Casper da Costa-Luis

> identify if the `cmd` needs to be reproduced I was intending it to always be run (unconditionally) but good point that this may be inefficient. Maybe we need a...

@dmpetrov definite misunderstanding here. In the issue here I'm referring to `filter` as the a command which extracts a subset of the file defined in `path`. This is completely and...

When you say > checksums from user defined functions Do you mean the hash of the UDF/`filter` itself (rather than its output)? Similar to the `md5` for `cmd` right now?...

Hmm I see your concern, but we definitely need to save the filter's output checksum (i.e. subset of file checksum) in the DVC-file for reproducibility across systems. How about appearing...

> when you run UDF (unconditionally) and it returns a checksum how you decide if the dependency changed and the stage needs to be re-run The stage's checksum *is* the...

> storing checksums in a local DVC DB does not affect reproducibility across systems. However, it affects performance Yes, that's what I meant; should have been more precise. I also...

Users can also call Python from MATLAB: https://uk.mathworks.com/help/matlab/call-python-libraries.html

released as `v24.0.1`

Indeed... the ethos is to use Python unless for performance reasons C++/CUDA is needed. I suppose STIR wouldn't be able to use NiftyPET, but SIRF would?

NiftyPET ([NIPET](https://github.com/NiftyPET/NIPET) and [NIMPA](https://github.com/NiftyPET/NIMPA)) use [CuVec](https://github.com/AMYPAD/CuVec) arrays, which are simultaneously both numpy arrays and also CUDA arrays :) I suppose [CIL](https://github.com/TomographicImaging/CIL) might indeed be an appropriate place for integrating NiftyPET....