Reproduce the experiment results of TransE
Opened this issue · 2 comments
With the default config in the code, I ran the FB15k_TransE.py to train a best model, and I found the evaluation result of that model was really different from what the paper said. The default epochs in the code if 500, and the paper said that the model was trained at most 1000 epochs. But current result was even much better than the result in paper. Does the paper used the code in the current repo? Below is my result.
MICRO:
-- left >> mean: 229.41149, median: 23.0, hits@10: 37.377%
-- right >> mean: 160.86706, median: 14.0, hits@10: 45.088%
-- global >> mean: 195.13927, median: 18.0, hits@10: 41.233%
MACRO:
-- left >> mean: 106.30351, median: 83.18991, hits@10: 55.557%
-- right >> mean: 84.51045, median: 63.63632, hits@10: 63.104%
-- global >> mean: 95.40698, median: 33.58689, hits@10: 59.331%
@eshijia Not sure why, but they've mention something about reaching better results on the FB15k.
I've got exact numbers too :)
$ virtualenv --system-site-package -p /usr/local/Cellar/python/2.7.11/bin/python venv-sme
$ pip install numpy scipy
$ pip install theano
$ source venv-sme/bin/activate
(venv-sme) $ python SME/FB15k/FB15k_TransE.py
(venv-sme) $ python SME/FB15k/FB15k_evaluation.py SME/FB15k/FB15k_TransE/best_valid_model.pkl
### MICRO:
-- left >> mean: 229.41149, median: 23.0, hits@10: 37.377%
-- right >> mean: 160.86706, median: 14.0, hits@10: 45.088%
-- global >> mean: 195.13927, median: 18.0, hits@10: 41.233%
### MACRO:
-- left >> mean: 106.30351, median: 83.18991, hits@10: 55.557%
-- right >> mean: 84.51045, median: 63.63632, hits@10: 63.104%
-- global >> mean: 95.40698, median: 33.58689, hits@10: 59.331%
I got the same result. two questions: the parameters in the code is different from paper,which one is optimal? even the result of *.out file on the official github page is different frome the paper's result ,is it not funny?