search

callsys / ControlCap

[ECCV 2024] ControlCap: Controllable Region-level Captioning
53 stars 0 forks source link

The minimal computational resources? #8

Open VoyageWang opened 2 months ago

VoyageWang commented 2 months ago

Hi, there!

Thanks for your nice work! I would like to know the minimal resources needed to train the overall pipeline of your model. I have 8 NVIDIA 3090 GPUs with 24GB, is it enough?

callsys commented 2 months ago

Hello! We've only done training on the 8 x A100 80G. ControlCap does not have much trainable parameters. By reducing the batch size and increases the gradient_accumulate_steps, it is possible to train ControlCap on 3090 24G.

VoyageWang commented 2 months ago

Hello! We've only done training on the 8 x A100 80G. ControlCap does not have much trainable parameters. By reducing the batch size and increases the gradient_accumulate_steps, it is possible to train ControlCap on 3090 24G.

Thank you for your quick reply. I want to know how to change the gradient_accumulate_steps in the configs or somewhere. I didn't find the specific parameters corresponding to this