ESC Meta - Environmental Sound Classification with KAN

A PyTorch-based implementation for environmental sound classification using Kolmogorov-Arnold Networks (KAN) and traditional CNN architectures on the FSC22 dataset.

Installation

1. Clone the repository:

git clone https://github.com/DilshanRakshitha/ESC_Meta.git
cd ESC_Meta

2. Install dependencies:

pip install -r requirements.txt

3. Preparee a data folder in the root directory, extract data in place wav44 folder and FSC22.csv file in the follwing way.. Then run then install requirements. Then extract features. Then tran the model.

ESC_Meta/
├── data/
│   └── fsc22/
│       ├── FSC22.csv          # Dataset metadata
│       ├── wav44/             # Audio files (.wav)
│    
└── ...

Usage

1. Feature Generation

Generate mel spectrogram features from audio files:

# Basic usage (uses default settings)
python generate_features.py

# With custom paths (relative to project root - recommended)
python generate_features.py --csv-path data/fsc22/FSC22.csv --audio-path data/fsc22/wav44 --output data/fsc22/Pickle_Files

# Different feature types
python generate_features.py --feature-type MEL    # Mel spectrograms (default)
python generate_features.py --feature-type MFCC   # MFCC features
python generate_features.py --feature-type MIX    # Mixed features

# Augmentation levels
python generate_features.py --augmentation 0   # No augmentation
python generate_features.py --augmentation 1   # Time stretch only
python generate_features.py --augmentation 2   # Pitch shift only
python generate_features.py --augmentation 3   # Time stretch + pitch shift (default)
python generate_features.py --augmentation 4   # All augmentations

# Use preset configuration
python generate_features.py --preset aug_ts_ps_mel_features_5_20

Available Arguments:

• --csv-path, -c: Path to FSC22.csv file (default: data/fsc22/FSC22.csv)
• --audio-path, -a: Path to audio files directory (default: data/fsc22/wav44)
• --output, -o: Output directory for features (default: data/fsc22/Pickle_Files)
• --feature-type, -f: Feature type - MEL, MFCC, MIX (default: MEL)
• --augmentation, -aug: Augmentation level 0-4 (default: 3)
• --preset, -p: Use preset configuration (aug_ts_ps_mel_features_5_20)

2. Model Training

Train models using cross-validation:

# Train different models
python main.py --model alexnet
python main.py --model kan  
python main.py --model ickan
python main.py --model rapid_kan_lite

Available Models:

CNN Architectures:

• alexnet - AlexNet CNN
• densenet - DenseNet-121
• efficientnet - EfficientNet-V2 B0
• inception - Inception-V3
• mobilenet - MobileNet-V3 Small
• mobilenetv3large - MobileNet-V3 Large
• resnet50 - ResNet-50 v2
• resnet18 - ResNet-18

KAN-Inspired Architectures (Fast, CNN-based):

• kan_inspired - High-performance KAN-inspired model
• ickan_inspired - ICKAN-inspired model
• wavkan_inspired - Wavelet KAN-inspired model

KAN Architectures:

• kan - Exact KAN implementation
• kan_fast - Fast KAN (balanced mode)
• kan_memory_safe - Memory-efficient KAN

ICKAN Architectures:

• ickan - Standard ICKAN
• ickan_light - Lightweight ICKAN
• ickan_deep - Deep ICKAN

Rapid KAN Architectures:

• rapid_kan - Efficient Rapid KAN
• rapid_kan_lite - Lightweight Rapid KAN
• rapid_kan_power - Powerful Rapid KAN

3. Hyperparameter Optimization

Hyperparameter optimization for KAN models is not still working. Optimize model hyperparameters using Optuna:

# Quick optimization (5 trials, 10 epochs)
python optimize_models.py --model alexnet --config quick

# Standard optimization (20 trials, 30 epochs)
python optimize_models.py --model kan --config standard

# Extensive optimization (100 trials, 50 epochs)
python optimize_models.py --model ickan --config extensive

# Optimize all available models
python optimize_models.py --model all --config quick

Configuration Options:

• quick - Fast testing with few trials
• standard - Balanced optimization
• extensive - Thorough hyperparameter search