Spaces:
Running
on
T4
Good results
Dear skytnt,
First and foremost, I'd like to express my gratitude for the work you've done on the AI model. The quality and capabilities of the base model are truly commendable.
However, I've recently attempted to train and fine-tune the model on my own dataset of 904 mid items using the following command:
"python train.py --data midi_data/train --lr 0.001 --batch-size-train 16 --max-step 10000 --val-step 25"
Despite these efforts, I've found that the quality of the fine-tuned model isn't meeting my expectations. I believe I might be missing some crucial steps or insights that could enhance the model's performance post fine-tuning.
Would it be possible for you to provide some documentation or guidance on the fine-tuning process? Specifically, I'm interested in understanding any best practices, recommended dataset sizes, and potential pitfalls to avoid.
Your expertise and guidance would be immensely valuable to me as I continue to work with the model.
Thank you in advance for your time and assistance.
"python train.py --data midi_data/train --lr 0.001 --batch-size-train 16 --max-step 10000 --val-step 25"
Did you forget use --ckpt to laod pretrained model?
lr seems too high, try 1e-4 or 5e-5
Reduce --data-val-split to use as much data as possible for training
In fact, I'm not sure about the quality of the fine-tuned model.