Support for Go-Nogo and deadlines with missing data #358
Conversation
| hddm-wfpt = "^0.1.1" | ||
| seaborn = "^0.13.0" | ||
| seaborn = "^0.13.2" | ||
| pytensor = "<2.17.4" |
There was a problem hiding this comment.
are we still held back by that? just as a sidenote
There was a problem hiding this comment.
Yes, especially in non-conda environments
There was a problem hiding this comment.
Ok let's follow gameplan (this is the wrong place to put it but nevertheless :)):
- add the new network types
- fix initialization issues
- get to conda
- finish RL piece
- web interface
src/hssm/distribution_utils/dist.py
Outdated
| def make_likelihood_callable( | ||
| loglik: pytensor.graph.Op | Callable | PathLike | str, | ||
| loglik_kind: Literal["analytical", "approx_differentiable", "blackbox"], | ||
| backend: Literal["pytensor", "jax", "cython", "other"] | None, |
There was a problem hiding this comment.
is there need to make 'cython' backend explicit at this level?
There was a problem hiding this comment.
We might want this. It is possible to implement analytical likelihoods in cython right? Our likelihood kinds might need some rethinking, since they are not mutually exclusive...
There was a problem hiding this comment.
agree that we may want to change the naming conventions entirely eventually.
For now, cython stuff falls under blackbox (since no gradients) and I am just questioning if it really falls under 'backend' category in any relevant sense.
It's just some pyhton callable in the end.
| The VJP of the log-likelihood function computed at gz. | ||
| """ | ||
| _, vjp_fn = vjp(vmap_logp_no_data, *dist_params) | ||
| return vjp_fn(gz)[1:] |
There was a problem hiding this comment.
just as a question. Can't immediately remember why the return here skips 0 index.
There was a problem hiding this comment.
Here's the Jax Doc
If has_aux is False, returns a (primals_out, vjpfun) pair, where primals_out is fun(*primals). If has_aux is True, returns a (primals_out, vjpfun, aux) tuple where aux is the auxiliary data returned by fun.
vjpfun is a function from a cotangent vector with the same shape as primals_out to a tuple of cotangent vectors with the same number and shapes as primals, representing the vector-Jacobian product of fun evaluated at primals.
Looks like we don't need the primals
There was a problem hiding this comment.
Gotta love how complicated this reads...
src/hssm/hssm.py
Outdated
| @@ -219,9 +217,17 @@ def __init__( | |||
| link_settings: Literal["log_logit"] | None = None, | |||
| prior_settings: Literal["safe"] | None = None, | |||
| extra_namespace: dict[str, Any] | None = None, | |||
| missing_data: bool = False, | |||
| deadline: bool = False, | |||
| na_value: float = -999.0, | |||
There was a problem hiding this comment.
maybe missing_data_value here instead of na_value?
There was a problem hiding this comment.
I am fine with both. na_value is shorter and is also a convention in pandas. Let me know which one you prefer
There was a problem hiding this comment.
If we have missing_data kwarg, then probably go with missing_data_value kwarg to make it consistent.
We can also allow both missing_data_value and na_value and pool internally, if you think it's worth the burden.
In general appreciate the convention argument, but locally I think going for missing_data AND na_value is confusing.
digicosmos86
deleted the
357-support-for-networks-built-in-jax-and-multiple-networks
branch
No description provided.