ModelPart is now a syntactic sugar class for these three abstract classes:
Parameterized - manages variables, name scopes, loading and saving
Feedable - has feed_dict, batch_size, and train_mode (in the future, may also contain some logic for creating placeholders or handling data in tf.dataset)
GenericModelPart - the get_dependencies function is now here and all the classes that handle something in the graph before runners and/or trainers are descendants of this class
as a bonus, this PR fixes the dependenies gradient blocking objects and provides a way of how to do
it e.g. for adversarial gradient layers introduced in #771
Read the commit messages.
ModelPart is now a syntactic sugar class for these three abstract classes:
Parameterized
- manages variables, name scopes, loading and savingFeedable
- has feed_dict, batch_size, and train_mode (in the future, may also contain some logic for creating placeholders or handling data in tf.dataset)GenericModelPart
- the get_dependencies function is now here and all the classes that handle something in the graph before runners and/or trainers are descendants of this classas a bonus, this PR fixes the dependenies gradient blocking objects and provides a way of how to do it e.g. for adversarial gradient layers introduced in #771