Why epochs is 5?

[gongshaojie12]

Hi,I noticed that the epochs is 5 during training, why not set it a bit larger? Is it because the epochs are set too large to cause overfitting? Thanks!

[finiteautomata]

@gongshaojie12 I just used a fairly standard number of epochs. I observed no increased performance by running them for a larger period of time.

You can try it to train them for more time and let me know if that yields better results. Ping me if you want to do so (particularly regarding downloading the data, I haven't documented that very much)

[gongshaojie12]

Hi @finiteautomata I set epochs to 30, eval_loss and train_loss are as follows:

As can be seen from the figure, from the third epoch, eval_loss starts to rise, while train_loss keeps falling. It shows that the model has been overfitting. What is the reason? I guess the reasons may be: 1, the data set is too small, 2, the pre-training model for fine-tuning is too large.

[finiteautomata]

@gongshaojie12 great, thanks for your experiments. I understand that you observed no performance increase by setting a larger number of epochs; moreover, it seems that ~5 epochs seems to be ok.

What dataset are you using? SemEval (English) or TASS(Spanish)?

[gongshaojie12]

Hi @finiteautomata The dataset I used is EmoEvent(dataset_emotions_EN.csv). The training command is: python bin/train_emotion.py "roberta-base" models/roberta-base-emotion-analysis/ --epochs 5 --lang en . Is the dataset used incorrectly?

[finiteautomata]

@gongshaojie12 that is the dataset used for English Emotion Analysis. As you can see, that dataset is small-sized – the same occurs with the Spanish version of EmoEvent, and the Sentiment Analysis in Spanish (TASS). We used 10 epochs for Sentiment Analysis in English as that dataset is bigger (around 50k instances)

[gongshaojie12]

Hi @finiteautomata Thanks for your reply. What if I want to get more EmoEvent data?

[finiteautomata]

@gongshaojie12 if you want to get more data, you should collect and annotate your own corpus. You can get more information on how the authors did so here https://github.com/fmplaza/EmoEvent-multilingual-corpus

[gongshaojie12]

OK,Thanks a lot

[finiteautomata]

Closing this. Feel free to reopen or send email if you have any doubt