View source on GitHub |

## Class `TFParser`

Callable taking subset of TFF AST constructs to CompiledComputations.

When this function is applied via `transformation_utils.transform_postorder`

to a TFF AST node satisfying its assumptions, the tree under this node will
be reduced to a single instance of
`building_blocks.CompiledComputation`

representing the same
logic.

Notice that this function is designed to be applied to what is essentially
a subtree of a larger TFF AST; once the processing on a single device has
been aligned at the AST level, and placement separated from the logic of
this processing, we should be left with a function wrapped via
`federated_map`

or `federated_apply`

to a federated argument. It is this
function which we need to reduce to TensorFlow, and it is to the root
node of this function which we are looking to apply `TFParser`

. Because of
this, we assume that there is a lambda expression at the top of the AST
we are looking to parse, as well as the rest of the assumptions below.

- All called lambdas have been converted to blocks.
- All blocks have been inlined; that is, there are no block/LET constructs remaining.
- All compiled computations are called.
- No compiled computations have been partially called; we believe this should be handled correctly today but we haven't reasoned explicitly about this possibility.
- The only leaf nodes present under
`comp`

are compiled computations and references to the argument of the top-level lambda which we are hoping to replace with a compiled computation. Further, every leaf node which is a reference has as its parent a`building_blocks.Call`

, whose associated function is a TF graph. This prevents us from needing to deal with arbitrary nesting of references and TF graphs, and significantly clarifies the reasoning. This can be accomplished by "decorating" the appropriate leaves with called identity TF graphs, the construction of which is provided by a utility module. - There is only a single lambda binding any references present in the AST,
and it is placed at the root of the AST to which we apply
`TFParser`

. - There are no intrinsics present in the AST.

`__init__`

```
__init__()
```

Populates the parser library with mutually exclusive options.

## Methods

`__call__`

```
__call__(comp)
```

Transforms `comp`

by checking all elements of the parser library.

This function is roughly performing intermediate-code generation, taking
TFF and generating TF. Calling this function is essentially checking the
stack and selecting a semantic action based on its contents, and *only one*
of these actions should be selected for a given computation.

Notice that since the parser library contains mutually exclusive options, it is safe to return early.

#### Args:

: The`comp`

`building_blocks.ComputationBuildingBlock`

to check for possibility of reduction according to the parsing library.

#### Returns:

A tuple whose first element is a possibly transformed version of `comp`

,
and whose second is a Boolean indicating whether or not `comp`

was
transformed. This is in conforming to the conventions of
`transformation_utils.transform_postorder`

.