Data to update G and D net is of the same batch?

[lhao0301]

As is known, data to update G and D net is of two different batches respectively in the classic gan (here I just consider one of these two gan nets). It's my first time to try TFRecord and so I have one question about it. Is the data of same batch here. If not, I guess data is updated once whether optimization of G or optimization of D is happened?

[vanhuyz]

Data is updated once after both optimization of G and optimization of D are finished here. But note that inputs of G and D are different. Data for G is a image obtained from TFRecord, but fake data for D is obtained from history of generated images.

[lhao0301]

Sorry, my question may be ambiguous. In one iteration, assume that the input data for G is Z_0 and the corresponding input data for D is X_fake (X_fake is generated from Z_0) and X_real, then we input X_fake and X_real to D for updating D. To update G, we still input X_fake generated from Z_0 to D and calculate the loss. The same Z_0 is used during the update of D and G net. In other word, we use one same batch of Z for the update of both G and D in one iteration. However, in classic gan, two different batch of Z are used to update D and G. Sample one batch Z (called Z_0 to distinguish ) used to generate fake data for updating D and then sample one other batch Z (called Z_1) for updating G in one iteration.

[vanhuyz]

I understand your point. Maybe for that reason, instead of using the same Z_0 every time, CycleGAN randomly uses a image Z_1 from history of generated images to update D in order to make the training process stabler.

[lhao0301]

Oh, it's right. According to the image pool, the input Z used to update D and G respectively is distinguished. It's same with classic gan (vanilla gan). Thank you very much.