Iteration -1 simplified midi scraping script (removed the use of selenium which requires a lot of setup and hence can add complexity in containerisation and multi-node processing) Implemented async scrapping - the scraping was multifold faster Added new midi data - from a different website - tripled the data size
Iteration-2 - MLOps implementation <= currently here
Use Airflow for data pipeline TFX for training, error analysis, and serving using dockers, ec2/gcp, kubeflow.
Iteration-3 - Improve the Model
Use various model architectures and hyperopt
1. Transformers
2. ...
Experiemnt with W&B, MLFlow, TFX to monitor, track and store experiments. To understand the pros and cons of each.
Based on Linan Chen's work Github https://github.com/linanpy/Music-Generation Using deep learning models to generate jazz music!