FunctionGraph

TODO How do we initialize a FunctionGraph?

TODO What does a FunctionGraph do that a variable given by a TensorVariable does not?

TODO What are Features?

TODO What functionalities of FunctionGraphs are needed where?

Need to draw a map of where the different FunctionGraph functionalities could be used.

Misc

From the documentation:

A `FunctionGraph` represents a subgraph bound by a set of input variables and a set of output variables, ie a subgraph that specifies an Aesara function. The inputs list should contain all the inputs on which the outputs depend. `Variable`\s of type `Constant` are not counted as inputs.

The `FunctionGraph` supports the replace operation which allows to replace a variable in the subgraph by another, e.g. replace ``(x + x).out`` by ``(2 * x).out``. This is the basis for optimization in Aesara.

This class is also responsible for verifying that a graph is valid (ie, all the dtypes and broadcast patterns are compatible with the way the `Variable`\s are used) and for tracking the `Variable`\s with a :attr:`FunctionGraph.clients` ``dict`` that specifies which `Apply` nodes use the `Variable`. The :attr:`FunctionGraph.clients` field, combined with the :attr:`Variable.owner` and each :attr:`Apply.inputs`, allows the graph to be traversed in both directions.

It can also be extended with new features using :meth:`FunctionGraph.attachfeature`. See `Feature` for event types and documentation. Extra features allow the `FunctionGraph` to verify new properties of a graph as it is optimized.

What I am wondering is: when is a FunctionGraph needed? When does it become cumbersome?

import aesara
from aesara.graph.fg import FunctionGraph
import aesara.tensor as at

x = at.vector('x')
z = at.log(at.exp(x + x)) + x + x

fg = FunctionGraph(outputs=[z])

aesara.dprint(fg)
print(fg.clients)

TODO What is the clients dictionary used for?

It looks like this is what allows to traverse the graph?

TODO Keep track of where a FunctionGraph is necessary, where it is produced, where it is cloned, etc.

Links to this note