musicnn-pytorch

MusicNN-PyTorch

This is a PyTorch reimplementation of the MusicNN library for music audio tagging.

It contains the model architecture and converted weights from the original TensorFlow 1.x checkpoints.

Supported Models

• MTT_musicnn: Trained on MagnaTagATune (50 tags) - Default model
• MSD_musicnn: Trained on Million Song Dataset (50 tags)
• MSD_musicnn_big: Larger version trained on MSD (512 filters)

Super Simple Usage (Hugging Face Transformers)

from transformers import AutoModel

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

# Use the model
tags = model.predict_tags("your_audio.mp3", top_k=5)
print(f"Top 5 tags: {tags}")

Embeddings (Optional)

from transformers import AutoModel

model = AutoModel.from_pretrained("oriyonay/musicnn-pytorch", trust_remote_code=True)

# Extract embeddings from any layer
emb = model.extract_embeddings("your_audio.mp3", layer="penultimate", pool="mean")
print(emb.shape)

Colab Example

# Install dependencies
!pip install transformers torch librosa soundfile

# Load with AutoModel
from transformers import AutoModel
model = AutoModel.from_pretrained("oriyonay/musicnn-pytorch", trust_remote_code=True)

# Use the model
tags = model.predict_tags("your_audio.mp3", top_k=5)
print(tags)

Traditional Usage

If you prefer to download the code manually:

from musicnn_torch import top_tags

# Get top 5 tags for an audio file
tags = top_tags('path/to/audio.mp3', model='MTT_musicnn', topN=5)
print(tags)

Installation

pip install transformers torch librosa soundfile

Credits

Original implementation by Jordi Pons. PyTorch port by Gemini.