Diverse Classifiers for SSRO #1336
Replies: 1 comment 4 replies
-
|
@jevillegasdatTII we already had most of those classifiers in Qibocal, until #1025 (including a NN-based one, for which we were using ONNX to store the classifier). They were not providing much benefit on top of the linear classifier, especially for properly amplified transmission lines (we love TWPAs ❤️). Moreover, non-linear classifiers are not supported for real-time classification. And that's the reason (together with limiting deps) why we decided to drop them in Qibocal. However, this does not mean we can not review this decision: do you have any application in mind? |
Beta Was this translation helpful? Give feedback.
-
|
The availability of other real time classifiers can be increased with work with Manarat, we talked a little about what could be implemented in an FPGA, but maybe this could be discussed also with them. It does boils down to
For nice working readouts I can see how there is no point but for weirder responses (let's say some instability in amplifiers) these could be useful. But for classification of more than one state these could become very useful. |
Beta Was this translation helpful? Give feedback.
-
|
@jevillegasdatTII can you report the different fidelities you found for reference? |
Beta Was this translation helpful? Give feedback.
-
Beta Was this translation helpful? Give feedback.
-
|
thanks a lot!! |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Since #1278 uses sklearn for classification, would it make sense to also allow for the user to point at other classifiers from scikit learn to allow fast evaluations of more classifying techniques. This could be done relativeluy easy just by changing where the classification model is instantiated (we may need to also add a Standardizer before to make sure that different classifiers still work).
Of course the threshold and angle of the linear classifier would no longer mean anything for some classifier, and this would need to be handled.
Yes I know that #1278 is called simplify single shot routines, and that the other classifiers that used to be supported (Naive-Bayes) are no longer there, but its rather easy to add this functionality using sklearn.
In #1335 I added the same list of classifiers that is in the sckiti-learn examples as a possible implementation of this. Here is a set of different classifiers being called (only the LDA computes angle and threshold)
k nearest neighbors:
random forest
adaboost
qda
lda
Beta Was this translation helpful? Give feedback.
All reactions