This PR fixes several issues which we ran into when testing out model in our last working group meeting (#56). They were partly listed in this comment.
Changes
Decrease RAM to 4GB
Decrease #CPU cores to 2
Disable shuffling for validation data
Rename batch_size to physical_batch_size
Make the replacement conditional, depending on is_training
Fix iterations bug, which led to very large training epochs which were in fact virtual_batch_size_factor many epochs.
Fix learning_rate to 1 (learning adjustment with alpha, beta, and gamma).
Extend logging significantly
Validation
Started a couple of validation runs for testing purposes.
This PR fixes several issues which we ran into when testing out model in our last working group meeting (#56). They were partly listed in this comment.
Changes
batch_size
tophysical_batch_size
is_training
virtual_batch_size_factor
many epochs.learning_rate
to1
(learning adjustment withalpha
,beta
, andgamma
).Validation
Started a couple of validation runs for testing purposes.