acl2020-transition-discontinuous-ner
acl2020-transition-discontinuous-ner copied to clipboard
daixiangau
bad training result
Hi, I'm implementing your model. But I get 0 f1, 0 recall not only on my own data but also on the pulic CADEC data. Strangely, the accuracy is above 0. Is there a reason why this happened?
result example: loss: nan; discontinuous-precision-overall: 0.0; discontinuous-recall-overall: 0.0; discontinuous-f1-overall: 0.0; precision-overall: 0.0; recall-overall: 0.0; accuracy: 0.6972145593136464; f1-overall: 0.0
Hi, I am not sure what do you mean by saying 'implementing'. Are you reimplement using your code base? or you just run my code? If later, this issue may help?
Hi, thanks for the quick reply. I'm running your code and I got nan loss and 0 f1. The issue you mentioned only said he solved the problem but didn't say how. As there is no "requirements.txt" file, so I'm training with python 3.8 and torch 1.7, is this the reason? As I know, different versions may fail the code to run, but is it the reason to cause bad results? I'll try again using the torch version you said.
Hi, there is a train.log in that issue, maybe get some clues by comparing yours and that one?
Hi, I compared your log with mine. I simply run your code on my data and the CADEC data, so the config is all the same. I can't get the result by changing python and torch version, is it possible you can share your environment setups with me?