Uint16 intensity images

[echon (Francisco José Tornay Mejías)]

Any ideas about how to train a resnet/resnext network with intensity images whose value ranges exceed 8 bits?

[machinethink (Matthijs)]

Haven’t tried it myself, but usually the first step when training is to normalize the images, which turns the 8-bits pixel values into floating point numbers. I don’t see why you can’t do the same thing with 16-bits pixel values. And as soon as you have floating point numbers, the neural network can be trained on those.

[echon (Francisco José Tornay Mejías)]

Yes, this is what I’m trying. I wanted to make sure that it’s the right approach

Thank you for the reply

[Pcamellon (Pedro Pablo Camellón Quintero)]

Hi !

@machinethink

do the same thing with 16-bits pixel values. And as soon as you have floating point numbers, the neural network can be trained on those

Does fastai do this for me?