i do not plan on publishing app.

Yes but I'd like to know what it means :)

you might enjoy some of my code here using dataset (currently under intense development) : https://github.com/ker2x/recolorizeNN/blob/master/main.py Caching, prefetching, lot of stuff here and there trying to get things faster.

Yes, this numpy version requirement can be a great source of troubles. Tensorflow won't work with a higher version. There was a post about it and apple answered that this was the version recommended by Tensorflow so they just followed the recommandation.

i have more now. I have some particularly bad result on some epoch, just like your. weird.

yes. tf.random is broken on 12.1

It's on a M1. don't use pip when dealing with Tensorflow as it's suppose to be installed using miniforge3

it's also worth adding that that simple/tiny model works faster on M1 cpu than gpu (even a discrete gpu). The google's TPU powerhouse is even more slower than anything on small model. The M1 CPU is surprisingly fast on small ML model.

you're welcome :)

models running faster on CPU than GPU is a very common occurence, it's not Mac specific. If there is a bug, identify it. saying that a cpu can be faster than gpu is not a bug, it's extremely common. It depend on the model, the cpu, the gpu, the input pipeline, etc ... I've also had the M1 CPU running faster than on a badass Nvidia Quadro or even a Nvidia P100. The M1 cpu is surprisingly good at this.

it works on 3.9 too (not sure if it worked on 3.9 when you posted this a month ago. But as of today, it works)

it doesn't works in a tf.function, it used to works, before 12.1, and a lot of exemple/guide/tutorial relying on random doesn't works anymore