Thanks for the tip! I'll have to try MLX out later. Right now I need some results. I'm taking a course and a full week behind because all the data and models that I generated in the previous week is now suspect.

@Xcode-K After reading a bunch of git issues, it looks like the problem resides in relu activation and possibly softmax. Apparently tanh and softplus are possibly not impacted. I managed to squeeze out some good results by using those. I don't know enough to verify this so I'm just abandoning running on the apple metal for the moment. I've got a second hand GPU and stuffed it in a linux box. I'll just work on that for the time being.

@Xcode-K I tried 2.18 TensorFlow with the metal plugin and got different bad results. fit() just coughed up a bunch of NaN for loss and the accuracy fell to almost zero after about 5 epochs. I also have a project requirement of Keras 2. I can probably work around that, but the uncertainty of the results leaves me very suspicious. Is there something about 2.16 that fixes whatever is broken in the other versions?