| license: mit | |
| ### Enviroment | |
| ``` | |
| tar -xzf piano_transcription.tar.gz -C piano_transcription | |
| source piano_transcription/bin/activate | |
| ``` | |
| ### Data | |
| Download the MAESTRO dataset and place it in a dataset folder within the piano_transcription directory. | |
| Mozart is already included. | |
| ### Train | |
| Modify the Config File： | |
| 1. Paths: Update root, ckpt_path, and configs_path to absolute paths on your system (e.g., replace /home/zheqid/workspace/ with your local directory). | |
| 2. Codec Vocab Size: Set vocab_size to match your audio codec (e.g., 65536 for trancodec_fsq). | |
| 3. Model Size: Adjust n_layer, n_head, and n_embd to scale the model (e.g., increase for a larger model). | |
| 4. Hardware: Modify device and batch_size_per_device based on your GPU setup. | |
| Then train as follow : | |
| ``` | |
| python train.py --config /home/zheqid/workspace/musictokenizer/configs/mozart.yaml | |
| ``` | |