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
You can notice that the third parameter of the feature n is actually the second parameter of feature n+1 and the first parameter (weight) of feature n+2.
if (stride == 2) { /*either adaptive or normalized*/}
else {/*adaptive and normalized*/}
This seems that the possibility stride=1 (stride_shift=0) is not expected in the code. (--sgd --save_resume is not a common use case, but someone could use it for the effect of aggregated average loss, so this still looks like a bug.)
I tried vw -d rcv1_small.dat --save_resume --adaptive --readable_model readable.model and here I would expect two parameters per feature (the weight and the AdaGrad sum of squared gradients), but I still see three parameters in readable.model.
Note that this probably affects also binary model and also reading the model (few lines above in the code), not only writing, which may have more dangerous consequences.
The text was updated successfully, but these errors were encountered:
You can notice that the third parameter of the feature n is actually the
second parameter of feature n+1 and the first parameter (weight) of feature
n+2.
if (stride == 2) { /either adaptive or normalized/}else {/adaptive and normalized/}
This seems that the possibility stride=0 is not expected in the code. (--sgd
--save_resume is not a common use case, but someone could use it for the
effect of aggregated average loss, so this still looks like a bug.)
I tried vw -d rcv1_small.dat --save_resume --adaptive --readable_model
readable.model and here I would expect two parameters per feature (the
weight and the AdaGrad sum of squared gradients), but I still see three
parameters in readable.model.
Note that this probably affects also binary model and also reading the
model (few lines above in the code), not only writing, which may have more
dangerous consequences.
Reply to this email directly or view it on GitHub #596.
Let's say I want non-adaptive, non-normalized sgd. So I use either
--sgd
or--invariant
(depending on whether I want invariant updates or not).Surprisingly, the readable.model still contains three parameters per features, although I would expect just one parameter (the weight):
You can notice that the third parameter of the feature n is actually the second parameter of feature n+1 and the first parameter (weight) of feature n+2.
At https://github.com/JohnLangford/vowpal_wabbit/blob/master/vowpalwabbit/gd.cc#L815
you can see
This seems that the possibility stride=1 (stride_shift=0) is not expected in the code. (
--sgd --save_resume
is not a common use case, but someone could use it for the effect of aggregated average loss, so this still looks like a bug.)I tried
vw -d rcv1_small.dat --save_resume --adaptive --readable_model readable.model
and here I would expect two parameters per feature (the weight and the AdaGrad sum of squared gradients), but I still see three parameters in readable.model.Note that this probably affects also binary model and also reading the model (few lines above in the code), not only writing, which may have more dangerous consequences.
The text was updated successfully, but these errors were encountered: