A small mistake about the implementation of "Unsupervised_BiHalf.py"

[liyunqianggyn]

Hey,

I am one author of paper "Unsupervised_BiHalf", and thanks a lot for adding our paper in this nice Implementation.

I notice a small mistake in implementing the model at test time in line 77-79: During training, we train to align continuous distribution p(u) with the half-half distribution of +1 and −1(bi-half layer). This allows us at test time to simply use the "sign function" as a deterministic function for quantization.

Could you help us change bi-half to sign at test time, thanks a lot!

[swuxyj]

Yes, I got the MAP 0.593, which is amazing. I am sorry for my carelessness.

[liyunqianggyn]

Thanks :)