fixed rbf kernel in cka #175
Conversation
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## master #175 +/- ##
==========================================
+ Coverage 76.40% 76.42% +0.02%
==========================================
Files 40 40
Lines 2051 2053 +2
Branches 261 261
==========================================
+ Hits 1567 1569 +2
Misses 400 400
Partials 84 84
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Sentry. |
| """Use an rbf kernel for computing the gram matrix. Sigma defines the width.""" | ||
| GX = X @ X.T | ||
| KX = np.diag(GX) - GX + (np.diag(GX) - GX).T | ||
| if sigma is None: | ||
| if self.sigma is None: | ||
| mdist = np.median(KX[KX != 0]) | ||
| sigma = math.sqrt(mdist) |
There was a problem hiding this comment.
Question, if self.sigma=None, you compute it's value, but shall it be stored? Or shall we recompute it every time?
There was a problem hiding this comment.
We could store it but the caveat here is that sigma depends on the input variable X which changes as soon as we use the feature space Y instead of X. I am not sure it's worth storing two sigmas, one for X and one for Y, if self.sigma=None. Also, self.sigma=None should be seen as a special case. Ideally, one assigns self.sigma a specific value which is then used for the bandwidth of the rbf kernel irrespective of the feature space.
fixed sigma arg in rbf kernel