WladimirSidorenko/word2vec

different results obtained when launched with multiple threads

WladimirSidorenko opened this issue · 1 comments

  1. Description:
    The results obtained from this reimplementation, when launched with multiple threads, are different from the ones produced by the original word2vec.
  2. Example:
# modified version
./bin/word2vec -train tokens.sample.tok -output changed.vec 

# original word2vec
./word2vec -train ~/Projects/word2vec/build/tokens.sample.tok -output ~/Projects/word2vec/build/original.vec

# diff
diff -q original.vec changed.vec 
Files original.vec and changed.vec difer
  1. Expected:
    no difference between the files

Partially fixed (critical sections introduced into the training method; full compliance guaranteed for single-thread applications)

No full fix is possible due to a thread schedule different from the original.

Test case added to tests/test_1.test