GANs with Multiple Constraints for Image Translation

Algorithm 1

MMGANs. We use default values of , , , , initial learning rate , and . Requirement is as follows: There are , , and samples in the , , and domains, the batch size is 1, and Adam optimizer hyperparameters are and . The multiscale parameter and the multilevel parameter .

(1) Initialization , , generator parameters , , , , , ,

discriminator parameters , , . Suppose = (, , , , , , , , ).

(2) while do

(3) Sample one sample at a time , , .

(4) The losses G1_loss, F1_loss, G2_loss, F2_loss, G3_loss and F3_loss of generators

are calculated according to formula (8).

(5) for i =1 to do

(6) for j =1 to do

(7) The losses DX_loss, DY_loss and DZ_loss of discriminators are respectively

calculated according to formula (1), (2), (3), (4), (5), (6).

(8) end for

(9) end for

(10) Adam( (G1_loss, F1_loss, G2_loss, F2_loss, G3_loss, F3_loss, DX_loss,

DY_loss, DZ_loss), , , )

(11) ++

(12) end while