Autodist¶

User Interface.

set_default_autodist(o)[source]¶: Set the AutoDist object the scope of which you are in.

get_default_autodist()[source]¶: Get the AutoDist object the scope of which you are in.

class AutoDist(*args, **kwargs)[source]¶

Bases: autodist.autodist._V1Graph, autodist.autodist._V2Graph, autodist.autodist._V2Eager

AutoDist is a scalable ML engine.

AutoDist provides user-friendly interfaces to distribute local deep-learning model training across multiple processing units with scalability and minimal code changes.

Parameters

resource_spec_file (str) – file path of a resource specification yaml
strategy_builder (base.StrategyBuilder) – (optional) a strategy builder object

scope()[source]¶

Create a context manager capturing the code block to be distributed.

Yields: AutoDist context

build_strategy()[source]¶

Build distributed strategy based on the default graph in the scope.

Returns: Distributed strategy representation object.
Return type: base.Strategy

create_distributed_session()[source]¶

Create a Session object to execute the default graph in a distributed manner.

Returns: A wrapped TensorFlow Session object.
Return type: WrappedSession

function(fn)[source]¶: Experimental interface similar to tf.function.

is_built()[source]¶

Whether the distributed graph is built for the most recent original graph.

Returns: True if the distributed graph is built by AutoDist
Return type: bool