“Dongxu Zhang”版本间的差异
来自cslt Wiki
| 第2行： | 第2行： | ||
* Did following experiments: | * Did following experiments: | ||
* pooling method.(have run 195 hours for 117 iterations. performance is still raising(0.49). but haven't exceeded transE yet.(0.53), one iteration takes 1h40min ) | * pooling method.(have run 195 hours for 117 iterations. performance is still raising(0.49). but haven't exceeded transE yet.(0.53), one iteration takes 1h40min ) | ||
| − | * pooling method with sampling and minibatch.(performance grew very slowly after several iteration.(0.37) It seems that sampling is not as good as origin pooling method. And because cost stops going down when it is still a quite big value, which means sampling may lead to a saddle point? one iteration takes 40 minutes) | + | * pooling method with sampling and minibatch.(performance grew very slowly after several iteration.(0.37) It seems that sampling is not as good as origin pooling method. |
| + | And because cost stops going down when it is still a quite big value, which means sampling may lead to a saddle point? one iteration takes 40 minutes) | ||
* attention pooling method.(very slow, one iteration may take two days or more. still waiting for the performance of first iteration.) | * attention pooling method.(very slow, one iteration may take two days or more. still waiting for the performance of first iteration.) | ||
* attention pooling method with sampling and minibatch.( still running. one iteration takes 1 hour.) | * attention pooling method with sampling and minibatch.( still running. one iteration takes 1 hour.) | ||
2015年10月5日 (一) 13:19的版本
Last Week:
- Did following experiments:
* pooling method.(have run 195 hours for 117 iterations. performance is still raising(0.49). but haven't exceeded transE yet.(0.53), one iteration takes 1h40min )
* pooling method with sampling and minibatch.(performance grew very slowly after several iteration.(0.37) It seems that sampling is not as good as origin pooling method.
And because cost stops going down when it is still a quite big value, which means sampling may lead to a saddle point? one iteration takes 40 minutes)
* attention pooling method.(very slow, one iteration may take two days or more. still waiting for the performance of first iteration.)
* attention pooling method with sampling and minibatch.( still running. one iteration takes 1 hour.)
This Week:
- waiting for results and also think about other improvements.
- pooling method with entity and relation is too slow. Another simplification is to ignore the around entities during pooling, only relation types left, which leads to a feature vector joining behind the orginal entity vector, describing the type of entity. This is more like the original transE.
- I feel that the chain rule should based on random walk. Mean pooling leads to equal weights. But random walk give high possibility to high page rank entities. So there should be improvements.
