Training does not converge - black image

menyifang / ADGAN

The Implementation of paper "Controllable Person Image Synthesis with Attribute-Decomposed GAN" CVPR 2020 (Oral); Pose and Appearance Attributes Transfer;

476 stars 90 forks source link

Training does not converge - black image #27

Open krips89 opened 3 years ago

krips89 commented 3 years ago

After a certain number of epochs, I am getting black image as output as shown in the image. The training seems to be converging in the beginning; however, it fails after a few epochs.

The losses also explodes during that time. Is this something known?

krips89 commented 3 years ago

Update: I get this black images when I train with the split used in many of the previous works (PoseGan, Coordinate Based Inpainting, DensePose Transfer, Neural Human Re-Rendering).

However when I train with the split used in this project, it seems to converge fine. Why was this new split introduced? What could be a good explanation? So much instability with different splits?