Problem in using xlnet-base model

[so-hyeun]

Hi, I'm trying to use the DialogXL you provide, but I have a problem.

At first, I used Transformers 3.0.2 as shown in the requirement, but a Segment fault error occurred, so I used the latest version of transformers.
This solved the above error, but "UnicodeDecodeError:'utf-8' codec can't decode byte 0x80 in position 0" error occurred.
Can you tell me how to fix that error or a link to download xlnet-base-case for transformers 3.0.2?

Thanks for conducting an interesting research.

[shenwzh3]

Hi, could you show me the full error description? It seems that this error usually occurs when the program reading the dataset files rather than when loading the model.

[so-hyeun]

Thanks for the answer.
When I used Transformers 3.0.2, the segment fault error mentioned above was resolved, but the following error occurred.

[Digimonseeker]

Hi, @so-hyeun, thanks for the interest in our work，we have updated the relevant files, you can kindly check it out.

[so-hyeun]

Thank you. It works fine for eval.py code.

By the way, when IEMOCAP is trained as the command in the Readme, it shows remarkably low performance. (Valid F-Score: 4.42 Test F-Score : 9.44 Test Acc: 23.86)
What is the cause?

[Digimonseeker]

We found some problems with the previous IEMOCAP files and we have now updated the relevant files. Thanks for your friendly reminder.

[so-hyeun]

Hello, first of all, thank you for always quick feedback.

1. By the way, the same phenomenon occurs when running the train using the newly uploaded IEMOCAP file. Can you check it again?

2. In addition, when I change line 131 of run.py to model =nn.DataParallel(model) to use 2 GPUs, an error occurs. If you have used multi-GPU, please advise.

Thank you.

[tenihasina]

Hello, first I'd like to thank you for working on such an interesting subject

I am also trying to use the model provided in the repository :
Training with MELD gives good results (comparable to your article), unfortunately, I also encountered the same issue as @so-hyeun where training with IEMOCAP gives bad performance.

Did you manage to find what was the issue ?