word2vec give inconsistent results

[samehif]

I have installed gensim using pip install gensim. the version i have is the latest 0.12.4.

I have created a word2vec model from the same data and using the same parameters but every time i create the model it give different results. I have tried to use seed with fixed int when creating the model but it still behaves in the same way.

Here is some example code:

`>>> from nltk.corpus import brown

from gensim.models import Word2Vec
sentences = brown.sents()[:100]
model = Word2Vec(sentences, size=10, window=5, min_count=5, workers=4, seed=128)
model[sentences[0][0]]
array([ 0.04913874, 0.04574081, -0.07402877, -0.03270053, 0.06598952,
0.04157289, 0.05075986, 0.01770534, -0.03796235, 0.04594197], dtype=float32)
model = Word2Vec(sentences, size=10, window=5, min_count=5, workers=4)
model[sentences[0][0]]
array([ 0.04907205, 0.04569579, -0.07379777, -0.03273782, 0.06579078,
0.04167712, 0.05083019, 0.01780009, -0.0378389 , 0.04578455], dtype=float32)
model = Word2Vec(sentences, size=10, window=5, min_count=5, workers=4)
model[sentences[0][0]]
array([ 0.04906179, 0.04569826, -0.07382379, -0.03274316, 0.06583244,
0.04166647, 0.0508585 , 0.01777468, -0.03784611, 0.04578935], dtype=float32)`

[gojomo]

As soon as you use more than one worker thread, scheduling jitter from the OS means examples are trained in slightly different order. And, sources of randomness in the algorithm – like frequent-word downsampling – will be applied to different examples, meaning slightly different words chosen each run. (And even further, in Python3, PYTHONHASHSEED-controlled randomization on each interpreter-launch will affect the iteration order of keys in the discovered vocabulary dictionary, which can again affect their sampling or ordering inside the model.)

So: you can't expect identical results without taking extra steps, including limiting yourself to just a single worker thread. A pending PR (#642) will make this clearer in the doc-comment.

See also this thread on the discussion forum – https://groups.google.com/d/msg/gensim/7eiwqfhAbhs/qC0pmbw5HwAJ – the same considerations apply to Word2Vec. (If you have other questions that are not likely to be bugs, it's better to discuss at that forum than in this issue-tracker.)