Got abnormal loss

[alphabet-lgtm]

I trained with custom data and observed a very high loss. Do you have any suggestions? I tried setting instance_classes to 2 or 1, but it didn't help.

Additionally, I noticed that the coordinate size of my custom data is around 1000, whereas the official data has coordinates around 10. Should I resize the coordinates?

[alphabet-lgtm]

I got a lower loss about 100 by resizing the coordinates but it still seems a bit high. And I still have a question: What is the relationship between semantic_classes, instance_classes, CLASSES, and BENCHMARK_SEMANTIC_IDXS? It confused me a lot. If it’s convenient, could you please tell me?