tfp.experimental.auto_batching.TensorFlowBackend

Implements the TF backend ops for a PC auto-batching VM.

safety_checks Python bool indicating whether we should use runtime assertions to detect stack overflow/underflow.
while_parallel_iterations Python int, the argument to pass along to tf.while_loop(..., parallel_iterations=while_parallel_iterations)
while_maximum_iterations Python int or None, the argument to pass along to tf.while_loop(..., maximum_iterations=while_maximum_iterations)
basic_block_xla_device Python str indicating the device to which basic blocks should be targeted (i.e. 'CPU:0' or 'GPU:0'); if not None.

variable_class

Methods

any

View source

assert_matching_dtype

View source

Asserts that the dtype of value matches expected_dtype.

Args
expected_dtype A numpy dtype
value Tensor or convertible.
message Optional diagnostic message.

Raises
ValueError If dtype does not match.

batch_size

View source

Returns the first (batch) dimension of value.

broadcast_to_shape_of

View source

Broadcasts val to the shape of target.

Attempts to match the dtype of broadcast_val to the dtype of target, if val is not a Tensor and target has a dtype.

Args
val The value to be broadcast. Must be broadcast-compatible with target.
target Tensor whose shape we will broadcast val to match.
name Optional name for the op.

Returns
broadcast_val A Tensor with shape matching val + target. Provided that val's dimension sizes are all smaller or equal to target's, the returned value will be the shape of target.

cond

View source

Implements a conditional operation for the backend.

Args
pred A boolean scalar Tensor indicating the condition.
true_fn A callable accepting and returning nests of Tensors having the same structure as state, to be executed when pred is True.
false_fn A callable accepting and returning nests of Tensors having the same structure as state, to be executed when pred is False.
name Optional name for the op.

Returns
state Output state, matching nest structure of input argument state.

create_variable

View source

Returns an intialized Variable.

Args
name Name for the variable.
alloc VariableAllocation for the variable.
type_ instructions.TensorType describing the sub-batch shape and dtype of the variable being created.
max_stack_depth Scalar int Tensor, the maximum stack depth allocated.
batch_size Scalar int Tensor, the number of parallel threads being executed.

Returns
var A new, initialized Variable object.

equal

View source

Implements equality comparison for TF backend.

fill

View source

Fill a fresh batched Tensor of the given shape and dtype with value.

Args
value Scalar to fill with.
size Scalar int Tensor specifying the number of VM threads.
dtype tf.DType of the zeros to be returned.
shape Rank 1 int Tensor, the per-thread value shape.
name Optional name for the op.

Returns
result Tensor of dtype values with shape [size, *shape]

full_mask

View source

Returns an all-True mask Tensor with shape [size].

merge_dtypes

View source

Merges two dtypes, returning a compatible dtype.

In practice, TF implementation asserts that the two dtypes are identical.

Args
dt1 A numpy dtype, or None.
dt2 A numpy dtype, or None.

Returns
dtype The common numpy dtype.

Raises
ValueError If dt1 and dt2 are not equal and both are non-None.

merge_shapes

View source

Merges two shapes, returning a broadcasted shape.

Args
s1 A list of Python int or None.
s2 A list of Python int or None.

Returns
shape A list of Python int or None.

Raises
ValueError If s1 and s2 are not broadcast compatible.

not_equal

View source

Implements inequality comparison for TF backend.

prepare_for_cond

View source

Backend hook for preparing Tensors for cond.

The TensorFlow backend uses this hook to apply tf.convert_to_tensor before entering the cond tree generated by virtual_machine._staged_apply. One could do this inside cond, but when this API element was defined there seemed to be a performance reason (for Eager mode) to do it once per cond tree rather than once per cond.

Args
state A state to be prepared for use in conditionals.

Returns
state The prepared state.

reduce_min

View source

Implements reduce_min for TF backend.

run_on_dummies

View source

Runs the given primitive_callable with dummy input.

This is useful for examining the outputs for the purpose of type inference.

Args
primitive_callable A python callable.
input_types list of instructions.Type type of each argument to the callable. Note that the contained TensorType objects must match the dimensions with which the primitive is to be invoked at runtime, even though type inference conventionally does not store the batch dimension in the TensorTypes.

Returns
outputs pattern of backend-specific objects whose types may be analyzed by the caller with type_of.

static_value

View source

Gets the eager/immediate value of t, or None if t is a Tensor.

switch_case

View source

Implements a switch (branch_selector) { case ... } construct.

type_of

View source

Returns the instructions.Type of t.

Args
t tf.Tensor or a Python or numpy constant.
dtype_hint dtype to prefer, if t is a constant.

Returns
vm_type instructions.TensorType describing t.

where

View source

Implements a where selector for the TF backend.

Attempts to match the dtypes of the value operands, if they are not yet both Tensors.

Args
condition A boolean Tensor, either a vector having length (x + y).shape[0] or matching the full shape of x + y.
x Tensor of values to take when condition is True. Shape must match that of y.
y Tensor of values to take when condition is False. Shape must match that of x.
name Optional name for the op.

Returns
masked A broadcast-shaped Tensor where elements corresponding to True values of condition come from x, and others come from y.

while_loop

View source

Implements while loops for TF backend.

wrap_straightline_callable

View source

Method exists solely to be stubbed, i.e. for defun + XLA compile.