So far I have only had marginal success with the TensorFlow-directml python library, however, it seems to be awful and maybe has a memory issue (small model and data yet trying to allocate 5-19GB into GPU RAM or RAM). I have tried to use ROCm, but This has led me to constant deadends even with older versions and docker methods.
At this point, I have no other ideas as to what I can do to get these programs to interact with my AMD RX580, and purchasing a new GPU is not an option.
Stop wasting time, use colab and kaggle for testing or run whatever you can on 8gb card anyways. If you need to run for more than 24 hrs, use paperspace pro subscription
I would sell the card and buy s 1080Ti to offload all the burden.
Maybe first try on a small subset of the data just to see if it actually works? I.e. avoid mem restrictrion