about lesson8:DCGAN

[backkon]

Why do I get the log as shown below:

[0/20][0/32]    Loss_D: 1.8010  Loss_G: 8.0963  D(x): 0.4317    D(G(z)): 0.3617 / 0.0005
[0/20][10/32]   Loss_D: 7.5290  Loss_G: 38.1183 D(x): 0.9269    D(G(z)): 0.9958 / 0.0000
[0/20][20/32]   Loss_D: 0.0425  Loss_G: 58.4481 D(x): 0.9731    D(G(z)): 0.0000 / 0.0000
[0/20][30/32]   Loss_D: 0.2492  Loss_G: 57.3842 D(x): 0.9787    D(G(z)): 0.0000 / 0.0000
[1/20][0/32]    Loss_D: 0.0000  Loss_G: 56.7433 D(x): 1.0000    D(G(z)): 0.0000 / 0.0000
[1/20][10/32]   Loss_D: 0.0000  Loss_G: 56.9849 D(x): 1.0000    D(G(z)): 0.0000 / 0.0000
[1/20][20/32]   Loss_D: 0.0001  Loss_G: 56.8024 D(x): 0.9999    D(G(z)): 0.0000 / 0.0000
[1/20][30/32]   Loss_D: 0.0000  Loss_G: 57.0740 D(x): 1.0000    D(G(z)): 0.0000 / 0.0000

The outputs have always been noise, I am running according to the source code, where is the problem?

[zhangxiann]

I haven`t encountered this problem. here is my log:

[0/20][0/32]    Loss_D: 2.2277  Loss_G: 15.5083 D(x): 0.5273    D(G(z)): 0.6209 / 0.0000
[0/20][10/32]   Loss_D: 0.2962  Loss_G: 46.8450 D(x): 0.9128    D(G(z)): 0.0000 / 0.0000
[0/20][20/32]   Loss_D: 0.0020  Loss_G: 52.9141 D(x): 0.9982    D(G(z)): 0.0000 / 0.0000
[0/20][30/32]   Loss_D: 0.2574  Loss_G: 51.1907 D(x): 0.9840    D(G(z)): 0.0000 / 0.0000
[1/20][0/32]    Loss_D: 0.0000  Loss_G: 51.1631 D(x): 1.0000    D(G(z)): 0.0000 / 0.0000
[1/20][10/32]   Loss_D: 0.0005  Loss_G: 51.2719 D(x): 0.9995    D(G(z)): 0.0000 / 0.0000
[1/20][20/32]   Loss_D: 0.0001  Loss_G: 50.5107 D(x): 0.9999    D(G(z)): 0.0000 / 0.0000
[1/20][30/32]   Loss_D: 0.0000  Loss_G: 50.6817 D(x): 1.0000    D(G(z)): 0.0000 / 0.0000
[2/20][0/32]    Loss_D: 0.0000  Loss_G: 50.4814 D(x): 1.0000    D(G(z)): 0.0000 / 0.0000
[2/20][10/32]   Loss_D: 0.0000  Loss_G: 50.3472 D(x): 1.0000    D(G(z)): 0.0000 / 0.0000
[2/20][20/32]   Loss_D: 0.0000  Loss_G: 50.2126 D(x): 1.0000    D(G(z)): 0.0000 / 0.0000

Perhaps you can try to run more epochs.

[backkon]

Your log is about the same as mine. Why is D(G(z)) 0/0? The second term of D(G(z)) should be getting closer and closer to 1?

[zhangxiann]

• D(G(z)) denotes the predicted probability of generate(fake) images.
• D(x) denotes the predicted probability of real images. I think, for the discriminator，it should try to distinguish the real and fake images. So the D(x) is getting closer to 1, the D(G(z)) is getting closer to 0.

[backkon]

I got results as shown below:

[zhangxiann]

Wait, I will try to find out waht`s happening.