search

MedARC-AI / fMRI-reconstruction-NSD

fMRI-to-image reconstruction on the NSD dataset.
MIT License
280 stars 39 forks source link

change batch size based on number of devices if using multi-gpu #43

Closed song-wensong closed 4 months ago

song-wensong commented 4 months ago

https://github.com/MedARC-AI/fMRI-reconstruction-NSD/blob/main/src/Train_MindEye.py#L189 Hello, why did you comment batch_size *= accelerator.num_processes? If I use multi-GPU, don't I need batch_size *= accelerator.num_processes?

PaulScotti commented 4 months ago

I'm not sure, we never got multi-gpu working better than single-gpu so can't offer advice on multi-gpu implementations