tf.types.experimental.PolymorphicFunction

Base class for polymorphic graph functions.

Inherits From: Callable

Graph functions are Python callable objects that dispatch calls to a TensorFlow graph. Polymorphic graph functions can be backed by multiple TF graphs, and automatically select the appropriate specialization based on the type of input they were called with. They may also create specializations on the fly if necessary, for example by tracing.

Also see tf.function.

function_type Returns a FunctionType describing this callable.

Methods

experimental_get_compiler_ir

View source

Returns compiler IR for the compiled function.

This API is intended only for debugging as there are no guarantees on backwards compatibility of returned IR or the allowed values of stage.

Args
*args compilation args supports inputs either: (1) all inputs are TensorSpec or (2) all inputs are tf.Tensor/Python variables.
**kwargs Keyword arguments used for compilation. Same requirement as compiliation args.

Returns
Function callable with the following kwargs:

  • stage at which the compiler IR should be serialized. Allowed values are:
    • hlo: HLO output after conversion from TF (https://www.tensorflow.org/xla/operation_semantics).
    • hlo_serialized: Like stage=hlo, but the output is a serialized HLO module proto (a bytes object).
    • optimized_hlo: HLO after compiler optimizations.
    • optimized_hlo_serialized: Like stage=optimized_hlo, but the output is a serialized HLO module proto (a bytes object).
    • optimized_hlo_dot: optimized HLO in DOT format suitable for Graphviz.
  • device_name can be either None, in which case the preferred device is used for compilation, or a device name. It can be a full device name, or a partial one, e.g., /device:CPU:0.

For example, for

@tf.function(jit_compile=True)
def f(x):
  return x + 1

f.experimental_get_compiler_ir(tf.random.normal([10, 10])(stage='hlo')

the output is:

HloModule a_inference_f_13__.9

ENTRY %a_inference_f_13__.9 (arg0.1: f32[10,10]) -> f32[10,10] {
  %arg0.1 = f32[10,10]{1,0} parameter(0), parameter_replication={false}
  %reshape.2 = f32[10,10]{1,0} reshape(f32[10,10]{1,0} %arg0.1)
  %constant.3 = f32[] constant(1)
  %broadcast.4 = f32[10,10]{1,0} broadcast(f32[] %constant.3)
  %add.5 = f32[10,10]{1,0} add(f32[10,10]{1,0} %reshape.2,
                               f32[10,10]{1,0} %broadcast.4)
  %reshape.6 = f32[10,10]{1,0} reshape(f32[10,10]{1,0} %add.5)
  %tuple.7 = (f32[10,10]{1,0}) tuple(f32[10,10]{1,0} %reshape.6)
  ROOT %get-tuple-element.8 = f32[10,10]{1,0}
    get-tuple-element((f32[10,10]{1,0}) %tuple.7), index=0
}

Here is another example using tf.TensorSpec inputs:

y = tf.Variable(tf.zeros([10, 20], dtype=tf.float32))

@tf.function(jit_compile=True)
def f(x):
  return x + y

hlo_str = f.experimental_get_compiler_ir(tf.TensorSpec(shape=(10,
20)))(stage='hlo')

The output is:

HloModule a_inference_f_120__.8,
entry_computation_layout={(f32[10,20]{1,0},f32[10,20]{1,0})->f32[10,20]{1,0} }

ENTRY %a_inference_f_120__.8 (arg0.1: f32[10,20], arg1.2: f32[10,20]) ->
f32[10,20] {
  %arg0.1 = f32[10,20]{1,0} parameter(0), parameter_replication={false},
  metadata={op_name="XLA_Args"}
  %reshape.3 = f32[10,20]{1,0} reshape(f32[10,20]{1,0} %arg0.1)
  %arg1.2 = f32[10,20]{1,0} parameter(1), parameter_replication={false},
  metadata={op_name="XLA_Args"}
  %add.4 = f32[10,20]{1,0} add(f32[10,20]{1,0} %reshape.3, f32[10,20]{1,0}
  %arg1.2), metadata={op_type="AddV2" op_name="add"
  source_file="<ipython-input-16-ea04879c1873>" source_line=4}
  %reshape.5 = f32[10,20]{1,0} reshape(f32[10,20]{1,0} %add.4),
  metadata={op_name="XLA_Retvals"}
  %tuple.6 = (f32[10,20]{1,0}) tuple(f32[10,20]{1,0} %reshape.5),
  metadata={op_name="XLA_Retvals"}
  ROOT %get-tuple-element.7 = f32[10,20]{1,0}
  get-tuple-element((f32[10,20]{1,0}) %tuple.6), index=0,
  metadata={op_name="XLA_Retvals"}
}
</td>
</tr>

</table>


The HLO module accepts a flat list of inputs. To retrieve the order of these inputs signatures, users can call the concrete_fn.structured_input_signature and concrete_fn.captured_inputs:

# Use concrete_fn to get the hlo_module flat_args.
concrete_fn = f.get_concrete_function(tf.TensorSpec(shape=(10, 20)))
flat_args = list(
    tf.nest.flatten(concrete_fn.structured_input_signature)
    ) + concrete_fn.captured_inputs

Raises
ValueError (1) If an invalid stage is selected (2) or if applied to a function which is not compiled (jit_compile=True is not set). (3) or if input shapes are not fully defined for tf.TensorSpec inputs
TypeError When called with input in graph mode.

get_concrete_function

View source

Returns a ConcreteFunction specialized to input types.

The arguments specified by args and kwargs follow normal function call rules. The returned ConcreteFunction has the same set of positional and keyword arguments as self, but their types are compatible to the types specified by args and kwargs (though not neccessarily equal).

@tf.function
def f(x):
  return x
f_concrete = f.get_concrete_function(tf.constant(1.0))
f_concrete = f.get_concrete_function(x=tf.constant(1.0))

Unlike normal calls, get_concrete_function allow type specifiers instead of TensorFlow objects, so for example tf.Tensors may be replaced with tf.TensorSpecs.

@tf.function
def f(x):
  return x
f_concrete = f.get_concrete_function(tf.TensorSpec([], tf.float64))

If the function definition allows only one specialization, args and kwargs may be omitted altogether.

@tf.function(input_signature=[tf.TensorSpec(None, tf.float32)])
def f(x):
  return x
f_concrete = f.get_concrete_function()

The returned ConcreteFunction can be called normally:

f_concrete(tf.constant(1.0))
<tf.Tensor: shape=(), dtype=float32, numpy=1.0>
f_concrete(x=tf.constant(1.0))
<tf.Tensor: shape=(), dtype=float32, numpy=1.0>

Args
*args inputs to specialize on.
**kwargs inputs to specialize on.

Returns
A ConcreteFunction.

__call__

View source

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.