.

• The convolutional filters we use are highly adaptive: they only refine the regions that produce large reconstructions errors.

• By utilizeing the above features, our network can progressively 'grow' itself and end up with the 'right' amount of parameters for characterzing the data. Different level of performances can be achieved at different stages of training.

Usage

• We provide three methods to train the architecture: train_net(), train_net_one_level() and train_net_one_stage() which corresponds to (i) end-to-end training, (ii) train one level only and (iii) train one stage only (after an architecture change).
• See demos under scripts/ for more details
• Notes: training on GPUs is highly recommended.

Citation

@article{liu2020multiresolution,
  title={Multiresolution Convolutional Autoencoders},
  author={Liu, Yuying and Ponce, Colin and Brunton, Steven L and Kutz, J Nathan},
  journal={arXiv preprint arXiv:2004.04946},
  year={2020}
}