View source on GitHub
|
Abstract interface for data backends.
tff.framework.DataBackend()
A data backend is a component that can resolve symbolic references to data
as URIs, and locally materialize the associated payloads. Data backends are
used in tandem with the data_executor that queries them as it encounters
the data building block.
Methods
materialize
materialize(
data, type_spec
)
Materializes data with the given type_spec.
The form of the materialized payload must be such that it can be understood
by the downstream components of the executor stack. For example, if the data
backend is plugged into a stack based on an eager TensorFlow executor, the
accepted forms of payload would include Numpy-like objects, tensors, as well
as instances of eager tf.data.Dataset, and structures thereof. It is the
responsibility of the code that constructs the executor stack with the given
data backend to ensure that the types of payload materialized are compatible
with what the downstream components of the executor stack can accept.
| Args | |
|---|---|
data
|
A symbolic reference to the data to be materialized locally. Must be
an instance of pb.Data.
|
type_spec
|
An instance of computation_types.Type that represents the
type of the data payload being materialized.
|
| Returns | |
|---|---|
| The materialized payload. |