View source on GitHub |
Base class for differentiable graph functions.
Inherits From: Callable
A ConcreteFunction
encapsulates the original graph function definition with
support for differentiability under tf.GradientTape
contexts. In the
process, it may generate new graph functions (using the original) to
efficiently perform forwards and backwards passes.
Attributes | |
---|---|
function_type
|
Returns a FunctionType describing this callable. |
inference_fn
|
Returns the original AtomicFunction owned by this ConcreteFunction.
|
Methods
__call__
__call__(
*args, **kwargs
)
Executes this callable.
This behaves like a regular op - in eager mode, it immediately starts
execution, returning results. In graph mode, it creates ops which return
symbolic TensorFlow values (like tf.Tensor
, tf.data.Dataset
,
etc.). For example, tf.function
callables typically generate a
tf.raw_ops.PartitionedCall
op, but not always - the
exact operations being generated are an internal implementation detail.
Args | |
---|---|
*args
|
positional argument for this call |
**kwargs
|
keyword arguments for this call |
Returns | |
---|---|
The execution results. |