search

muzairkhattak / multimodal-prompt-learning

[CVPR 2023] Official repository of paper titled "MaPLe: Multi-modal Prompt Learning".
https://muzairkhattak.github.io/multimodal-prompt-learning/
MIT License
619 stars 43 forks source link

Why batch_size is setted to 4? #16

Closed EdmunddzzZ closed 1 year ago

EdmunddzzZ commented 1 year ago

When train the model in a RTX 3090，found that the memory-usage was about 11G, so why not make it larger? The dataset is ImageNet

muzairkhattak commented 1 year ago

Hi @EdmunddzzZ,

Thank you for showing interest in MaPLe.

In our initial experiments, we found batch size of 4 to provide stable results and thus we fixed it. However, BS is an hyper-parameter and it can be further tuned for more optimal GPU utilization and performance.

Feel free to adjust the BS according to your needs. Just make sure you also adjust/scale the learning rate accordingly.

Kindly let us know incase you have any other query.

Thank you!

EdmunddzzZ commented 1 year ago

Thank you for patiently answering