cat-alan

Cat Audio Emotion Classification using M5 Neural Network

Classification of domestic cat audio files into 10 emotional states: Angry, Defense, Fighting, Happy, HuntingMind, Mating, MotherCall, Paining, Resting, and Warning.

The model trains on raw waveforms using the M5 architecture (PyTorch), with time stretching, pitch shifting, and Gaussian noise augmentation.

---

🎯 Recent Updates (2026)

This repository has been updated with new training and inference scripts that work with modern Python modules (PyTorch 2.9+, Python 3.13+) while maintaining compatibility with the original M5 model architecture.

New Scripts:

1. train_modern.py - Modernized Training Script

Based on the original script.py, this updated version:

• ✅ Uses the same M5 architecture as the original model
• ✅ Works with PyTorch 2.9+ and Python 3.13+ (no need for old versions!)
• ✅ Matches original training approach: Same hyperparameters (lr=0.01, weight_decay, LR scheduler), same augmentations (TimeStretch, PitchShift, Gaussian noise), same data pipeline
• ✅ Uses Kaggle data to train(https://www.kaggle.com/datasets/yagtapandeya/cat-sound-classification-dataset)

What you can do:

# Train a new model using the fixed pipeline with modern PyTorch
python train_modern.py --data data/CAT_DB --epochs 30

2. inference.py - Universal Inference Script

A new script that can work with both old and new models:

• ✅ Loads the original pretrained model (MLflow format from examples/model/)
• ✅ Also loads any new models trained with train_modern.py (.pt/.pth files)
• ✅ Works with modern PyTorch 2.9+ (no downgrade needed!)
• ✅ Auto-detects model architecture from checkpoint
• ✅ Two modes:
  • Evaluation on dataset (calculate accuracy)
  • Prediction on single audio file

What you can do:

# Evaluate original pretrained model with modern PyTorch
python inference.py --checkpoint examples/model --data data/CAT_DB --mode eval

# Predict emotion from a single cat audio file
python inference.py --checkpoint examples/model --audio cat_meow.mp3 --mode predict

# Also works with newly trained models
python inference.py --checkpoint my_model.pt --data data/CAT_DB --mode eval

Why These Updates Matter:

The original repository requires PyTorch 1.9 and Python 3.8 (from 2022). These new scripts let you:

• ✅ Use the latest PyTorch and Python versions
• ✅ Train new models with the same proven M5 architecture
• ✅ Load and use the original pretrained model without downgrading
• ✅ Fix bugs that caused training failures

---

📋 Requirements

• Python 3.8+
• PyTorch 2.0+
• See requirements.txt for full dependencies

---

🎓 Training

Option 1: Use the New Modern Training Script (Recommended)

Train a new model with the fixed pipeline that works with modern PyTorch:

python train_modern.py --data data/CAT_DB --epochs 30 --batch_size 4 --lr 0.01

This script:

• Uses the same M5 architecture as the original model
• Fixes bugs that prevented proper training
• Works with PyTorch 2.9+ and Python 3.13+
• Follows the same training methodology (augmentations, hyperparameters)

Option 2: Use the Original Script

For reference, the original training script is still available:

python script.py

Note: The original script was designed for PyTorch 1.9 and may require older dependencies.

Training Parameters:

• --data: Path to dataset directory
• --epochs: Number of training epochs (default: 10)
• --batch_size: Batch size (default: 4)
• --lr: Learning rate (default: 0.01)
• --filters: Filter sizes for each conv layer (default: [64, 64, 128, 128])
• --early_stop_patience: Early stopping patience (default: 5)
• --checkpoint_dir: Directory to save checkpoints

Example Output:

Using device: mps
Train Epoch 0: 100%|██████████| 20/20 [00:08<00:00, loss=0.612, accuracy=0.047]
Test  Epoch 0: 100%|██████████| 5/5 [00:07<00:00, loss=0.485, accuracy=0.000]
...
Train Epoch 8: 100%|██████████| 20/20 [00:07<00:00, loss=0.406, accuracy=0.359]
Test  Epoch 8: 100%|██████████| 5/5 [00:06<00:00, loss=0.402, accuracy=0.250]

---

🏗️ Model Architecture

M5 Network:

• 4 convolutional blocks
• Each block: Conv1D → BatchNorm → ReLU → MaxPool → Dropout(0.3)
• Global average pooling
• Fully connected layer → LogSoftmax
• Input: Raw waveform at 8kHz
• Output: 10 emotion classes

Default Configuration:

• Filters: [64, 64, 128, 128]
• Kernel sizes: [80, 3, 3, 3]
• Strides: [16, 1, 1, 1]

---

📊 Model Performance

Real Performance on CAT_DB Dataset (100 samples, 10 per class):

We evaluated both the original pretrained model and a newly trained model on the same CAT_DB test set (20 samples):

Model	Overall Accuracy	Test Samples	Best Classes	Notes
Original Pretrained (examples/model/)	10.00%	2/20 correct	MotherCall (100%), Happy (50%)	Trained on different dataset from research paper
Newly Trained (train_modern.py)	35.00%	7/20 correct	HuntingMind (100%), Fighting (75%), Paining (50%)	Trained specifically on CAT_DB

Detailed Per-Class Accuracy:

Original Pretrained Model:

Angry          : 0.00%
Defense        : 0.00%
Fighting       : 0.00%
Happy          : 50.00%  ✓
HuntingMind    : 0.00%
Mating         : 0.00%
MotherCall     : 100.00% ✓
Paining        : 0.00%
Resting        : 0.00%
Warning        : 0.00%

Newly Trained Model (train_modern.py):

Angry          : 0.00%
Defense        : 0.00%
Fighting       : 75.00%  ✓
Happy          : 0.00%
HuntingMind    : 100.00% ✓
Mating         : 33.33%  ✓
MotherCall     : 0.00%
Paining        : 50.00%  ✓
Resting        : 0.00%
Warning        : 33.33%  ✓

Analysis:

Important: Both models use the SAME M5 architecture AND the SAME training pipeline!

The performance difference is NOT due to different model structures or different training methodology. Both use:

• ✅ Identical M5 architecture
• ✅ Same augmentations (TimeStretch, PitchShift, Gaussian noise)
• ✅ Same hyperparameters (lr=0.01, weight_decay=0.0001, LR scheduler)
• ✅ Same training approach (raw waveform input, 8kHz sampling)

Class-Specific Performance Patterns:

• Original model: Best at MotherCall (100%), Happy (50%)
• New model: Best at HuntingMind (100%), Fighting (75%), Paining (50%)
• Same architecture, same pipeline design, but different learned patterns
• Shows that training data distribution matters more than architecture

Next steps:

1. Collect More Data: Aim for 100-150 samples per class
2. Use less class: Considering this paper and employed less sound classes to train the model

Good to know:

This kaggle author employed 500 epochs to train CNN model using the same dataset and reach 45% accuracy. https://www.kaggle.com/code/muqaddasejaz/cat-emotion-classification-eda/notebook

How to Reproduce These Results:

# Evaluate original pretrained model
python inference.py --checkpoint examples/model --data data/CAT_DB --mode eval

# Evaluate newly trained model
python inference.py --checkpoint path/to/your_model.pt --data data/CAT_DB --mode eval

---

📚 Citation

Credit to the original dataset and augmentation techniques:

Domestic Cat Sound Classification Using Transfer Learning Yagya Raj Pandeya, Dongwhoon Kim and Joonwhoan Lee https://doi.org/10.3390/app8101949