You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, thanks for publishing this code. It is helping us a lot.
I want to ask about the part where you calculate the overall covariance for GMM.
If my understanding is correct, the overall covariance for GMM with weight w_i is:
So I think the corresponding change to your current implementation is:
---a/python/gps/utility/gmm.py+++b/python/gps/utility/gmm.py@@ -89,7+89,7 @@ classGMM(object):# For some reason this version works way better than the "right"# one... could we be computing xxt wrong?diff=self.mu-np.expand_dims(mu, axis=0)-diff_expand=np.expand_dims(diff, axis=1) * \
+diff_expand=np.expand_dims(self.mu, axis=1) * \
np.expand_dims(diff, axis=2)
wts_expand=np.expand_dims(wts, axis=2)
sigma=np.sum((self.sigma+diff_expand) *wts_expand, axis=0)
The comment in your code suggests that you're aware that the current code is not theoretically derived, so I'd like to know if my modified version is what you tried as the "right" version, and if so, how it performed "worse" compared to the current version.
I did try the above modification with mcj_badmm_example, and I couldn't find any significant difference between the original version.
Thanks!
The text was updated successfully, but these errors were encountered:
We're aware that the code is mathematically incorrect. If you would like, you can make a PR and we can run the benchmark tasks to make sure there is no performance degradation.
We've been planning on replacing this GMM with a version based on the scikit-learn implementation, so this code will likely be removed in the future.
Hi, thanks for publishing this code. It is helping us a lot.
I want to ask about the part where you calculate the overall covariance for GMM.
If my understanding is correct, the overall covariance for GMM with weight
w_i
is:So I think the corresponding change to your current implementation is:
The comment in your code suggests that you're aware that the current code is not theoretically derived, so I'd like to know if my modified version is what you tried as the "right" version, and if so, how it performed "worse" compared to the current version.
I did try the above modification with
mcj_badmm_example
, and I couldn't find any significant difference between the original version.Thanks!
The text was updated successfully, but these errors were encountered: