jessie-taylor/periodicnetwork

Cyclic-permutation invariant networks for classifying periodic time-series

Cyclic-Permutation Invariant Neural Networks

Code associated with the paper:

Classification of Periodic Variable Stars with Novel Cyclic-Permutation Invariant Neural Networks
Keming Zhang & Joshua Bloom
Submitted to MNRAS
arXiv:2011.01243
Accepted to ICLR 2020: FSAI (spotlight talk)
Accepted to NeurIPS 2020: ML4PS

This repository contains pytorch implementations of the cyclic-permutation invariant networks used in our study. It also contains code to reproduce our main results.

Neural network implementations can be found under ./model/

Reproducing our results

Description of files:
train.py: train neural networks on variable star light curve datasets
ppmnist.py: train neural networks on the periodic permuted MNIST task
run_variablestar.sh: commmands for reproducing our variable star experiements
run_ppmnist.sh: commmands for reproducing our PP-MNIST experiements
data/download.sh: script to download the datasets
trained_models.tar: trained models of Table 1 of the paper

We provide two options for reproducing our results. To test on our provided trained models, decompress trained_models.tar into the ./results first. Alternatively, you may opt to train these models from stretch simply by removing the --test option in the run_variablestar.sh commands. If you do so, we suggest that a new conda environment be created from environment.yml to replicate the identical software environment.

You may use the --ngpu and the --njob options to facilitate parallel training/testing with multiple gpus or processes. If GPU device is not available for your device, you can still specify --njob for parallel training/testing. The code automatically detects the availability of GPUs.

The variable star light curve datasets used in this study have been uploaded to zenodo, and can be downloaded using the script in ./data/download.sh. These datasets have been constructed from publicly available databases. If you use this dataset, please cite the original papers, the citation of which can be found in https://zenodo.org/record/3903015

Adapting to your own data

Code for iTCN and iResNet can be found under model. To implement cyclic-permutation invariance in your custom CNN archetecture written in pytorch, add the wrap module under model/padding.py before any nn.Conv1D module and remove existing padding.

You might find times_to_lags under util.py useful for transforming times to time intervals which accounts for periodicity.

def times_to_lags(x, p=None):
    lags = x[:, 1:] - x[:, :-1]
    if p is not None:
        lags = np.c_[lags, x[:, 0] - x[:, -1] + p]
    return lags

While the training code has been optimized for reproducebility rather than flexibility, you can use our training code train.py and data structure light_curve.py as a starter code. The data structure has orginally been adapted from Naul et al. 2018. To train a model, create a LightCurve() object for each of your light curves, and save the list of LightCurve() objects to a pickle file. From there, follow instructions for reproduction.