Repository navigation
add optimise as option in transform to match other flucoma behaviour - #357
Conversation
|
actually maybe the best is to stop returning the embeddings with fittransform. at the moment this is our only object that has a different result with fit -> transform vs fittransform. so what we should probably do is to return the embeddings with fit, or getXXX like (s)kmeans, and run fittransform as fit then transform... in all cases, a UX discussion is needed |
|
ok, overnight wisdom makes for another proposal: we do not touch fittransform, as it gives us the embeddings, but we do add this optimize bool to transform, and document the fit->tranform ISNOT fittransform in the reference. for most usecase, fit then transform with the switch off is the flucoma behaviour, and fittransform returns the embeddings. |
|
|
||
| MessageResult<void> transform(InputDataSetClientRef sourceClient, | ||
| DataSetClientRef destClient) | ||
| DataSetClientRef destClient, bool optimize = false) |
There was a problem hiding this comment.
No benefit in having a default value here (and I don't generally like them), and could rename like algorithm
tremblap
left a comment
There was a problem hiding this comment.
ok I am resolving comments, but please give me your opinion on the general discussion with the 3 ideas on what to do with fittransform not being fit+transform
Co-authored-by: Owen Green <gungwho@gmail.com>
I thought leaving |
this adds a flag to
transformto harmonise the behaviour flucoma style (transform and fittransform yielding the same results) while keeping the canonical option of SGD refinement if the user wants it.TODO: document and SC-class change if approved