Confusions about Joint_W.py under folder Two_Circle

[yixiaoshenghua]

Compared with the pseudo codes presented in the paper, the training process in Joint_W.py is a little different.

1. In paper the models take the Adam optimizer and in code it utilizes the RMSProp.
2. In paper, it takes 5 rounds to train the stein critic, but in code it just train one round.
3. In paper, the sample generator's loss is consist of two parts, but in code it just minize the discriminator loss term.
   Can you give me answers about these confusions? Thank you very much!

[qitianwu]

Hi, thank you for the interests in our work. The pseudo code presented in our paper is for the general case (using the neural Stein critic) which is implemented with CIFAR experiment. For the two synthetic datasets (Two-Circle and Two-spiral), as stated in our paper, we use the kernelized Stein method whose implementation (the optimizer, the training scheme and loss computation) is a bit different.