Confusing about the Deepfillv2 training d_loss

JiahuiYu / generative_inpainting

DeepFill v1/v2 with Contextual Attention and Gated Convolution, CVPR 2018, and ICCV 2019 Oral

http://jiahuiyu.com/deepfill/

Other

3.27k stars 787 forks source link

Confusing about the Deepfillv2 training d_loss #467

Closed ewrfcas closed 3 years ago

ewrfcas commented 4 years ago

Sorry to bother you but I have some problems while reimplement training the deepfillv2 model. The d_loss is converged to 1.0 which means that the discriminator output the same value for both real and fake samples. Should I train the netD for few times (maybe 5?) after once train the netG? For some reason I can't use the tensorboard. Here is the manual loss pic.

The model results are not good

JiahuiYu commented 3 years ago

D_loss is not stable which is common in GAN training. I think the loss you plotted is reasonable.