Structural and Statistical Audio Texture Knowledge Distillation (SSATKD) for Passive Sonar Classification

Structural and Statistical Audio Texture Knowledge Distillation (SSATKD) for Passive Sonar Classification

Jarin Ritu, Amirmohammad Mohammadi, Davelle Carreiro, Alexandra Van Dine, Joshua Peeples

Note: If this code is used, cite it: Jarin Ritu, Amirmohammad Mohammadi, Davelle Carreiro, Alexandra Van Dine, Joshua Peeples. (2025, January 3) Peeples-Lab/SSATKD_Lighting: Initial Release (Version v1.0) Zendo. https://doi.org/10.5281/zenodo.14593898.

arXiv

BibTeX

In this repository, we provide the paper and code for " Structural and Statistical Audio Texture Knowledge Distillation (SSATKD) for Passive Sonar Classification."

Installation Prerequisites

The requirements.txt file includes all the necessary packages, and the packages will be installed using:

pip install -r requirements.txt

Or, a pre-configured environment file, ssatkd.yml, is available in the project directory. Follow these steps to set it up:

Open a terminal and navigate to the project directory.

Create the Conda environment: conda env create -f ssatkd.yml

Activate the environment: conda activate ssatkd

Run SSATKD!

Demo

To get started, please follow the instructions in the Datasets folder to download the dataset. Next, run demo.py in Python IDE (e.g., Spyder) or command line to train, validate, and test models. To evaluate performance, run View_Results.py (if results are saved out).

Main Functions

The target classification code uses the following functions.

1. Intialize model

   model = intialize_model(**Parameters)

2. Prepare dataset(s) for model

   data_module = DeepShipDataModule(**Parameters)

3. Train model

   trainer.fit(**Parameters)

4. Test model

   trainer.test(**Parameters)

Parameters

The parameters can be set in the following script:

Demo_Parameters.py

Inventory

https://github.com/Peeples-Lab/HLTDNN 

└── root directory
    ├── demo.py                           // Run this. Main demo file.
    ├── Demo_Parameters.py                // Parameter file for the demo.
    ├── View_Results.py                   // Run this after the demo to view saved results. 
    └── Datasets                
        ├── Get_Preprocessed_Data.py      // Resample the audio data and generate segments for the dataset.
        └── DeepShipDataModules.py.py     // Create a custom dataset for DeepShip.
    └── Utils                     
        ├── Class_information.py          // Class names for dataset for the results script.
        ├── Compute_FDR.py                // Compute Fisher’s discriminant ratio.
        ├── Compute_EHD.py                // Compute edge responses. 
        ├── Confusion_mats.py             // Create and plot confusion matrix.
        ├── Generate_Learning_Curves.py   // Generate learning curves for training and validation.
        ├── Generate_TSNE_visual.py       // Generate TSNE visuals.
        ├── Generate_Spatial_Dims.py      // Compute spatial dimensions for histogram layer models.
        ├── Get_Optimizer.py              // Define a set of optimizers to choose from.
        ├── Histogram_Model.py            // Load histogram model with TDNN (or any backbone network).
        ├── Network_functions.py          // Contains functions to initialize, train, and test the model.
        ├── RBFHistogramPooling.py        // Create histogram layer.
        ├── Save_Results.py               // Save results from the demo script.
        ├── pytorchtools.py               // Implement early stopping to terminate training based on validation metrics.
        └── Feature_Extraction_Layer.py   // Extract and transform different features from the audio files.
        └── TDNN.py                       // Baseline TDNN model.
        └── DTIEM_Model_RBF.py            // Statistical module.
        └── EDM.py                        // Structural module.
        └── EMD_loss.py                   // Compute Earth Mover Distance loss.
        └── Lightning_Wrapper.py          // simplifies training by organizing code and automating tasks like logging and checkpoints
        └── Loss_function.py              // Define all the loss functions.
        └── PANN_models.py                // Define pretrained audio neural networks.
        └── SSTKAD_v2.py                  // Define SSATKD model.

License

This source code is licensed under the license found in the LICENSE file in the root directory of this source tree.

This product is Copyright (c) 2025. Jarin Ritu, Amirmohammad Mohammadi, Davelle Carreiro, Alexandra Van Dine, Joshua Peeples. All rights reserved.

Citing SSATKD

If you use the SSATKD code, please cite the following reference using the following entry.

Plain Text:

Jarin Ritu, Amirmohammad Mohammadi, Davelle Carreiro, Alexandra Van Dine, Joshua Peeples,"Structural and Statistical Audio Texture Knowledge Distillation (SSATKD) for Passive Sonar Classification", 2025 arXiv preprint,doi:10.48550/arXiv.2501.01921.

BibTex:

@article{ritu2025structural,
  title={Structural and Statistical Audio Texture Knowledge Distillation (SSATKD) for Passive Sonar Classification},
  author={Ritu, Jarin and Mohammadi, Amirmohammad and Carreiro, Davelle and Van Dine, Alexandra and Peeples, Joshua},
  journal={arXiv preprint arXiv:2501.01921},
  year={2025}
}