training often crashes due to Out of Memory (OOM) errors, this seems to be a fairly novel problem after I moved some functions into classes in a utils.py script and made slight changes to the archtiecture, suddenly things stopped working (I think it may even be a TF library update) I have an atypical experimental setup where a new instance of the model is made for a full training and prediction, and the non-gpu ram steadily increases despite not needing past versions of the object