The codes of paper "Long Text Generation via Adversarial Training with Leaked Information" on AAAI 2018. Text generation using GAN and Hierarchical Reinforcement Learning.
GAN models usually generate samples from latent space. It enables the GAN to generate distinct samples and generate samples based on some information. However, in your paper and your code, I cannot find a latent space input. In the "Generation Process" sub section of the paper, you claimed that the Manager and Worker both start from zero hidden state. And also, I found that in your code, the initial g(goal vector) is set to a trainable tensor, which means in evaluation or a single training batch it's fixed. Moreover, the initial input, say x_0, is fixed. These are common latent space inputs in RNN GANs, but none of them is set as latent vector input in your model, and this fact really confuses me.
GAN models usually generate samples from latent space. It enables the GAN to generate distinct samples and generate samples based on some information. However, in your paper and your code, I cannot find a latent space input. In the "Generation Process" sub section of the paper, you claimed that the Manager and Worker both start from zero hidden state. And also, I found that in your code, the initial g(goal vector) is set to a trainable tensor, which means in evaluation or a single training batch it's fixed. Moreover, the initial input, say x_0, is fixed. These are common latent space inputs in RNN GANs, but none of them is set as latent vector input in your model, and this fact really confuses me.