Closed bdzyubak closed 1 year ago
The issue is twofold. The existing file I was using did not have 'epoch_time' as it was from a version of training before the callback. This will now throw a warning saying that a train_fresh=True is required to generate a valid training history for comparison, and epoch time won't be written. The issue of appending to an existing column was patched by integrating with CSVLogger more organically. Now, the custom callback will not open and write to a CSV on its own. It will simply update the callback logs which are used by CSVLogger. This appends to old files and introduces the initial condition of epoch 0 (training start parameters) organically. 3ec13ef067ed15ccb332f09f91b7f2b63dd87d38
When using train_fresh=False (default), the epoch timing callback cannot save history to a csv log causing a crash. It needs to append, not insert the epoch time column.