Keras is an open-source neural network library written in Python. It is capable of running on top of TensorFlow, Microsoft Cognitive Toolkit, R, Theano, or PlaidML. One of the significant benefits of Keras is allowing developers to export models at specific epochs, which can be a great asset in model tuning and performance evaluation.
Problem
When training a machine learning model, we usually monitor the loss or metric function’s performance for some validation data. Different epochs can result in a different model performance. Sometimes, the best results achieve at an epoch that doesn’t correspond to the end of the training process. In such situations, it would come in handy if we could save/export the keras model at specific epochs.
Solution
The solution to this problem lies in Keras Callbacks. A Callback is an object (a class instance implementing specific methods) that is passed to the model in the call to fit and that is called by the model at various points during training. It has access to all the available data about the state of the model and its performance.
A custom callback function allows us to specify actions at various stages of training, such as at the beginning or end of an epoch, before or after a single batch, etc. One such action could be to save the model at specific epochs.
Step-by-step Explanation of the Code
First, we define a custom callback to save the model at specific epochs.
class CustomSaver(keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs={}):
if epoch == 9: # or save after some epoch, each k-th epoch etc.
self.model.save("model_{}.hd5".format(epoch))
Next, we add this callback to the model fitting process.
model = ... # create model model.compile(optimizer='...', loss='...') # compile model saver = CustomSaver() model.fit(..., callbacks=[saver]) # put your X_train, Y_train ...
In the code above, we create an instance of the CustomSaver class, and then pass it to the model’s fit method as part of the callbacks list.
By modifying the “if” condition within our ‘on_epoch_end’ method, we can make the callback save after each ‘k’ epochs, or whenever a particular condition is met.
Keras Libraries and Their Functions
Keras’s sleek, user-friendly design simplifies deep learning model creation and modification. It comes with several tools, like Sequential and Model classes for building models, various layers for neural networks (Convolutional, Pooling, Dense, etc.), and callbacks for monitoring training.
One of the most useful aspects of callbacks is stopping training early, or saving the best model according to validation performance. It’s such a frequently asked feature that Keras already contains built-in callbacks for them, known as ModelCheckpoint and EarlyStopping.
That’s it! Now apply this knowledge to your scenario and save the Keras model at the epoch where it offers the best performance. Happy model training!
