Loss not decreasing

[SrinidhiBharadwaj]

Is it expected to see the loss not go down during training? Loss kind of stagnates after the first epoch which does not make any sense. What am I missing?

[WuTanKun]

I encountered the same problem, where the score loss did not decrease, but the sdf loss did decrease, and I have trained for 10,000 steps. I look forward to the author's answer and other ideas.

[robotgradient]

In our implementation, the loss function for the diffusion part was not very trustworthy.

Did you try running the model and visualize the outcome? In our experience, even if the loss was not decreasing much, the generation was good.

[WuTanKun]

Thanks for your response! I have run the model and visualized the results. Although the diffusion loss is not reduced much, the results seem to perform well. I will do more exploration.