Using the calculation comparison loss method in the source code, the calculated loss may be negative, My input is(batchsizedim, batchsizeclass_num*dim, class_num),And Lz and Lθ may be negative at the same time
Hi, thank you for your exciting work.
When I want to save model get this error:
Transformer object has no "save_pretrained"
How to save the model after we train it with your code?
In fact, I want to save the model and upload it in Huggingface so that I can load and use it later.
Hi thank you for your exciting work. I've noticed a potential problem regarding to the evaluation procedure. To best of my knowledge currently the best model is selected based on the test data. However this is not desirable since in real conditions it is not possible to chose the model based on the testing data. One probable issue rather than getting comparable performances is possibility for overfitting. Altough test data is not used for gradient updates, model is chosen based on the best performing test data. Therefore, we have no way of knowing if the proposed model is just better at leaking the information via model selection. One extreme case is if you randomly guess enough times on test set you can get 100%. That's generally why the validation split is used in prior works 1.
Dear author, your framework is valid on the English dataset, but when I used dual-loss deficiency on my Chinese dataset, gradient collapse occurred. My Chinese label is two characters, is it related to this? Or do I have to adjust somewhere? Thank you very much. Look forward to hearing from you soon
Your work is very good and effective. But I have some questions about the baseline approach. I tried different hyperparameters to adjust supervised contrastivelearning or unsupervised contrastive learning to fine-tune BERT, and then to classify. But I've never been able to do anything better than just Cross-Entropy. I wonder what I didn't take into account? I've seen a lot of papers that contrastive learning can help improve classification results, but here I always get the opposite. Maybe I want to know the hyperparameters you set when you ran the comparison.
Hi there!
I think these code and paper awesome!
when I run this code, I can see increasing accuracy.
but I want to see moving that the class representation and sentence feature representation too.
Could you please upload the tSNE visualization code to github as well?
It's a great job. But I have a question, why do you use DualCL to perform out-of-order operation specifically for labels? This operation will not change the real label in binary classification, but it will change the real label in multi-classification. I don't understand the significance of this.
In fact, I followed this setup and then trained it on my own dataset, a binery classification task like dialogue intention recognition, and trained it for 30 epochs using Roberta, with very poor results, isn't DualCl suitable for this kind of task? I hope you can help me to point out my misunderstanding.