Mistral_MidiTok_Transformer_Single_Instrument_Small
This model is trained from scratch using tokenized midi music.
I have trained a MidiTok tokeniser (REMI) and its made by spliting multi-track midi into a single track.
We then trained in on a small dataset.
Its using the Mistral model that has been cut down quite a bit.
What else needs to be done
Update model training to use small positional embeddings for the model 1024 + a padding amount like 8
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 30
- eval_batch_size: 30
- seed: 444
- gradient_accumulation_steps: 3
- total_train_batch_size: 90
- optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: cosine_with_restarts
- lr_scheduler_warmup_ratio: 0.3
- training_steps: 20000
Framework versions
- Transformers 4.46.2
- Pytorch 2.1.0+cu121
- Datasets 3.1.0
- Tokenizers 0.20.3
