Need python API support for performing all PCA operations for correspondence model currently present in Studio.
The API should calculate the following:
- Eigenvalues
- Eigenvectors
- PCA Loading
- PCA Component Scores / Variance
All these quantities should be saved as numpy arrays.
iyerkrithika21
All 7 comments
Related to #995
iyerkrithika21
on 23 Feb 2021
@akenmorris: Currently, which library does Studio use to perform the principal component analysis?
@sheryjoe Should we use the same library for the API or use python libraries like sklearn to perform the PCA analysis?
If we want to use python libraries, following the same suite of ShapeCohortGenPackage/DataAugmentationsPackage I can start working on PCAPackage in the python library.
iyerkrithika21
on 2 Mar 2021
Studio uses the ParticleShapeStatistics class:
https://github.com/SCIInstitute/ShapeWorks/blob/master/Libs/Particles/ParticleShapeStatistics.h
In general, my recommendation is to use a single library for both. This keeps results and behavior consistent across tools. I worry that having multiple implementations is a step backward into a disjoint set of tools that we've been trying to consolidate over time.
FYI, there are some different plans for the future of statistical analysis including using R.
akenmorris
on 2 Mar 2021
Are we planning on creating a pybind wrapper of ParticleShapeStatistics?
archanasri
on 2 Mar 2021
Are we planning on creating a pybind wrapper of ParticleShapeStatistics?
Yes, as part of the shapeworks python library.
sheryjoe
on 3 Mar 2021
830 #1172
cchriste
on 30 Mar 2021
It’ll be 1000x easier to get this feature out by simply exporting Eigen::Matrix that already has all these functions.
cchriste
on 30 Mar 2021
Related issues
jadie1
·
8Comments
akenmorris
·
22Comments
cchriste
·
3Comments
akenmorris
·
23Comments
iyerkrithika21
·
12Comments
Most helpful comment
Studio uses the ParticleShapeStatistics class:
https://github.com/SCIInstitute/ShapeWorks/blob/master/Libs/Particles/ParticleShapeStatistics.h
In general, my recommendation is to use a single library for both. This keeps results and behavior consistent across tools. I worry that having multiple implementations is a step backward into a disjoint set of tools that we've been trying to consolidate over time.
FYI, there are some different plans for the future of statistical analysis including using R.