Closed ady-yu closed 2 years ago
I see. The author of DocRED has updated dev.json after I downloaded it. Please check the original version of dev.json (with 1000 documents): https://drive.google.com/drive/folders/1owp7ZRbrMl_s1ljIh6AvnmniLJSliV6h?usp=sharing Thanks for raising this issue!
It's actually normal to get F1=61.15 after training. You'll likely see another performance gain after running test_bert.sh
I see. The author of DocRED has updated dev.json after I downloaded it. Please check the original version of dev.json (with 1000 documents): https://drive.google.com/drive/folders/1owp7ZRbrMl_s1ljIh6AvnmniLJSliV6h?usp=sharing Thanks for raising this issue!
It's actually normal to get F1=61.15 after training. You'll likely see another performance gain after running test_bert.sh
Thanks you for you reply!!! I wonder the command bash scripts/test_bert.sh eider test hoi occupy how much GPU memory. It seems 16G is not enough!
I see. The author of DocRED has updated dev.json after I downloaded it. Please check the original version of dev.json (with 1000 documents): https://drive.google.com/drive/folders/1owp7ZRbrMl_s1ljIh6AvnmniLJSliV6h?usp=sharing Thanks for raising this issue! It's actually normal to get F1=61.15 after training. You'll likely see another performance gain after running test_bert.sh
Thanks you for you reply!!! I wonder the command bash scripts/test_bert.sh eider test hoi occupy how much GPU memory. It seems 16G is not enough!
Can your set evi_eval_mode=none in test_bert.sh? The memory is 6571M on my computer after setting it.
evi_eval_mode is to evaluate the evidence F1 (which is not necessary)
when running the command bash scripts/test_bert.sh eider test hoi, it occurs the following mistakes.
Besides, when running the command bash scripts/train_bert.sh eider test hoi, the best F1 only get 0.61158.
![image](https://user-images.githubusercontent.com/69707855/161212074-83b36c07-722b-4292-bb02-32e7d5f2fb10.png)