I am currently using pytorch-lightning to perform some deep learning tasks. The code is currently working fine. However, I wanted to perform some pre-training tasks (mostly vizualisation). I implemented theses tasks in the on_fit_start hook. Inside this hook, the val_dataloaders attribute from trainer is always None. Later on in the script (for instance in the on_fit_end hook, I call it exactly the same way, and it actually yields my dataloader). Is the dataloader not available at this stage of the fit ? I could not find anything ressembling elsewhere. When I commented the on_fit_start method, everything worked well even though the validation dataloader is still called the same way. I could refactor my code to do the preprocessing elsewhere, but it would not make sense from a logical point of view... Any idead what might be wrong ? Thanks in advance! Continue reading...
