Skip to content

Latest commit

 

History

History
124 lines (88 loc) · 4.76 KB

README.md

File metadata and controls

124 lines (88 loc) · 4.76 KB
Raw

Classification of Long Sequential Data using Circular Dilated Convolutional Neural Networks

paper: https://doi.org/10.1016/j.neucom.2022.10.054.

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 xor_generation.py;
  2. Run xor_main.py for one model or run xor_all.sh for all models.

The generator will create 3 files for each sequence length and store them in the ./xor_datasets/ folder in the following format: xor_{length}_train.pt xor_{length}_test.pt xor_{length}_val.pt

The ./xor_log/ folder will save all results. The ./xor_model/ folder will save all best models.

To reproduce the dissimilar experiment results, you should:

  1. Run dissimilar_generation.py;
  2. Run dissimilar_main.py for one model or run dissimilar_all.sh for all models.

The generator will create 4 files for sequence length of 2048 and store them in the ./dissimilar_datasets/ folder in the following format: dissimilar_2048_train.pt dissimilar_2048_test.pt dissimilar_2048_val.pt dissimilar_2048_dtest.pt

The ./dissimilar_log/ folder will save all results. The ./dissimilar_model/ folder will save all best models.

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

The ./lra_log/ folder will save all results. The ./lra_model/ folder will save all best models.

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 or for one experiment or run time_all.sh for all experiments.

The generator will create 3 files for each dataset and store them in the ./time_datasets/ folder in the following format: {dataset}_train.csv {dataset}_val.csv {dataset}_test.csv

The ./time_log/ folder will save all results. The ./time_model/ folder will save all best models.

To reproduce the noisy RightWhaleCalls results, you should:

  1. Run noise_generation.py;
  2. Run noise_main.py for one model or run noise_all.sh for all models.

The generator will create 8 files for each dataset and store them in the ./noise_datasets/ folder in the following format: RightWhaleCalls_train_data.csv RightWhaleCalls_train_label.csv RightWhaleCalls_val_data.csv RightWhaleCalls_val_label.csv RightWhaleCalls_test_data.csv RightWhaleCalls_test_label.csv RightWhaleCalls_dtest_data.csv RightWhaleCalls_dtest_label.csv

The ./noise_log/ folder will save all results. The ./noise_model/ folder will save all best models.

We provide our used configurations in time_config.py.

Cite

@article{cheng2022classification,
  title={Classification of long sequential data using circular dilated convolutional neural networks},
  author={Cheng, Lei and Khalitov, Ruslan and Yu, Tong and Zhang, Jing and Yang, Zhirong},
  journal={Neurocomputing},
  year={2022},
  publisher={Elsevier}
}