Fine-tuning ShapeHD always generates same weird result

[rozentill]

Hi,

I am fine-tuning shapeHD with a pre-trained 3D GAN, but it seems all the test model will lead to this result.

Any idea about this?

Thank you!

[ztzhang]

this might be an artifact due to the bad weight balancing between the Gan loss and the cross entropy loss. The way we balance those two losses is looking at corresponding gradient variance. Maybe @xiuming can share more insights into this.

[xiumingzhang]

Is this a visualization issue? Looks to me the voxels to the cube sides are all occupied. Could be that your threshold is too low, or you took log when you shouldn't, or you didn't take log when you should?

[rozentill]

Thanks, I changed the weight to 1e-5 and works now.