search

lonePatient / BERT-NER-Pytorch

Chinese NER(Named Entity Recognition) using BERT(Softmax, CRF, Span)
MIT License
2.05k stars 425 forks source link

在train过程中注释掉了evaluate 会影响模型的结果 #73

Open nicholasbao opened 2 years ago

nicholasbao commented 2 years ago

30/2022 18:14:35 - INFO - root - acc: 0.9514 - recall: 0.9552 - f1: 0.9533 - loss: 2.1387 01/30/2022 18:14:35 - INFO - root - Entity results 01/30/2022 18:14:35 - INFO - root - * CONT results **** 01/30/2022 18:14:35 - INFO - root - acc: 1.0000 - recall: 1.0000 - f1: 1.0000 01/30/2022 18:14:35 - INFO - root - * EDU results **** 01/30/2022 18:14:35 - INFO - root - acc: 0.9720 - recall: 0.9811 - f1: 0.9765 01/30/2022 18:14:35 - INFO - root - * LOC results **** 01/30/2022 18:14:35 - INFO - root - acc: 1.0000 - recall: 1.0000 - f1: 1.0000 01/30/2022 18:14:35 - INFO - root - * NAME results **** 01/30/2022 18:14:35 - INFO - root - acc: 0.9910 - recall: 1.0000 - f1: 0.9955 01/30/2022 18:14:35 - INFO - root - * ORG results **** 01/30/2022 18:14:35 - INFO - root - acc: 0.9316 - recall: 0.9369 - f1: 0.9342 01/30/2022 18:14:35 - INFO - root - * PRO results **** 01/30/2022 18:14:35 - INFO - root - acc: 0.9000 - recall: 1.0000 - f1: 0.9474 01/30/2022 18:14:35 - INFO - root - * RACE results **** 01/30/2022 18:14:35 - INFO - root - acc: 1.0000 - recall: 0.9333 - f1: 0.9655 01/30/2022 18:14:35 - INFO - root - * TITLE results **** 01/30/2022 18:14:35 - INFO - root - acc: 0.9551 - recall: 0.9551 - f1: 0.9551 上面的结果是在train的时候，没有去掉evaluate的结果 下面的是在train的时候，去掉了evaluate的结果 30/2022 18:17:20 - INFO - root - Eval results 01/30/2022 18:17:20 - INFO - root - acc: 0.9485 - recall: 0.9593 - f1: 0.9538 - loss: 2.1727 01/30/2022 18:17:20 - INFO - root - Entity results 01/30/2022 18:17:20 - INFO - root - * CONT results **** 01/30/2022 18:17:20 - INFO - root - acc: 1.0000 - recall: 1.0000 - f1: 1.0000 01/30/2022 18:17:20 - INFO - root - * EDU results **** 01/30/2022 18:17:20 - INFO - root - acc: 0.9906 - recall: 0.9906 - f1: 0.9906 01/30/2022 18:17:20 - INFO - root - * LOC results **** 01/30/2022 18:17:20 - INFO - root - acc: 1.0000 - recall: 1.0000 - f1: 1.0000 01/30/2022 18:17:20 - INFO - root - * NAME results **** 01/30/2022 18:17:20 - INFO - root - acc: 0.9910 - recall: 1.0000 - f1: 0.9955 01/30/2022 18:17:20 - INFO - root - * ORG results **** 01/30/2022 18:17:20 - INFO - root - acc: 0.9375 - recall: 0.9465 - f1: 0.9420 01/30/2022 18:17:20 - INFO - root - * PRO results **** 01/30/2022 18:17:20 - INFO - root - acc: 0.9000 - recall: 1.0000 - f1: 0.9474 01/30/2022 18:17:20 - INFO - root - * RACE results **** 01/30/2022 18:17:20 - INFO - root - acc: 1.0000 - recall: 0.9333 - f1: 0.9655 01/30/2022 18:17:20 - INFO - root - * TITLE results **** 01/30/2022 18:17:20 - INFO - root - acc: 0.9414 - recall: 0.9551 - f1: 0.9482

想问下，为什么去掉了evaluate会对模型的训练结果产生影响，按照一般的模型训练，在训练过程中evaluate添加与否，应该不会影响模型的最终预测效果，这个是什么原因？谢谢

nicholasbao commented 5 months ago

您好，您的来信已收到，谢谢！