Retrieval fine-tuning resources requirement

Open yfb-xieyu opened this issue 1 year ago • 1 comments

I'm very interested in your work. The chapter on retrieval fine-tuning provides the corresponding script at https://github.com/OpenGVLab/InternVL/tree/main/internvl_g, which uses 32 A100 GPUs. May I ask how long it would take to train on Flickr30K with 32 A100 GPUs, and whether it's possible to replicate this training if I only have 4 A100 GPUs?

Mar 27 '24 14:03 yfb-xieyu

Do you have the answer yet?

How many GPUs do you need for fine-tuning in your tests?

May 09 '24 07:05 lexilii

Hi, you can see this document for retrieval fine-tuning: https://internvl.readthedocs.io/en/latest/internvl1.0/internvl_g.html#retrieval-fine-tuning-head

The GPU memory requirement is shown here:

Jul 30 '24 13:07 czczup