Three questions about this model ~

[redinton]

1. Is this model fixed? Because when the training of this model is finished, users can only change the output by changing the "start token"?
2. What if I want to generate something else?
3. Compared to the raw-gan in CV, people can use the gan to generate the picture from some noise input. So can we use this SeqGAN model to do that? Generate the sentence from some noise input?

[TobiasLee]

1. You can give the model some "hint" by changing the start token or the initial hidden state, which is initiated all zeros.
2. Change the positive file to some others, like word tokens of Tang poems as the paper described, instead of tokens generated by oracle to implement your own SeqGAN.
3. The noise in CV such as gaussian distribution is used to imitate the real data distribution. GAN minimizes the KL-divergence of generated data and real data to achieve the goal of producing vivid images. However, since the text distribution may be much more complicated than image, I'm not sure whether it is effective to do that, but it is worth exploring.