psklight / keras_one_cycle_clr

Keras callbacks for one-cycle training, cyclic learning rate (CLR) training, and learning rate range test.

One Cycle & Cyclic Learning Rate for Keras

This module provides Keras callbacks to implement in training the following:

• One cycle policy (OCP)
• Cyclic learning rate (CLR)
• Learning rate range test (LrRT)

(Documentation at https://psklight.github.io/keras_one_cycle_clr/)

Highlights

• Learning rate & Weight decay range test.
• Using callbacks, the module works for datasets of numpy arrays or data generator.
• Common usage as callbacks for both model.fit and model.fit_generator where epochs is intuitively interpreted as cycle lengths.

In detail:

This is inspired by how well fastai library implements this for PyTorch. By the time this module was made, a few options to implement these learning policies in Keras have two limitations: (1) They might not work with data generator; (2) They might need a different way to train (rather than passing a policy as a callback). This module addresses both limitation by defining these training policies as Keras callbacks in such a way that both model.fit and model.fit_generator can be called. For OPC, the number of epochs (argument for fitting) directly represents a cycle length. For LrRT and CLR, epochs necessary to complete a training with a particular policy can be calculated from the policy callback's .find_n_epoch.

Additionally, the utils submodule defines some useful functions such as:

• plot_from_history plots train and validation loss (if any) as a function of epochs.
• concatenate_history concatenates training and validation losses and metrics from a list of keras.callbacks.History which can be obtained from model.history after training. This is helpful in connecting histories from multiple one-cycle policy trainings.

Dependencies:

• tensorflow
• (optional) keras
• matplotlib, numpy, pandas, tqdm
• (optional) numba

Example of LrRT

lrrt_cb = clr.LrRangeTest(lr_range=(1e-3, 1),
                          wd_list=[0, 1e-4, 1e-3, 1e-2, 1e-1], # grid test for weight decay
                          steps=steps,
                          batches_per_step=b,
                          validation_data=(x_test, y_test), # good to find weight decay
                          batches_per_val=5,
                          threshold_multiplier=5.,
                          verbose=False)

n_epoch = lrrt_cb.find_n_epoch(train_gen)
# n_epoch = lrrt_cb.find_n_epoch(x_train, batch_size) # for numpy array as train set
        
model.fit_generator(generator=train_gen,
                       epochs=n_epoch,
                       verbose=0,
                       callbacks=[lrrt_cb])

lrrt_cb.plot()

Example of OCP

ocp_1_cb = clr.OneCycle(lr_range=(0.01, 0.1),
                     momentum_range=(0.95, 0.85),
                     reset_on_train_begin=True,
                     record_frq=10)

ocp_cb.test_run(1000)  # plot out values of learning rate and momentum as a function of iteration (batch). 1000 is just for plotting. The actual iteration will be computed when model.fit or model.fit_generator is run.

# setting ``epochs`` to 20 means a cycle length is 20 epochs.
hist1 = model_kr.fit_generator(generator=train_gen,
                      epochs=20,
                      validation_data=val_gen,
                      callbacks=[ocp_1_cb, checkpoint, tensorboard_cb],
                      verbose=2)

# train for another cycle
ocp_2_cb = clr.OneCycle(lr_range=(0.001, 0.01),
                     momentum_range=(0.95, 0.85),
                     reset_on_train_begin=True,
                     record_frq=10)

hist2 = model_kr.fit_generator(generator=train_gen,
                      epochs=20,
                      validation_data=val_gen,
                      callbacks=[ocp_2_cb, checkpoint, tensorboard_cb],
                      verbose=2)

hist_all = concatenate_history([hist1, hist2], reindex_epoch=True)

plot_from_history(hist_all) # plot train and validation losses versus epochs

Example of CLR

clr_cb = ktool.CLR(cyc=3,
                   lr_range=(1e-2/5, 1e-2),
                   momentum_range=(0.95, 0.85),
                   verbose=False,
                   amplitude_fn=lambda x: np.power(1.0/3, x))

clr_cb.test_run(600) # see that a new cycle starts at 0th, 200th, and 400th iteration.

clr_hist = model.fit(x_train, y_train,
                     epochs=60,
                     validation_data=(x_test, y_test),
                     verbose=2,
                     callbacks=[clr_cb])

plot_from_history(clr_hist)