I preprocessed 42 of the 43 subjects in my dataset with fmriprep 1.5.1rc2 singularity container without issue, but when I called it on my final subject this morning, this terrifying error was at the top of my standard error output:
WARNING: Version 1.5.1rc2 of fMRIPrep (current) has been FLAGGED
(reason: Nipype unstable).
That means some severe flaw was found in it and we strongly
discourage its usage.
Does this mean that I need to re-run fmriprep for the entire dataset? With 6 runs of BOLD containing 320 images each, this is a somewhat lengthy process…
The issue is with the pipeline engine (nipype). If it runs without error, then I would not worry about this. If you do run into errors, then the solution will be to upgrade.
Thanks much. It appears to have run without additional errors (at least not any that I can ascertain from the error log or output), but that error message put the fear in me…
Yeah. It’s a little bit unfortunate that it’s so blunt an instrument. It should be scary if, for example, we flip the phase encoding direction and need to invalidate a version. Perhaps we should add a link to a FAQ, or a per-version discussion, that will help people make informed choices.
I upgraded and started re-preprocessing with 1.5.2 yesterday, and it seems to be running much faster, although that may be a product of the relatively light load on the cluster at the moment. If there’s anything I can do to help w/rt FAQ or per-version discussion, let me know!
Since the reason for that version being flagged is that nipype was unstable, as long as you manage to reach the end of processing for all your subjects you’d be fine.
In principle, we try to adhere to (somewhat) semantic versioning principles. This means that, if a version has been flagged it can be replaced with some later release of the same series (1.5.x in your case). If the minor part of the version changes, then the upgrade should be more carefully done.
I think this answers the question I came here to ask, but just to confirm. So if we ran 9,000 or so subjects through fmriprep 1.5.0 and 2/3 of them completed successfully, running the remaining 1/3 that failed through 1.5.3 should be fine, versus running the full 9,000 through 1.5.3? The .0 to .3 releases are generally just differing on bugfixes and minor things like that, not changes in actual processing steps?
In general, yes. However, if you are processing 9,000 subjects it feels like you are in “production mode” to use industry terminology. Therefore, I’d recommend the upgrade in a targeted way (i.e., only if that new minor version addresses specifically the problem impeding 1/3 of the subjects to successfully run).