xlvector / sw2v

sparse word2vec

Word2Vec with ps-lite

Implement CBOW with negative sampling by:

• OpenMP : using -DLOCAL
• Parameter Server: using ps-lite

This project is under develop now. So, does not suggest you use it in production enviroment.

Next Steps

• More tests on large dataset
• More tests on multiple machines
• More tests with docker

sw2v's Issues

model init problem?

int model_size = nword * nhidden_;
vectorps::Key keys(nword, 0);
vector vals(nword, 0);
for(int i = 0; i < nword**(maybe change to model_size)**; i++) {
keys[i] = rand() % model_size;
vals[i] = (RAND01() - 0.5) / sqrt(float(nhidden_) + 1.0);
}

auc calculation

src/sw2v.cc中auc的计算是不是不对，p的算法结果是负的