MPS Graph Neural Network Training Produces NaN Loss on Xcode 15.0 beta 8 + iOS 17.0

Hello,

I've been working on an app that involves training a neural network model on the iPhone. I've been using the Metal Performance Shaders Graph (MPS Graph) for this purpose. In the training process the loss becomes Nan on iOS17 (21A329).

I noticed that the official sample code for Training a Neural Network using MPS Graph (link) works perfectly fine on Xcode 14.3.1 with iOS 16.6.1. However, when I run the same code on Xcode 15.0 beta 8 with iOS 17.0 (21A329), the training process produces a NaN loss in function updateProgressCubeAndLoss. The official sample code and my own app exhibit the same issue.

Has anyone else experienced this issue? Is this a known bug, or is there something specific that needs to be adjusted for iOS 17?

Any guidance would be greatly appreciated.

Thank you!

Replies

I found that the issue disappear on iOS 17.1