ENH: Allow on_unused_input="ignore" for string keys in eval

[tvwenger]

Before

Passing on_unused_input="ignore" to eval when the inputs_to_values dictionary is indexed by the pytensor variables properly ignores any unused inputs since on_unused_input is passed along to pytensor.function. When the dictionary is indexed by strings, however, the helper function convert_string_keys_to_variables will raise a ValueError when a given input name is not found in the graph, regardless of the value of on_unused_input.

After

Both of these eval statements should work, but the latter raises a ValueError.

import pytensor.tensor as pt

x = pt.dscalar("x")
y = pt.dscalar("y")
z = x + 1.0

print(z.eval({x: 1.0, y: 0.0}, on_unused_input="warn"))

print(z.eval({"x": 1.0, "y": 0.0}, on_unused_input="warn"))

Context for the issue:

Users may wish to evaluate a pytensor function after having lost references to the underlying pytensor variables, in which case the easiest (or only) solution is to index the inputs_to_values dictionary by the variable names. The user then has to determine which variables are needed to evaluate a given function, which is cumbersome.

[ricardoV94]

I don't think that's possible because pytensor.function requires actual variables to be initialized, even when unused.

In your example it would need to know what type to use as x or y (such as scalar(dtype=int) or matrix(shape=(a, b))), but it can't know because it only has a string.

There is a utility to know what root variables are needed to evaluate a graph: pytensor.graph.basic.required_explicit_inputs or something like that.

Eval is only meant for debug, we have utilities to do everything eval can do and more without needing it.

[tvwenger]

Thanks for the lightning fast response, @ricardoV94! I submitted a PR with this feature.

My specific use-case is related to evaluating variables in a pymc model for the purposes of simulating data. For example, consider:

import pymc as pm

with pm.Model() as model:
    x = pm.Normal("x")
    y = pm.Normal("y")

    obs1 = pm.Normal("obs1", mu=x, observed=1.0)
    obs2 = pm.Normal("obs2", mu=x+y, observed=2.0)

If the user is only given access to model at this point, then I want there to be an easy way to evaluate obs1 and obs2 for given values of x and y. Here's what I have been doing in the past:

params = {
    "x": 3.0,
    "y": 2.0,
}

sim_obs1 = model["obs1"].eval(params, on_unused_input="ignore")
sim_obs2 = model["obs2"].eval(params, on_unused_input="ignore")

This fails with the ValueError since y isn't in the graph of obs1. With my PR, this works as intended.

Please let me know if there is a smarter way to do this! Thanks for your help!

[ricardoV94]

The statement eval is only meant for debugging still stands. In your case you probably want to make a function with inputs x and y and outputs obs1 and obs2.

This will save on repeated computations between the two outputs.

Then pymc specifically has the pm.do transformation for this kind of graph operations. After do you could call sample_prior_predictive or pm.draw. Also these will make sure to seed and update the rngs where your eval will return the same values if called repeatedly.

[ricardoV94]

https://www.pymc.io/projects/docs/en/stable/api/model/generated/pymc.model.transform.conditioning.do.html

[tvwenger]

Thanks for the suggestion about pm.do! I'll use that from now on, but I'll leave this feature request and PR open since I still think this would be a good addition to pytensor.