hub.create_module_spec

View source on GitHub

Creates a ModuleSpec from a function that builds the module's graph.

hub.create_module_spec(
    module_fn, tags_and_args=None, drop_collections=None
)

The module_fn is called on a new graph (not the current one) to build the graph of the module and define its signatures via hub.add_signature(). Example:

# Define a text embedding module.
def my_text_module_fn():
  text_input = tf.placeholder(dtype=tf.string, shape=[None])
  embeddings = compute_embedding(text_input)
  hub.add_signature(inputs=text_input, outputs=embeddings)

See add_signature() for documentation on adding multiple input/output signatures.

Multiple graph variants can be defined by using the tags_and_args argument. For example, the code:

hub.create_module_spec(
    module_fn,
    tags_and_args=[({"train"}, {"is_training":True}),
                   (set(), {"is_training":False})])

calls module_fn twice, once as module_fn(is_training=True) and once as module_fn(is_training=False) to define the respective graph variants: for training with tags {"train"} and for inference with the empty set of tags. Using the empty set aligns the inference case with the default in Module.init().

THIS FUNCTION IS DEPRECATED.