convertvec output inconsistent with word2vec ouput

[skbach]

Hi,

I think I found an issue with the code.

1.) Run the word2vec/demo-word.sh script.

Generates vectors.bin

2.) Run the following (modified version of above, but the output mode is text not binary)

time ./word2vec -train text8 -output vectors.txt -cbow 1 -size 200 -window 8 -negative 25 -hs 0 -sample 1e-4 -threads 20 -binary 0 -iter 15

3.) Convert the original binary into text with convertvec

./convertvec bin2txt vectors.bin vectors-converted.txt

4.) diff vectors.txt with vectors-converted.txt and the files are not the same. They have the same words in each file, and each word has 200 vectors, but the vectors do not correspond to the original vectors generated by word2vec in text mode.

Am I crazy? :)

[marekrei]

Word2vec initialises the vectors randomly. So even if you just run the same command twice, you're not going to get the same vectors.

[skbach]

Ah, that makes sense. Thanks!