About the performance on Semantickitti.

[a18700]

Hello, thanks for amazing works.

Do you have any idea on what is the main factor for the degradation of performance when to exploit pytorch codes?

I want to boost your model performance to the level of the one of paper.

[qiqihaer]

Hello, thanks for amazing works.

Do you have any idea on what is the main factor for the degradation of performance when to exploit pytorch codes?

I want to boost your model performance to the level of the one of paper.

I have no idea about the degradation. As for the performance, I think the performance on the validation set is similar to the origin paper. Maybe you can try to train more epoches and add some different augmentation methods.

[a18700]

Thank you for reply!

Except for the performance, i have an issue on the memory costs of this implementation.

Can i ask how much memory is used for experimenting on semantickitti on default setting?

I'm experimenting on 8gb rtx2080 with batch size 6, and even i reduced it to 2 but still OOM errors come up.

[qiqihaer]

8G seems small for the default setting. You can conduct one back propagation for several iterations.