finetune adapointr on PCN vs costume dataset

[m-kafiyan]

Hi,
First of all, thanks for the great work and for sharing this project.

I'm fine-tuning AdaPoinTr (trained on the PCN dataset) on my own custom dataset. Before opening this issue, I carefully read through most of the existing ones but couldn't find a satisfying answer to my concern.

Here’s what I’m observing:

• During fine-tuning, the scale of the loss looks different. For example, in dense the loss starts around 28, while in another it starts closer to 4 (considering the loss is multiplied by 1000). Is such variation expected or natural?
• Also, I noticed a gap between the dense training loss and the dense test loss. Surprisingly, the test loss is lower than the train loss. I saw a few issues discussing similar results, but I’m still not convinced why this would happen.

Could you please clarify:

1. Is this behavior normal or a red flag?
2. If this is expected, what’s the reasoning behind it?
3. How can I reduce the dense test error further? Would adding weighting terms to the loss help?

As a sanity check, I also fine-tuned AdaPoinTr on the original PCN dataset, and I observed similar results there too.

Any insights or suggestions would be greatly appreciated!

Here is the results for my costume dataset

and this is the results of fine-tuning over pcn datasset