search

Pointcept / PointTransformerV2

[NeurIPS'22] An official PyTorch implementation of PTv2.
356 stars 25 forks source link

Model weights of PTv2m1 #20

Closed chrockey closed 1 year ago

chrockey commented 1 year ago

Hi,

Thank you for releasing the source code of your excellent work :) Could you provide the model weights of PTv2m1, which is used for the camera-ready version of your paper on either S3DIS or ScanNetV2?

Gofinge commented 1 year ago

Hi, thank you for being interested in our work. Due to a change in internship, we no longer have access to the weights that were trained in the first half of last year. Additionally, the released model code has been completely rewritten from the prototype code to make it easier to understand, making it incompatible with the original weights. At the moment, we do not have access to computing resources that contain 4 x 48G memory GPUs. However, we will share a PTv2m1 training log as soon as these computing resources become available.

chrockey commented 1 year ago

Thanks for your timely reply.

Due to a change in internship, we no longer have access to the weights that were trained in the first half of last year.

I see. Actually, I trained the PTv2m1 model on ScanNetV2 validation set from scratch using this codebase for 450 epochs (half of the original config) and found that the results seem to be reasonable (mIoU = 74.9).

Thanks again for releasing this well-organized codebase :)

Gofinge commented 1 year ago

Repeating 2-3 times should have a higher performance than I reported. PTv2m2 is also well worth trying.