Dongxu Zhang 2015-12-28
来自cslt Wiki
Last Week
- change classifier to liblinear. slightly improve the performance.
- dropout some words from input can improve the performance. And dropout from both input and output results in similar performance.(try 1/10->1/5)
- with initialization and dropout, the performance improved a little.
- idf weight on sum of pooling(less), on sum of cost(more) hurt the performance.
This Week
- Tune the dropout ratio(1/2).
- summarize the classification accuracy on knn and liblinear.