psychoinformaticslab / nipymc Goto Github PK

Hi,
PyMC3 has been renamed PyMC. If this affects you and you have questions, or you want someone to direct your rage at I'm available! Do let me know how i, or any of the PyMC devs can help.

Ravin

P.S Ive made a PR to the readme as well pinning the install to the PyMC3 branch. Hope this helps

Currently split_by only takes a single string (the name of a design variable) to split on. It would be useful to allow a list, and split the target variable based on the product of all variables in the list.

Suppose we pass a variable to add_term() and set split_by='condition'. If the target variable has no valid values for one or more levels of the splitting variable, a constant term will be added to the model. This is typically not desirable. We should check for this and, at minimum, issue a warning. We should probably also implement a general check to make there is (at most) a single constant column in the design.

Currently we represent the design matrices for each term added to the model as its own array, but we don't have a single representation of all variables in a single design matrix. This makes it hard to apply most validations (e.g., to test for rank deficiency), and is also generally clunky. We should probably move to an architecture that has a separate DesignMatrix class, with each BayesianModel class collecting all DesignMatrix classes. Then we can implement a finalize() call that concatenates all DMs, runs a series of validations, applies convolution, etc. This will also make it easier to abstract the fMRI-specific parts of the package into separate classes so that nipymc can inherit from psypymc.

Currently the HRF convolution is applied to the full design matrix that spans all subjects and runs. This means there could theoretically be some bleeding over across different runs and subjects. In practice this is unlikely to matter much given the designs we've been working with, but it needs to be addressed. Recommended approach is to refactor the modeling code so that all design matrix specification and convolution is done on a run-by-run basis, and then processed runs are concatenated into the full dataset just prior to sampling. This will also have the added benefit of making it easier to handle different numbers of volumes/events in each scanning run.

Currently event onsets are simply rounded to the nearest volume; we should add the ability to interpolate events (and use interpolation by default).