Why the G and D train at the same time in your code?

YuanWu3 / Dual-Mixup-Regularized-Learning-for-Adversarial-Domain-Adaptation

Implementation of "Dual Mixup Regularized Learning for Adversarial Domain Adaptation" in Pytorch

https://arxiv.org/pdf/2007.03141.pdf

11 stars 5 forks source link

Why the G and D train at the same time in your code? #1

Open wangyujie413 opened 3 years ago

wangyujie413 commented 3 years ago

Hello, Thanks for your code. In the Generative Adversarial Network (GAN), it always maintain the parameters of the G while update the parameters of D firstly, and then maintain the parameters of the D while update the parameters of G. And the description of the training procedure in your paper follows this, but the code does not do as this training procedure. Why?

YuanWu3 commented 3 years ago

Hi, thanks for your attention to our work. I recommend a classic unsupervised domain adaptation paper: Domain-adversarial training of neural networks (https://dl.acm.org/doi/abs/10.5555/2946645.2946704). I adopt the training fashion proposed in this paper, we used a reversal layer to deal with the training signal propagated from the discriminator D such that we don't need to use the GAN loop.