/hft

Pytorch implementation of automatic music transcription method that uses a two-level hierarchical frequency-time Transformer architecture (hFT-Transformer).

Primary LanguagePythonMIT LicenseMIT

hFT-Transformer

This repository contains the official PyTorch implementation of "Automatic Piano Transcription with Hierarchical Frequency-Time Transformer" presented in ISMIR2023 (arXiv 2307.04305).

Development Environment

  • OS
    • Ubuntu 18.04
  • memory
    • 32GB
  • GPU
    • corpus generation, evaluation
      • NVIDIA GeForce RTX 2080 Ti
    • training
      • NVIDIA A100
  • Python
    • 3.6.9
  • Required Python libraries

Usage

  1. corpus generation (MAESTRO-V3)
$ ./corpus/EXE-CORPUS-MAESTRO.sh
  1. training
$ ./training/EXE-TRAINING-MAESTRO.sh
  1. evaluation

If you want to avoid training models from scratch, you can download and put the model under the checkpoint/MAESTRO-V3 directory.

model_016_003.pkl is the model for MAESTRO.

$ wget https://github.com/sony/hFT-Transformer/releases/download/ismir2023/checkpoint.zip
$ unzip checkpoint.zip
$ ./evaluation/EXE-EVALUATION-MAESTRO.sh model_016_003.pkl test

If you want to evaluate the trained model using the validation set, you can change the second argument as below.

$ ./evaluation/EXE-EVALUATION-MAESTRO.sh model_016_003.pkl valid

Citation

Keisuke Toyama, Taketo Akama, Yukara Ikemiya, Yuhta Takida, Wei-Hsiang Liao, and Yuki Mitsufuji, "Automatic Piano Transcription with Hierarchical Frequency-Time Transformer," in Proceedings of the 24th International Society for Music Information Retrieval Conference, 2023.

@inproceedings{toyama2023,
    author={Keisuke Toyama and Taketo Akama and Yukara Ikemiya and Yuhta Takida and Wei-Hsiang Liao and Yuki Mitsufuji},
    title={Automatic Piano Transcription with Hierarchical Frequency-Time Transformer},
    booktitle={Proceedings of the 24th International Society for Music Information Retrieval Conference},
    year={2023}
}

Contact

Reference