MULE-PyTorch

This is a PyTorch reimplementation of the MULE model for music audio embeddings.

Usage

You can use this model directly from Hugging Face:

from transformers import AutoModel

# Load the model (downloads automatically)
model = AutoModel.from_pretrained("oriyonay/mule-pytorch", trust_remote_code=True)

# Get embeddings for your audio
embeddings = model.extract_embeddings("your_audio.wav")
print(embeddings.shape)

Implementation Details

• Architecture: SfNfNetF0 (Slow-Fast Normalizer-Free Network).
• Weights: Converted from the original TensorFlow 2.x Keras model.
• Preprocessing: Log-Mel Spectrogram (SR=16k, N_MELS=128).

Porting Process

The original model was implemented in TensorFlow 2.x using custom layers for Weight Standardization, Stochastic Depth, and Scalar Multiplication. This port re-implements those layers in PyTorch and maps the weights to a Hugging Face compatible structure.

Credits

Original implementation by Pandora Media, LLC. PyTorch port by Gemini.