## Usage of callbacks
A callback is a set of functions to be applied at given stages of the training procedure. You can use callbacks to get a view on internal states and statistics of the model during training. You can pass a list of callbacks (as the keyword argument `callbacks`) to the `.fit()` method of the `Sequential` or `Model` classes. The relevant methods of the callbacks will then be called at each stage of the training.
—
{{autogenerated}}
—
# Create a callback
You can create a custom callback by extending the base class `keras.callbacks.Callback`. A callback has access to its associated model through the class property `self.model`.
Here’s a simple example saving a list of losses over each batch during training:
“`python
class LossHistory(keras.callbacks.Callback):
def on_train_begin(self, logs={}):
self.losses = []
def on_batch_end(self, batch, logs={}):
self.losses.append(logs.get(‘loss’))
“`
—
### Example: recording loss history
“`python
class LossHistory(keras.callbacks.Callback):
def on_train_begin(self, logs={}):
self.losses = []
def on_batch_end(self, batch, logs={}):
self.losses.append(logs.get(‘loss’))
model = Sequential()
model.add(Dense(10, input_dim=784, kernel_initializer=’uniform’))
model.add(Activation(‘softmax’))
model.compile(loss=’categorical_crossentropy’, optimizer=’rmsprop’)
history = LossHistory()
model.fit(x_train, y_train, batch_size=128, epochs=20, verbose=0, callbacks=[history])
print(history.losses)
# outputs
”’
[0.66047596406559383, 0.3547245744908703, …, 0.25953155204159617, 0.25901699725311789]
”’
“`
—
### Example: model checkpoints
“`python
from keras.callbacks import ModelCheckpoint
model = Sequential()
model.add(Dense(10, input_dim=784, kernel_initializer=’uniform’))
model.add(Activation(‘softmax’))
model.compile(loss=’categorical_crossentropy’, optimizer=’rmsprop’)
”’
saves the model weights after each epoch if the validation loss decreased
”’
checkpointer = ModelCheckpoint(filepath=’/tmp/weights.hdf5′, verbose=1, save_best_only=True)
model.fit(x_train, y_train, batch_size=128, epochs=20, verbose=0, validation_data=(X_test, Y_test), callbacks=[checkpointer])
“`