Default target_accept not mentioned anywhere in the docs

[ricardoV94]

Description

...

[kdotmanoj]

can you provide more info about this?

[ricardoV94]

can you provide more info about this?

I think it's pretty self explanatory.

[ricardoV94]

It's a keyword argument to pm.sample

[kdotmanoj]

so the value is 0.8 right, found here:

i am finding it difficult to find where to edit as it is autogenerated from the code.

[ricardoV94]

so the value is 0.8 right, found here:

i am finding it difficult to find where to edit as it is autogenerated from the code.

You're looking at nutpie sample function not pymc

[cluhmann]

Is target_accept a kwarg to pm.sample() or just another **kwargs? The argument isn't even included in the API docs, so if it's the former, there's a bigger problem.

I mean, it's detailed here, but I assumed that the sampling defaults would be specified deeper in the code, possibly allowing for different default for different sampling algorithms.

[Armavica]

There are at least a couple places where a default is provided deep in the code:

pymc/pymc/sampling/mcmc.py

Line 730 in 747fda3

target_accept=kwargs.pop("nuts", {}).get("target_accept", 0.8),

pymc/pymc/sampling/jax.py

Line 478 in 747fda3

target_accept: float = 0.8,

pymc/pymc/step_methods/hmc/hmc.py

Line 119 in 747fda3

kwargs.setdefault("target_accept", 0.65)

pymc/pymc/step_methods/hmc/base_hmc.py

Line 72 in 747fda3

target_accept=0.8,

[cluhmann]

Exactly. But those are each specified separately and, it turns out, in a way that creates different defaults for different algorithms. So I'm not sure what's missing from the docs.

[ricardoV94]

Exactly. But those are each specified separately and, it turns out, in a way that creates different defaults for different algorithms. So I'm not sure what's missing from the docs.

I couldn't figure out what the default is as an advanced user. Since it's common advice to increase it (even automatic in case of divergences) it should be easy to find what the default is.

[cluhmann]

This is caused by pm.sample() being a franken-wrapper that dispatches to a bunch of actual sampling methods, but seems like it does something useful all on its own. Where would you expect the default value to be documented?

[ricardoV94]

pm.sample 🤷‍♂️🫠🥹? Or do you have a better idea?

[cluhmann]

But ... it's not an argument to pm.sample(). I do have a better idea, but you didn't like it (e.g., deprecate pm.sample). 😂

[ricardoV94]

But ... it's not an argument to pm.sample(). I do have a better idea, but you didn't like it (e.g., deprecate pm.sample). 😂

It is, just hidden inside **kwargs.

It's actually a special kwarg, the only one we accept both directly and as usual step kwargs

[ricardoV94]

See #5622

[kdotmanoj]

@ricardoV94 do you suggest mentioning the default values of target_accept here ?