This fixes Issue #24, where setting the learning rate decay would crash because it requires a value for total_training_samples; this value, however, is never set before starting training and having the input data set split and parsed.
The fix removes the requirement for that value in set_learning_rate_decay by moving the calculation for lr_decay_epochs into the private learning rate setter (__set_learning_rate), which gets called when starting training but after constructing the graph and splitting & parsing the input data set.
The test suite was also updated to check that the functions can run successfully for good inputs and that calling the private learning rate setter treats the existence of decay settings properly.
This fixes Issue #24, where setting the learning rate decay would crash because it requires a value for
total_training_samples; this value, however, is never set before starting training and having the input data set split and parsed.The fix removes the requirement for that value in
set_learning_rate_decayby moving the calculation forlr_decay_epochsinto the private learning rate setter (__set_learning_rate), which gets called when starting training but after constructing the graph and splitting & parsing the input data set.The test suite was also updated to check that the functions can run successfully for good inputs and that calling the private learning rate setter treats the existence of decay settings properly.