Classification of Long Sequential Data using Circular Dilated Convolutional Neural Networks

arXiv preprint: https://arxiv.org/abs/2201.02143.

Architecture

CDIL-CNN is a novel convolutional model for sequence classification. We use symmetric dilated convolutions, a circular mixing protocol, and an average ensemble learning.

Symmetric Dilated Convolutions

Circular Mixing

CDIL-CNN

Experiments

Synthetic Task

To reproduce the synthetic data experiment results, you should:

1. Run syn_data_generation.py;
2. Run syn_main.py for one experiment or run syn_all.sh for all experiments.

The generator will create 6 files for each sequence length and store them in the syn_datasets folder in the following format: adding2000_{length}_train.pt adding2000_{length}_train_target.pt adding2000_{length}_test.pt adding2000_{length}_test_target.pt adding2000_{length}_val.pt adding2000_{length}_val_target.pt

By default, it iterates over 8 sequence lengths: [2**7, 2**8, 2**9, 2**10, 2**11, 2**12, 2**13, 2**14].

You can run different models for different lengths. The syn_log folder will save all results.

We provide our used configurations in syn_config.py.

Long Range Arena

Long Range Arena (LRA) is a public benchmark suite. The datasets and the download link can be found in the official GitHub repository.

To reproduce the LRA experiment results, you should:

1. Download lra_release.gz (~7.7 GB), extract it, move the folder ./lra_release/lra_release into our create_datasets folder, and run all_create_datasets.sh.
2. Run lra_main.py for one experiment or run lra_all.sh for all experiments.

The dataset creators will create 3 files for each task and store them in the lra_datasets folder in the following format: {task}.train.pickle {task}.test.pickle {task}.dev.pickle

You can run different models on different tasks. The lra_log folder will save all results.

We provide our used configurations in lra_config.py.

Time Series

The UEA & UCR Repository consists of various time series classification datasets. We use three audio datasets: FruitFlies, RightWhaleCalls, and MosquitoSound.

To reproduce the time series results, you should:

1. Download the datasets, extract them, move the extracted folders into our time_datasets folder, and run time_arff_generation.py.
2. Run time_main.py for one experiment or run time_all.sh for all experiments.

The generator will create 2 files for each dataset and store them in the time_datasets folder in the following format: {dataset}_train.csv {dataset}_test.csv

You can run different models on different datasets. The time_log folder will save all results.

We provide our used configurations in time_main.py.