请教复现问题 about m3tl

[JayYip]

没看出来什么问题, 是否没有贴全

[18140663659]

没看出来什么问题, 是否没有贴全

训练得acc 一直很低

[JayYip]

明白, 我有空看看

[hoangthangta]

I have the same with a low accuracy. It seems using "|" for the problems will have better results than "&".

problem = 'c2_cls|c3_cls|c4_cls|c5_cls|c6_cls'

[JayYip]

I have the same with a low accuracy. It seems using "|" for the problems will have better results than "&".

problem = 'c2_cls|c3_cls|c4_cls|c5_cls|c6_cls'

Thanks for the info. Will take a look tomorrow. BTW, are the results significantly better(meaning it can reach expected accuracy) when chaining with "|" or just slightly better?

[hoangthangta]

It is slightly better only. Btw, thank for your hard works but it seems there is a problem with the accuracy. When using your code for my case (offensive language detection), the accuracy gets stuck about 63% while I can obtain 83% with mt-dnn.

[JayYip]

In that case, I think there's a bug. Will further investigate.

[JayYip]

没看出来什么问题, 是否没有贴全

训练得acc 一直很低

好像无法复现. 我在我自己这边跑是正常的, mean_acc计算有点问题, 但是不影响效果.

Epoch 2/20
48/48 [==============================] - 5s 95ms/step - mean_acc: 0.5634 - weibo_ner_acc: 0.9591 - BertMultiTaskTop/weibo_ner/losses/0: 0.1731
Epoch 3/20
48/48 [==============================] - 4s 74ms/step - mean_acc: 0.5293 - weibo_ner_acc: 0.9703 - BertMultiTaskTop/weibo_ner/losses/0: 0.0904
Epoch 4/20
48/48 [==============================] - 4s 74ms/step - mean_acc: 0.5181 - weibo_ner_acc: 0.9772 - BertMultiTaskTop/weibo_ner/losses/0: 0.0574
Epoch 5/20
48/48 [==============================] - 4s 74ms/step - mean_acc: 0.5119 - weibo_ner_acc: 0.9814 - BertMultiTaskTop/weibo_ner/losses/0: 0.0445
Epoch 6/20
48/48 [==============================] - 4s 73ms/step - mean_acc: 0.5135 - weibo_ner_acc: 0.9845 - BertMultiTaskTop/weibo_ner/losses/0: 0.0412
Epoch 7/20
48/48 [==============================] - 5s 95ms/step - mean_acc: 0.5121 - weibo_ner_acc: 0.9863 - BertMultiTaskTop/weibo_ner/losses/0: 0.0391

不知道是否tf和transformers版本导致.

我测试的版本:

tf.__version__: 2.6.2
transformers.__version__: 4.19.2