Feeding 256x256 images as arrays for high GPU utilization

Open sraut-scellbio opened this issue 4 months ago • 2 comments

Feeding a list of 256x256 grayscale images to .eval ignores the batch_size and defaults to a specific number. I am unable to increase GPU memory utilization by increasing batch_size at the moment. Currently I am attempting to feed my images as an array (512 x 3 x 256 x 256) but it forces me to specify channel_axis(which I set to be 1 in this case), z_axis (I don't have a z axis since these aren't z-stacks or confocal slices, regardless I set it to be 1) and I set do_3D=True. This seems to be taking even longer than feeding a list of images. Is these a better way to feed same sized images as an array (not from z-stacks and confocal) to CellposeSAM and increase GPU utilization?

Aug 19 '25 22:08 sraut-scellbio

Hi @sraut-scellbio We will implement this shortly. It was previously implemented in CP3, but lost in the transition to CP4.

Thanks

Aug 27 '25 20:08 mrariden

Hi, When will the new release with improved GPU utilization be available?

Sep 22 '25 09:09 PiyushGaik