TimDettmers
commented
6 years ago Make sure that you run it for a long time. You will see gradual improvement over time. I ran the model that I reported in the paper for 2 days I think. At first, it may seem to stagnate but still improves slowly.
I used the following parameters:
- input dropout 0.2
- feature dropout 0.2
- hidden dropout 0.3
- learning rate 0.003
- learning rate decay 0.99
- label smoothing 0.1 (I think, this is not clear from my logs)
I will close this for now. Let me know if you have any more problems with replication.
hey! To get the best results on the FB15k, what should the parameters be set to? I use the same settings as the FB15k-237 and can only get hit@10: 86 %.