Custom TFlite and Raven formats output very different results in MacOS M3 Max Chip

[nnbuainain]

Hey guys,

I trained a custom linear classifier for 154 species. When I make predictions over the same file I get very different results when using the tflite and Raven model formats on MacOS with M3 Max chip. When I move to linux, the same model and sound files outputs different results from the mac output, but identical output between tflite and Raven formats, which is the expected results. So I guess the problem is not related to any quantization and compressions done in tflite, but something related to making inferences in MacOS (?). Is there any known problem with the models and the Apple silicon chips? I read that Raven does not support the classifier models in the new Apple chips, is that the reason?

I'm in a conda environment using python3.10 and tensorflow 2.15.1 (I was not able to run train or inference with 2.16.1).

[hooge104]

Same here. M3 Max with python3.10 and tensorflow 2.15.1. I've not been able to figure out why this is happening either