About the design of non-translation

[xiaolongcheng]

In your paper, you have designed non-translation , self-translation, and cycle-translation. In non-translation generator, it just encode the input, e = encoder(input), and then decode e, output = decoder(e). Certainly you have already achieve great results with HiSD designed like this. I am a beginner of GAN, and I am just wondering why don't you set stylecode as 0, and use a translator in your non-translation, which means, e = encoder(input), e_trg = translator(e, 0), output = decoder(e_trg), I have seen other works did like this. Could it make any difference?

[imlixinyang]

1. use e_trg = translator(e, 0) can not help the training because that our style is absolute while the style or label of methods you mentioned (such as STGAN or RelGAN) is relative. However, it's worth if you have a try and edit the cycle-translation to be e_trg = translator(e, s_trg - s_src) if you want to guide the non-translation by 0.
2. The non-translation and self-translation paths are designed for the consistency of features before and after passing through the translator (please see modulargan) and therefore implicitly support multi-tag task.

[xiaolongcheng]

Thank you for your reply!