resolved it: The problem was with the GPU for a weird reason. Once I got the latest driver for the GPU, it worked. I wasn't using GPU for NetTrain but for some reason it needed fixing.
@Szabolcs Yeah. Goal is to move this to the python C API then move the core JLink classes over for now.
I assume Todd Gayley has thought a decent amount about how best to write a Mathematica <-> Java link so I'm hoping to make use of that. And people know how to work with JLink so it'd hopefully lower some of the initial barriers to entry.
Anyone happen to know if ImageAugmentationLayer will do exactly the same augmentation to my input as to my target mask, or do I need to explicitly feed my target to it as well?
from the way my network is acting, it seems like it doesn't
seems like that layeris more useful for classification tasks rather than segmentation ones :'(
@psimeson I ran into interesting problems with NetTrain - and even just invoking NetChain (!) on my laptop. They crashed the kernel right away, even when everything was supposed to be running on the CPU. My OS installation was inherited from an earlier laptop which had a Nvidia GPU and a CUDA driver installed, and after removing that driver, everything worked again!
The curious thing was that nothing in the code was supposed to touch GPU, especially on NetChain phase. Still plain existence of this driver (probably as a loaded module) was enough to completely prevent usage of neural networks functionality.
@kirma I kinda believe that NetTrain is not that stable in MMA. In my case installing the updated Nvidia GPU driver solved the problem. In my code also, I never invoked any GPU computing but NetTrain was just failing.
@halirutan. Sorry to bother you, but am having problems with your editor buttons javascript. The buttons appear in the editor toolbar, the tooltips work, but clicking on any button produces no action. Any advice?
