Fine-tuning question

[jsadjasjdjas]

Thanks for the great work.
I want to fine-tune the model on my own dataset (such as anime). Could you please tell me how to handle it? Should the model be fine-tuned with lora or only the last few layers of parameters be trained? It would be great if there were relevant fine-tuning tutorials.
Looking forward to your reply!

[hkchengrex]

Hello, we haven't experimented with model fine-tuning. I think you can try full model fine-tuning or LoRA -- fine-tuning only the last few layers is unusual for these setups.

[hboxgames]

If you manage to do Lora fine tuning, please let me know how it goes.. It's very interesting

[Zane0227]

If you manage to do Lora fine tuning, please let me know how it goes.. It's very interesting

I tried training LoRA with different layer combinations, hoping to improve MMAudio's capabilities in certain specific scenarios, but ultimately, every attempt led to overfitting and failure. Finally, I successfully trained the desired result on thinksound. I'm not entirely sure of the specific reasons, but thinksound seems to have a similar structure to MMAudio.

[mz2sj]

If you manage to do Lora fine tuning, please let me know how it goes.. It's very interesting

I tried training LoRA with different layer combinations, hoping to improve MMAudio's capabilities in certain specific scenarios, but ultimately, every attempt led to overfitting and failure. Finally, I successfully trained the desired result on thinksound. I'm not entirely sure of the specific reasons, but thinksound seems to have a similar structure to MMAudio.

Can you share your Thinksound lora training script.It will be great!