callbackaun 0.0.2

Callbackaun: Custom Callbacks that will control training

This notebook contains a custom callback you may wish to copy and use

It is a combination of the Keras callbacks Reduce Learning Rate on Plateau, Early Stopping and Model Checkpoint but eliminates some of the limitations of each. In addition, it provides an easier to read summary of the model's performance at the end of each epoch. It also provides a handy feature that enables you to set the number of epochs to train for until a message asks if you wish to halt training on the current epoch by entering H or to enter an integer which will determine how many more epochs to run before the message appears again. This is very useful if you are training a model and decide the metrics are satisfactory and you want to end the model training early. Note the callback always returns your model with the weights set to those of the epoch which had the highest performance on the metric being monitored (accuracy or validation accuracy) The callback initially monitors training accuracy and will adjust the learning rate based on that until the accuracy reaches a user specified threshold level. Once that level of training accuracy is achieved the callback switches to monitoring validation loss and adjusts the learning rate based on that. the callback is of the form: callbacks=[LRA(model, base_model, patience, stop_patience, threshold,factor, nspace, batches, initial_epoch, epochs, ask_epoch )] where:

• model is your compiled model

• base_model is the name of your base_model if you are doing transfer learning. for example you might have in your model base_model=tf.keras.applications.EfficientNetB1(include_top=False, weights="imagenet",input_shape=img_shape, pooling='max') base_model.trainabel=False During training you will be asked if you want to do fine tuning If you enter F to the query, the base_model will be set to trainable by the callback If you are not doing transfer learning set base_model==None

• patience is an integer that determines many consecutive epochs can occur before the learning rate will be adjusted (similar to patience parameter in Reduce Learning Rate on Plateau)

• stop_patience is an integer that determines hom many consecutive epochs for which the learning rate was adjusted but no improvement in the monitored metric occurred before training is halted(similar to patience parameter in early stopping)

• threshold is a float that determines the level that training accuracy must achieve before the callback switches over to monitoring validation loss. This is useful for cases where the validation loss in early epochs tends to vary widely and can cause unwanted behavior when using the conventional Keras callbacks

• factor is a float that determines the new learning rate by the equation lr=lr*factor. (similar to the factor parameter in Reduce Learning Rate on Plateau)

• nspace is a boolean. It is used in the callback as part of an experiment on training models. If on a given epoch the metric being monitored fails to improve it means your model has moved to a location on the surface of Nspace (where N is the number of trainable parameters) that is NOT as favorable (poorer metric performance) than the position in Nspace you were in for the previous epoch. If dwell is set to True the callback loads the model with the weights from the previous (better metric value) epoch. Why move to a worse place if the place you were in previously was better. Then the learning rate is reduced for the next epoch of training. If dwell is set to false this action does not take place.

• batches is an integer. It should be set to a value of batches=int(number of traing samples/batch_size). During training the callback provides information during an epoch of the form 'processing batch of batches accuracy= accuracy loss= loss where batch is the current batch being processs, batches is as described above, accuracy is the current training accuracy and loss is the current loss. Typically the message would appear as processing batch 25 of 50 accuracy: 54% loss: .04567. As each batch is processed these values are changed.

• initial_epoch is an integer. Typically set this to zero Itis used in the information printed out for each epoch. In the case where you train the model say with the basemodel weights frozen say you train for 10 epochs. Then you want to fine tune the model and train for more eppochs for the second training session you would reinstantiate the callback and set initial_epoch=10.

• epochs an integer value for the number of epochs to train

• ask_epoch is either set to an integer value or None. If set to an integer it denotes the epoch number at which user input is requested. If the user enter H training is halted. If the user inters an integer it represents how many more epochs to run before you are asked for the user input again. If the user enters F the base_model is made trainable If ask_epoch is set to None the user is NOT asked to provide any input. This feature is handy is when training your model and the metrics are either unsatisfactory and you want to stop training, or for the case where your metrics are satisfactory and there is no need to train any further. Note you model is always set to the weights for the epoch that had the beset metric performance. So if you halt the training you can still use the model for predictions. ** Example of Use: callbacks=[LRA(model=my_model, base_model=base_model, patience=1,stop_patience=3, threshold=.9, factor=.5, dwell=True,batches=85, initial_epoch=0 , epochs=20, ask_epoch=5)] this implies:

• your model is my_model

• base_model is the name of your base_model if you are doing transfer learning

• after 1 epoch of no improvement the learning rate will be reduced

• after 3 consecutive adjustment of the learning rate with no metric improve training terminates

• once the training accuracy reaches 90% the callback adjust learning rate based on validation loss

• when the learning rate is adjusted the new learning rate is .5 X learning rate

• if the current epoch's metric value did not improve, the weights for the prior epoch are loaded and the learning rate is reduced

• 85 batches of data are run to complete an epoch

• the initial epoch is 0

• train for 20 epochs

• after the fifth epoch you will be asked if you want to halt training by entering H or enter an integer denoting how many more epochs to run before you will be prompted again or enter F to make the base_model=trainable