Day 9 - Using Callbacks in Keras

• A list of callbacks can be passed to the fit() method.

• Keras calls callbacks at:
  • start and end of training
  • start and end of an epoch
  • before and after processing each batch

ModelCheckpoint callback

• Saves checkpoints of model at regular intervals during training.

• If a validation set is used during training, save_best_only=True will save weights only if performance on validation set is best so far.

• No need to worry about training too long and overfitting on training set. Restore last model after training. This will be best model on training set.

EarlyStopping callback

• Training is stopped when there is no progress on validation set for a given number of epochs (defined by the patience argument).

• Optionally, can roll back to the best model.

• Set number of epochs to large value since training will stop automatically when there is no progress.

Custom callbacks

• Extend the keras.callbacks.Callback class.

• For callback to be used during training, implement methods: on_train_begin(), on_train_end(), on_epoch_begin(), on_epoch_end(), on_batch_begin() and on_batch_end().

• For callbacks to be used during evaluation, implement methods: on_test_begin(), on_test_end(), on_test_batch_begin(), and on_test_batch_end().

• For callbacks to be used during prediction, implement methods: on_predict_begin(), on_predict_end(), on_predict_batch_begin() and on_predict_batch_end().

• Using callbacks during evaluation and prediction can be useful for debugging.