SVC (kernel=linear) JS prediction logic

[anmolshkl]

SVC (kernel=linear) JS prediction logic

[nok]

Hello @anmolshkl , do you get odd predictions between the original and transpiled estimator? Can you post the data and your params? Nevertheless there isn't any error in the regression tests of that classifier (e.g. log). But that doesn't mean, that there isn't an error.

[anmolshkl]

@nok thanks for the quick response. I'm getting the right predictions, and, the accuracy of the transpiled estimator is same as the scikit estimator (frankly, that's why I'm a little confused). To give you a little context about my confusion - I was trying to implement my own estimator for SVC and I tried following libsvm's code for prediction but I just couldn't get it right. Then, I looked into your estimator and found out that sklearn-porter's transpiled estimator uses decisions[d] = tmp + inters[d] while libsvm uses sum -= model->rho[p]. I replaced subtraction with an addition, and, my code started predicting values as expected.

I was just curious and wanted to know if you also encountered this problem while writing the algorithm for the estimator? Also, do you happen to have any reference material you used for writing the SVC estimator algorithm?

[nok]

Today I don't know how I figured it out in the past, but in general I add console outputs in the original scikit-learn classes and the generated classes. Then I compared each output and fix any differences on the fly. Finally I wrote regression tests to cover changes in the future.