Closed rom1504 closed 2 years ago
https://www.tensorflow.org/guide/data#consuming_python_generators probably would be faster with this (Already 100k sample/s but could be much faster, it's not using resources well)
This is WIP. I'm not even sure I really want to put this here, I need to think some more about it
works better, but still memleak
I think it's due to tf though and not embedding reader
something like https://stackoverflow.com/questions/64199384/tf-keras-model-predict-results-in-memory-leak
works now and no memleak
computing safety predictions on top of clip embeddings