Arxiv10 dataset split and code issues

[LaviRoars]

Hi Ashkan,

1. I'm trying to reproduce the Arxiv10 test results for my learning but the dataset shared on your github page does not specify the train(80,000), validation(10,000), and test split(5,000). The example in the dataloader.py code was for IMDB.csv.

2. Also, there is some issues with the code base. In trainer.py, df_embeddings was not defined anywhere at all. Plus, I can't seem to locate the multi-objective self-learning part that uses the similarities of embedding prototypes to train the Protoformer FW after fine tuning. Could you please point me in the right direction?

[ashfarhangi]

Hi, please use the sklearn.model_selection import train_test_split to perform 5 fold cross validation. The training time took around 1:45h on A100 GPU. You also might see the following accuracy if there is not enough computational resources to perform the 5 fold.

[LaviRoars]

Thank you, Ashkan. Circling back to question #2 above, I still could not replicate the training. Could you please share the latest version of the code that runs without error? Highly appreciate your help.

[sunying00]

谢谢你，Ashkan。回到上面的问题 #2，我仍然无法复制训练。您能否分享运行无误的最新版本的代码？非常感谢您的帮助。

您好，我也同样遇到了如您所说的问题#2，请问您现在对这个问题有解决的办法吗？感谢您的帮助。

[sunying00]

您好，这段代码说明为Protoformer的开源代码，可是为什么在代码中并没有Protoformer神经网络模型的定义呢？