I have recently noticed wierd behavior of Keras Tuner. The elapsed time is around 6 houres but the algorythm is running around 20 houres. Also I have noticed that there is really long time between trials.
I am using code bellow:
tuner_neurons = kt.BayesianOptimization(
model_builder,
max_trials = 1000,
num_initial_points = 500,
seed = 420,
objective='val_mean_absolute_percentage_error',
directory='hyperopt',
project_name='hyperopt_neurons_bayesian_1'
)
history_hyperopt = tuner_neurons.search(
X_train_dict[0],
y_train_dict[0],
batch_size = BATCH,
epochs = 1000,
validation_data = (X_valid_dict[0], y_valid_dict[0]),
verbose = 1,
steps_per_epoch=X_train_dict[0].shape[0] // BATCH,
callbacks=[learning_rate_reduction, early_stopping]
)
The cell bellow is running without any log from calculation:
Also the processor is on full speed so it is calculating something, just don't have any idea what. To find what parameters to find in BO should be pretty fast. Any ideas?