The kld loss in UPDATE GENERATOR process

[MapleSpirit]

I noticed that you use the kld = torch.mean(-z_log_stddev + 0.5 * (torch.exp(2 * z_log_stddev) + torch.pow(z_mean, 2) - 1)) in UPDATE GENERATOR . But I don't understand that why you chose this as a part of your loss function. And it seems that it was not mentioned in the original paper. Could you please tell me its intention here? And apart from this, the z_log_stddev and z_mean are just got from two different Linear+LeakyReLU layer. Emmm... why did you use the Linear+LeakyReLU layer rather than calculate mean and std directly?

Thanks for your help~

[woozzu]

@MapleSpirit It is actually mentioned in the original paper. The authors used the term 'text embedding augmentation', which is same as the conditioning augmentation in StackGAN. Also, I personally got a part of the original implementation from the authors. Therefore, I have no opinion about this specific implementation choice.

[MapleSpirit]

OK~ I'll go to see the contents in StackGAN. Thanks again~