CarlinLiao
commented
1 year ago Upon limited review of the differences between cg-gnn and the histocartography packages, I've come to the conclusion that, while reverting to depending on histocartography is very possible, its ease or lack thereof is not commensurate with its urgency at this point in time. Most changes I've made in my fork are either expanded docstrings or cosmetic changes which, combined with a few but significant functional abstractions and the incorporation of functionality from hact-net and patho-quant-explainer, make it so that there might be some significant details that would be lost in the reversion to histocartography dependency. To resolve this issue, I'll simply move the code over as-is, and leave reversion to depending on histocartography as future work.
After discussion, it was decided that we'd prefer not to have to manage two different Python packages: SPT and cg-gnn. To that end, we're aiming to completely move cg-gnn into SPT, following the partial move of graph creation and plotting functionality in #244.
Code not primarily developed by us will be located in a subdirectory/submodule, e.g.,
cggnn.external, and not required to be maintained to the same standard as other code in SPT.To minimize the amount of less-maintained code, I'll also explore whether it's possible to revert some of the edits made to the histocartography repository to create PIP cg-gnn. If so, we'll add histocartography as a dependency and rely on imports from there to power the model training and inference portions of cg-gnn that will be moved into SPT either way. (The provenance of this code is hact-net, which is no longer maintained and not published on PIP, unlike histocartography.)