cle.scale() returns a clipped image (in z axis)

[wbrett87]

When I try to use cle.scale() as a drop-in replacement for skimage.transform.rescale(), I get a clipped image, in which the tops and bottom of my objects (nuclei) are cut off. skimage.transform.rescale() works as expected, and outputs an image which has a z dimension appropriately scaled. The image returned from cle.scale() is of the same shape as the input image.

This is my line of code :

rescaled = cle.scale(input_image,factor_x=1,factor_y=1,factor_z=4.398)

[wbrett87]

It seems that cutting my (132,2048,2048) image into smaller blocks fixed the problem... Must be running into some sort of vram shortage?

[StRigaud]

Hi @wbrett87

Yes, this behaviour is expected by default. Working in GPU requires to allocate the output memory in advance so, by default cle.scale() allocate the same size as the input.

You would need to set the resize argument to true so that the output dimensions is pre-compute based on the transformation you want to do

rescaled = cle.scale(input_image,factor_x=1,factor_y=1,factor_z=4.398, resize=True)

More info on the function parameter in the API reference

An alternative, if the output dimensions is still wrong, is to provide yourself the output array with the correct shape. You have an example in the 3D segmentation demos notebook

A quick warning though, by rescaling, you might increase your memory footprint and you device (GPU) will indeed need to have enough RAM to be able to process the data.

Hope that solve the issue!

[wbrett87]

Thanks @StRigaud!

I did try resize=True before I posted, but it didn't seem to change anything. I'll attempt to provide the output array next (thanks for the link to the notebook!). Indeed, it might be a good sign that my GPU is not big enough and maybe I should just use smaller images.

[StRigaud]

Can you do a cle.info() and post your devices info?

Usually, lack of VRAM lead to an OUT_OF_RESSOURCES errors, not an incomplete output, but this might change with hardware. In you case, you would need a minimum of 24Gb VRAM device.