denizetkar / treegan Goto Github PK

Hi, thanks for you share. I'm doing a similar research on code generation and I also tried this model. But I haven't got a useful result yet. What I find is that generator only works on first 50 timesteps, and then it starts to generate some meaningless codes. So could you please share some details about this model, for example, the generated results, training settings and parameters? I think it is the problem of policy loss. Even in Seq GAN, they used sequences of length 20 to train the model, so it may be hard for the Tree GAN to work on a longer sequence. In my case, the median and mean length of training data are 837 and 900 respectively. Thanks.