Question about the back translations. about mixtext

[jiaaoc]

We use back-translation to create paraphrases for unlabeled data and perform consistency training. You could use other ways to generate paraphrases.

[callmeYe]

So I have to create paraphrases, right? In addition, When I look at the code, I find that only the first 100,000 pieces of data in the data set have been back translated. Do I not need to perform back translation for all the datasets？

[jiaaoc]

It depends on the size of the unlabeled data you are going to use. In this work, we used 100,000 unlabeled data, so we just did back translations on them, not the whole dataset.

[callmeYe]

Sorry, I'm still a little confused.
When I test with:
python ./code/train.py --gpu 0,1 --n-labeled 10 \ --data-path ./data/yahoo_answers_csv/ --batch-size 2 --batch-size-u 4 --epochs 20 --val-iteration 1000 \ --lambda-u 1 --T 0.5 --alpha 16 --mix-layers-set 7 9 12 \ --lrmain 0.000005 --lrlast 0.0005
The number of unlabeled data per class seems to be 5,000. Do they add up to exactly 100,000?

[jiaaoc]

You could use up to 100,000

[jiaaoc]

10,000

[jiaaoc]

Anyway, the number of data you need to paraphrase only depends on the number of unlabeled data you are going to use.

[callmeYe]

Are they one-to-one correspondence？

[jiaaoc]

one unlabeled data could be associated with multiple paraphrases. Please refer to the paper/codes for details.