Conflicts between hyperparameters for negative sampling?

[datistiquo]

Hi,

I wonder if there are possible interactions/conflicts when you use negative sampling with negative>0 and have hierarchical softmax accidentally activated hs=1? The docs says that only if hs=0 negative sampling will be used (negative>0). So I can hope that still if hs=1 and negative>0 hopefully no negative sampling is used?

Python 3.6
Win 10
NumPy 1.18.1
SciPy 1.1.0
gensim 3.8.1

[gojomo]

No, if both hs and negative are both non-zero, two output layers will be allocated, and both will be trained, backpropagating their corrections to the same shared input-vectors. This is an odd mode of operation, unlikely to be optimal in any situation, but matches the original word2vec.c from Google on which the gensim implementation was based. See also #2550 for more discussion. And, the project discussion list at https://groups.google.com/forum/#!forum/gensim is a better place for questions/support, reserving this issue-tracker for bugs & feature-requests.