Fix to bug 1729: no bias regularization for BFGS not working #1794
Conversation
| if (all.no_bias) | ||
| { | ||
| if (b.regularizers == nullptr) | ||
| { | ||
| (&weights[constant])[W_GT] -= regularization * weights[constant]; | ||
| ret -= 0.5 * regularization * (weights[constant]) * (weights[constant]); | ||
| (&weights.strided_index(constant))[W_GT] -= regularization * (weights.strided_index(constant)); |
There was a problem hiding this comment.
Is there a use case to NOT using strided_index? If not then perhaps we should make the operator[] private to avoid this bug in future?
There was a problem hiding this comment.
I don't know vw well enough to comment on that but like we discussed in our conversation earlier today, it would be a big change, probably out of scope of this PR
There was a problem hiding this comment.
This is definitely out of scope of the PR and if we decide we want to look into this I can create and issue to track it
There was a problem hiding this comment.
Unfortunately, there is. Currently, in setup_example all features are multipled by a stride which allows direct access to the parameters via [] as per https://github.com/VowpalWabbit/vowpal_wabbit/blob/master/vowpalwabbit/gd_predict.h#L33
There was a problem hiding this comment.
So does offset correspond with the stride? Maybe the default should be to access using the stride and then explicitly state when you want unstrided? It depends on what the more common/intuitive usage is
There was a problem hiding this comment.
Not exactly.
Stride controls floats per parameter. Offset indexes into logically distinct sets of parameters.
|
Pavithra, thanks for this. Let's merge as soon as tests pass. |
…abbit#1794) * fix to bug 1729 * fix to bug 1729
fixes #1729
Cause of problem: weights[constant] always returns 0, so there was no subtraction happening at all with the below two lines:
(&weights[constant])[W_GT] -= regularization * weights[constant]; ret -= 0.5 * regularization * (weights[constant]) * (weights[constant]);This is because we weren't accessing it the right way.
weights.strided_index(constant)gives us what we want. Thus, the correct code should probably be:(&weights.strided_index(constant))[W_GT] -= regularization * (weights.strided_index(constant)); ret -= 0.5*regularization*(weights.strided_index(constant))*(weights.strided_index(constant));