Deep Learning with Keras

By LiterallyTheOne

7: Callbacks

Introduction

• Previous Tutorial: Preprocessing and Data Augmentation
• This Tutorial: Callbacks

Callbacks

• A function
• Pass to Fit function
• Gets called in a specific moment

TensorBoard

from keras.callbacks import TensorBoard

log_dir = "logs/fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = TensorBoard(log_dir=log_dir)

history = model.fit(
    ...,
    callbacks=[tensorboard_callback],
)

EarlyStopping

• Stops the training
• No improvement

EarlyStopping Example

from keras.callbacks import EarlyStopping

early_stopping = EarlyStopping(
    monitor="val_loss",
    patience=5,
    restore_best_weights=True,
    verbose=1,
)

history = model.fit(
    ...,
    epochs=200,
    callbacks=[tensorboard_callback, early_stopping],
)

ModelCheckpoint

• Saves the model
• During training
• Helps us
  • Not to loose the model
  • Have the previous weights to continue from them

ModelCheckpoint Example

from keras.callbacks import ModelCheckpoint

model_checkpoint = ModelCheckpoint(
    filepath="checkpoints/best_model.weights.h5",
    monitor="val_loss",
    save_best_only=True,
    save_weights_only=True,
    verbose=1,
)

history = model.fit(
    ...,
    callbacks=[tensorboard_callback, early_stopping, model_checkpoint],
)

Save only Weights

• If save_weights_only=True:
  • file name should end with: .weights.h5
• Helps us:
  • Capacity
  • Multi-platform

Load the best weights

model.load_weights("checkpoints/best_model.weights.h5")

Link to the tutorial and materials