# Possible NetTrain[] memory leak when training in GPU

I've just moved to Mathematica 12.0 (from 11.3) and sometimes NetTrain[] will fail with the following error:

NetTrain::gpumemex: Computation aborted: GPU memory exhausted.


This never happens the first time I call NetTrain[], which makes me believe that there is some sort of memory leak. Is there a way to free up the memory of GPU, without quitting the kernel? (Because then I have to run all my preprocessing code again). I tried to call Remove[] on the trained network, but that didn't make any difference.

This did not happen with 11.3. Sorry that I don't have a MWE at the moment.