Strange curvature loss

[nicetryzz]

Thanks for your work! I test the meeting room in tank and temples datasets. The training curvature loss is strange. It increases a lot when reduces curvature weight. Should I increase the curvature weight?

curvature loss:

I test in 4 16G Tesla P100. Parameter as below:

• dict_size: 21
• dim: 8
• training batch size: 2

[mli0603]

Hi @nicetryzz

This is expected. As the numerical gradient step becomes smaller, more geometric details are generated due to coarse-to-fine control, which means larger curvature weight. If you want smoother surface, increasing the curvature weight would help at the cost of details. I hope this helps!