shaofanl / cyclegan-keras Goto Github PK

this project is great, i just search for a CycleGAN implemented in keras, when i run the demo, i encounter a error "expected input_3 to have shape (None, 3, 128, 128) but got array with shape (1, 128, 128, 3)" accured in "fake_A_pool.extend(self.BtoA.predict(real_B))" , it seems there are some problems with the array dimensions conversion, can you confirm it, thank you very much.

In resnet.py:

def resnet_6blocks(input_shape, output_nc, ngf, **kwargs):
ks = 3
f = 7
p = (f-1)/2 --> p = (f-1)//2
to avoid float argument causing error later

isn't there freezing of the D weights in your implementation when training G , or am I missing it?

I found in cyclegan.py, you are using opt.idloss as the loss weight of identity loss

if opt.idloss > 0:
            G_trainner = Model([real_A, real_B], 
                     [dis_fake_B,   dis_fake_A,     rec_A,      rec_B,      fake_B,     fake_A])
            
            G_trainner.compile(Adam(lr=opt.lr, beta_1=opt.beta1,),
                loss=['MSE',        'MSE',          'MAE',      'MAE',      'MAE',      'MAE'],
                loss_weights=[1,    1,              opt.lmbd,   opt.lmbd,   opt.idloss  ,opt.idloss])

but, in the original pytorch version, they are using

if lambda_idt > 0:
            # G_A should be identity if real_B is fed.
            idt_A = self.netG_A(self.real_B)
            loss_idt_A = self.criterionIdt(idt_A, self.real_B) * lambda_B * lambda_idt  # loss part?

So I think the weight of identity loss should be opt.idloss*opt.lmbd?

Hi,

Thanks for your valuable code. But I don't quite follow the logic for calculating the gan loss in the cyclegan.py. My understanding is the discriminator tries to tell the real one is real(1), fake one is fake(0). so in cyclegan.py

            for _ in range(opt.d_iter):
                _, D_loss_real_A, D_loss_fake_A, D_loss_real_B, D_loss_fake_B = \
                    self.D_trainner.train_on_batch([real_A, fake_A, real_B, fake_B],
                        [zeros, ones*0.9, zeros, ones*0.9])

should be

            for _ in range(opt.d_iter):
                _, D_loss_real_A, D_loss_fake_A, D_loss_real_B, D_loss_fake_B = \
                    self.D_trainner.train_on_batch([real_A, fake_A, real_B, fake_B],
                        [ones*0.9, zeros, ones*0.9, zeros])

While the generator tries to fool the system to treat fake one as real. so in cyclegan.py

                _, G_loss_fake_B, G_loss_fake_A, G_loss_rec_A, G_loss_rec_B = \
                    self.G_trainner.train_on_batch([real_A, real_B],
                        [zeros, zeros, real_A, real_B, ])

should be

                _, G_loss_fake_B, G_loss_fake_A, G_loss_rec_A, G_loss_rec_B = \
                    self.G_trainner.train_on_batch([real_A, real_B],
                        [ones, ones, real_A, real_B, ])

Have you tried same experiment e.g. horse2zebra using this code? How many epochs can you obtain reasonable results? I tried to run the code with tensorflow backend with slightly modification but never obtain correct results. I am wondering if this logic mentioned above has something to do with it. Please advice. Thanks.

ValueError: Dimensions must be equal, but are 3 and 4 for 'gen_A/instance_normalization2d_21/mul' (op: 'Mul') with input shapes: [1,3,1,1], [?,4,128,64].