lakh_jazz_transfer_model

Lakh MIDI Model with Jazz Transfer Learning

Model Overview

This repository contains a model retrained on the Lakh MIDI Dataset with additional transfer learning applied to the Jazz MIDI Dataset. The base model was trained from scratch on lakh MIDI and fine-tuned using a curated dataset of jazz MIDI segments. These segments consist of 20-second excerpts taken from the middle of full-length jazz compositions, focusing on the most musically interesting and dynamic portions.

Training Details

1. Base Training:

   • Dataset: lakh MIDI Dataset (cleaned and filtered).
   • Training: The model was trained from scratch on this dataset to learn general musical structures and styles.

2. Transfer Learning:

   • Dataset: A subset of 549 jazz MIDI segments.
   • Epochs: 2 epochs.
   • Purpose: Fine-tuning the model to specialize in generating and understanding jazz-specific musical patterns, leveraging the most intricate and dynamic parts of jazz compositions.

This two-step approach ensures that the model retains its general understanding of MIDI data while being optimized for jazz music-specific tasks.

Files in the Repository

The repository includes the following files:

1. config.json:

   • Contains the configuration of the model architecture. This includes details such as the number of layers, hidden dimensions, attention heads, and other parameters used to define the model.

2. generation_config.json:

   • Contains generation-specific settings, such as maximum sequence length, temperature, top-k, and top-p sampling parameters. These configurations are crucial for controlling the behavior of the MIDI sequence generation process.

3. model.safetensors:

   • The model weights saved in the safetensors format for efficient and secure loading. This format ensures safe deserialization of model weights.

4. training_args.bin:

   • Stores the training arguments and hyperparameters used during both base training and transfer learning. This file can be useful for reproducing the training setup or understanding the specifics of the training process.

Dataset Details

Lakh MIDI Dataset

• Focus: General MIDI compositions across various genres.
• Cleaning Process: Removed duplicates, ensured proper formatting, and filtered out noisy data.

Jazz MIDI Dataset (Transfer Learning)

• Focus: Jazz genre-specific MIDI compositions.
• Size: 549 compositions (20-second segments from the middle of each composition).
• Epochs: 2 epochs.
• Purpose: Fine-tuning the model for improved jazz music generation, emphasizing the most dynamic and musically rich portions of jazz pieces.

Usage

This model is suitable for:

• General MIDI music generation.
• Specialized jazz music generation tasks.
• Experimentation with transfer learning techniques in music AI.

Original Code Base

The original model and architecture are based on the repository AI-Guru/MMM-JSB. The base training and transfer learning were performed to adapt this architecture for diverse and genre-specific tasks.

License

This model follows the licensing terms of the original repository. Please review the license for more details.