Mihnea Aleman comments

Results 11 comments of


                                            Mihnea Aleman

test augmentation in scannet segmentation

Hi, I am not the author, but I will try to explain it. You do not need to do test-time augmentation, but sometimes it does help to offer a more...

mIoU

Me personally, I like to think that one dataset is treated as one single BIG scene.

CUDA Error: Device-Side Assert Triggered

Mine crashes after 3 epochs on the same line `self.scaler.scale(loss).backward()`.

How to run code saved in an exp file?

If you run the provided bash script, it will use the code from that specific folder.

Is it workable on Colab?

I have the same issue. However I dont think it is becasue collab. I think it might because of the spconv library, because I tried running PTv2 and it works....

Is it workable on Colab?

I actually used the same environment. I noticed that I get the same issue for scannet. I downgraded to cuda 11.8 and it works. Try using PTv2 and see if...

How to run multi-node training ?

Thanks for the reference. I will try it and see if it works.

Where can I get the preprocess data about ScanNet?

Usually you can find the preprocessed data here: https://huggingface.co/Pointcept However, I noticed too that they removed the scannet preprocessing data.

TypeError: list indices must be integers or slices, not numpy.float64

I will add the change later today and create a PR.

TypeError: list indices must be integers or slices, not numpy.float64

I see now why this was happening to me and you are right! > Aha, for PPT, I rewrote the trainer and only used ConcatDataset datasets to initialize multiple data...