Excellent work, I would like to know what kind of machine configuration is needed if I need to fine tune it？

[tms2003]

Excellent work, I would like to know what kind of machine configuration is needed if I need to fine-tune it?
I do some of my own data collection on the system and the results are sometimes not satisfactory. How can I fine-tune it? Required hardware

[vimar-gu]

It will be dependent on the amount of your data. If you only have a limited amount of data, parameter-efficient fine-tuning might be optimal. As ViT-L/14 is still not a large model compared with LLMs, it should be feasible to run on any recent machine learning system (e.g., A6000). Oppositely, if you have a large amount of data that has a distribution gap with TreeOfLife-200M, it might be better to run full fine-tuning. In such cases, I would recommend that you use at least 4 A100/H100s for the fine-tuning.

Unfortunately, I don't have any experience fine-tuning BioCLIP/BioCLIP 2 with custom data. There might be some hyper-parameter tuning needed to find the best setting.

@egrace479 I remember we were once talking about fine-tuning BioCLIP. Do you have any information from the previous users?

[tms2003]

@vimar-gu How do I make a small dataset for fine-tuning? The full 200m dataset seems daunting. I can't find instructions for making a dataset, what should I do?