Bijector error during data transformation in Turing about sequentialsamplingmodels.jl HOT 5 CLOSED

On a user-experience note:
The fact that one needs to pass a data object to the model formatted differently (as a tuple of individual tuples) than the data object expected by the SSM distributions (as a tuple of vectors) makes it feel a bit clunky.

I wonder if it would not be easier to (additionally) allow for an independent specification of choice and rt in the distributions. For instance, add the rt=none keyword arguments to LBA(data; **kwargs), and in case data is a vector (i.e., someone just passed choice alone) & rt != none, we just re-format data = (choice=data, rt=rt) as a first step in the function.

This would give some flexibility and hopefully make the practical usage smoother (also it would remove the need to do extra data transformation to make a vector of tuples).

An alternative version would be to also define

LBA(choice::Vector, rt::Vector; *kwargs)
    LBA((choice=choice, rt=rt), ...)

This would probably be more elegant than having an optional keyword argument

from sequentialsamplingmodels.jl.

Wait no actually no this wouldn't work since we need something that can be indexed in the for i in n loop of Turing... Okay nevermind then

EDIT: (and sorry for the spam, I wrote too quickly before giving a good thought) unless we allow data to be a matrix with 2 columns (one for choice and for RT) (in which case it would also return a matrix) . We could then do X[i, :] ~ LBA(...)

from sequentialsamplingmodels.jl.

@DominiqueMakowski, no problem. These things can be a bit tricky. The convention I adopted with NamedTuples follows the same convention used throughout Distributions.jl. For example, rand(Normal()) returns a scalar while rand(Normal(), 1) returns a vector. In this package, we return a NamedTuple of scalars for rand(LBA()) and a NamedTuple of vectors for rand(LBA(), 1).

You are right that the lack of indexing in the case of vectors is inconvenient. I did consider using a 2D array of floats instead, but the potential problem is that the choice indices are actually integers, and are often used for logical indexing (e.g. [rt[choice .==1]). Unfortunately, logical indexing could potentially fail due to float point error. I'm not sure how likely it would be in the scope of these models, but it could potentially create a silent bug which could invalidate someone's analyses. Depending on what you are doing, destructuring the NamedTuple can convenient, e.g., choice,rt = rand(LBA(), n) . As shown the doc strings, this works with broadcasting. Although the system is not perfect, I think some degree of pre-processing of data is almost unavoidable.

By the way, if you want a simpler way to generate a vector of NamedTuples, you can use a list comprehension or map:

n_obs =50
data1 = [rand(LBA(ν=[3.0, 2.0], A=0.8, k=0.2, τ=0.3)) for _ ∈ 1:n_obs]
data2 = [rand(LBA(ν=ν=[1.0, 1.5], A=0.8, k=0.2, τ=0.3)) for _ ∈ 1:n_obs]
data = vcat(data1, data2)

from sequentialsamplingmodels.jl.

I think we can close this

from sequentialsamplingmodels.jl.

I think there still might be a bug in Turing or Bijectors. We circumvented the problem by writing better code, but it seemed to be performing an unnecessary transformation. When I have time, I still raise the issue with devs at Turing.

from sequentialsamplingmodels.jl.