TRAIN epoch_params - WHOIGit/ifcb_classifier GitHub Wiki

Epoch Parameters

These training flags adjust how many epochs or training iterations are performed during training.

Once training is complete, regardless of how many epochs are performed, the model with the best validation loss epoch (not the last-run epoch) is selected as the final output model. This is to prevent model over-fitting.

If estop is disabled, the training routine will train for emax epochs. Otherwise, estop is used to stop training early, the assumption being that any further training will only result in over-fitted models that will ultimately be rejected.

Sometimes, it takes a number of epochs for the model training to stabilize. emin ensures that a minimum number of epochs are attempted before estop is allowed to end training early.

Epoch Parameters:
  --emax MAX            Maximum number of training epochs. Default is 60.
  --emin MIN            Minimum number of training epochs. Default is 10.
  --estop STOP          Number of epochs following a best-epoch after-which to stop training. 
                        AKA Early Stopping. Set STOP=0 to disable. Default is 10.