训练完了，感觉哪里有问题。检查了下loss计算和acc也没发现，生成的基本不行呢。楼主有测试过吗

[hzp1991]

Epoch / Total epochs: 41 / 200 100%|██████████| 568/568 [11:36<00:00, 1.23s/it]INFO - 12/15/21 18:11:38 - 8:55:34 - Train Epoch: 41, avg loss: 9.9027, accuracy: 6.20% INFO - 12/15/21 18:12:56 - 8:56:52 - Validation. Epoch: 41, avg dev loss: 9.9009, accuracy: 6.24%

Epoch / Total epochs: 42 / 200 100%|██████████| 568/568 [11:35<00:00, 1.23s/it]INFO - 12/15/21 18:24:32 - 9:08:28 - Train Epoch: 42, avg loss: 9.9026, accuracy: 6.21% INFO - 12/15/21 18:25:50 - 9:09:46 - Validation. Epoch: 42, avg dev loss: 9.9009, accuracy: 6.25%

Epoch / Total epochs: 43 / 200 100%|██████████| 568/568 [11:35<00:00, 1.23s/it]INFO - 12/15/21 18:37:26 - 9:21:22 - Train Epoch: 43, avg loss: 9.9025, accuracy: 6.22% INFO - 12/15/21 18:38:44 - 9:22:40 - Validation. Epoch: 43, avg dev loss: 9.9010, accuracy: 6.24%

Epoch / Total epochs: 44 / 200 100%|██████████| 568/568 [11:36<00:00, 1.23s/it]INFO - 12/15/21 18:50:20 - 9:34:16 - Train Epoch: 44, avg loss: 9.9026, accuracy: 6.20% INFO - 12/15/21 18:51:38 - 9:35:34 - Validation. Epoch: 44, avg dev loss: 9.9009, accuracy: 6.24% INFO - 12/15/21 18:51:38 - 9:35:34 - No best dev loss, stop training.

[hzp1991]

默认参数，只改了bert base 换成了roberta Chinese

[Shougong]

您好 我测试过 在变体字识别转换为正常语句的任务中，大概有69这样的BLUE分值吧，还是比较低的。相比其他的例如UNiLM这种，差挺多的。

------------------ 原始邮件 ------------------ 发件人: "Shougong/bert-transformer-pointer-generator-networks" @.>; 发送时间: 2021年12月16日(星期四) 上午9:34 @.>; @.***>; 主题: Re: [Shougong/bert-transformer-pointer-generator-networks] 训练完了，感觉哪里有问题。检查了下loss计算和acc也没发现，生成的基本不行呢。楼主有测试过吗 (Issue #2)

默认参数，只改了bert base 换成了roberta Chinese

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

[hzp1991]

难怪了哈 谢谢

[ithieund]

I faced with the same issue, the loss doesn't go down :(