Parameter-efficient Dual-encoder Architecture with Differentiable Choquet Integral Fusion for Underwater Acoustic Classification

Amirmohammad Mohammadi, Alexandra Van Dine and Joshua Peeples

Installation Prerequisites

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

pip install -r requirements.txt

A pre-configured environment with the essential packages already installed is available as well: environment.yml

Demo

Run demo.py in Python IDE or command line to train the models.

Inventory

└── root directory
    ├── demo.py                                // Entry point for training (Single/Dual modes)
    ├── datasets/
    │   ├── DeepShipDataModule.py              // Lightning DataModule for DeepShip audio dataset loader.
    │   └── ShipsEarDataModule.py              // Lightning DataModule for ShipsEar audio dataset loader.
    ├── src/
    │   ├── model_module.py                    // LightningModule wrapper around HF models (Single mode)
    │   ├── multiencoder_module.py             // Dual-branch LightningModule wrapper (Dual mode)
    │   ├── audio_augmentation.py              // Ship-specific and waveform augmentations (SpecAugment, SNR noise)
    │   ├── audio_filtering.py                 // Feature extraction transforms (Mel, MFCC, CQT, VQT) and normalization
    │   ├── hpt.py                             // HPT integration
    │   ├── RBFHistogramPooling.py             // HistogramLayer (RBF histogram)
    │   ├── adapters.py                        // Parallel Adapters and integration
    │   └── callbacks/
    │       ├── hpt_gate_logger.py             // Callback to log per-epoch HPT gate values.
    │       └── print_epoch_losses.py          // Callback to print epoch train/val loss summaries.
    └── utils/
        ├── metrics.py                         // aggregate_and_save: aggregates per-run metrics.
        ├── profiling.py                       // Standalone profilers (FLOPs, latency, throughput).
        └── run_evaluation_plots.py            // Evaluation plotting.

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 A. Mohammadi, A. Dine and J. Peeples. All rights reserved.

Citation

If you use this work, please cite:

arXiv preprint

@article{mohammadi2026parameter,
  title={Parameter-efficient Dual-encoder Architecture with Differentiable Choquet Integral Fusion for Underwater Acoustic Classification},
  author={Mohammadi, Amirmohammad and Peeples, Joshua and Van Dine, Alexandra},
  journal={arXiv preprint arXiv:2606.02341},
  year={2026}
}