jellis18 / learning-rate

Implementation of Learning Rate Finder, SGDR and Cyclical Learning Rate in Keras

Learning Rate

Keras Callbacks to :

• Find the optimal Learning Rate
• Use Stochastic Gradient Descent with Restart
• Use Cyclical Learning Rate

Learning Rate Finder

Usage

lr_finder = LRFinder(min_lr=1e-5, max_lr=1e-2, steps_per_epoch=np.ceil(X_train.shape[0]/batch_size), epochs=1)

model.fit(X_train, y_train, callbacks=[lr_finder] )

# Plot the raw losses (they can be noisy)
lr_finder.plot_loss()

# Plot the smoothed losses
lr_finder.plot_avg_loss()

Stochastic Gradient Descent with Restart

schedule = SGDRScheduler(min_lr=1e-4,max_lr=1e-3,steps_per_epoch=np.ceil(X_train.shape[0]/batch_size), lr_decay=0.9, cycle_length=3,mult_factor=1.5)

model.fit(X_train, y_train, epochs=10, batch_size=128, validation_data=(X_test, y_test), callbacks=[schedule])

Cyclical Learning Rate

clr_triangular = CyclicLR(mode='triangular', step_size=np.ceil(X_train.shape[0]/batch_size))

model.fit(X_train, y_train, batch_size=128, epochs=10, validation_data=(X_test, y_test), callbacks=[clr_triangular])

References

This code is based on:

• The method described in the 2015 paper "Cyclical Learning Rates for Training Neural Networks" by Leslie N. Smith
• The implementation of the algorithm in fastai library by Jeremy Howard.
• This implementation of CLR