about the pretrained ConvNeXt model
wyf0414 opened this issue · 3 comments
Hi, thanks for open sourcing this brilliant work!
I read your paper in detail. You use ConvNeXt as the pretrained model of PointPillars. This is very pioneering.
I want to know:
Does the pretrained model has the same architecture as your lidar backbone?
Does the pretrained model also has large kernal (like 7*7 in the paper) ?
If possible, can you provide the way to download your pretrained model?
I'm so sorry for the late response. Recently, I've been busy with my doctoral thesis.
Sure, the pretrained weights include a convolution with a 7x7 kernel.
If you need the weights immediately, the ConvNeXt weights I used are from mmcls==0.22.1.
More specifically, the image weights are pretrained on ImageNet-1k.
https://github.com/open-mmlab/mmpretrain/tree/master/configs/convnext