size mismatch when loading pretraind model in mmdet3d framework[Solved]

[beyoursel]

Hello, thank you very much for your excellent work. I want to import LitePT to the MMDet3d framework and use pre-trained weights of backbone, but when I load the model weights, I find that the sparse convolution weight sizes are mismatched, the warning is as follow. My current spconv version is spconv-cu118. Could this be related to the spconv version, or have you encountered a similar problem? I look forward to your reply!

[ywyue]

Hi @beyoursel, thanks for your interest and kind words! For object detection, in line with previous approaches (SST, FlatFormer, PTv3), we avoid spatial downsampling, thus turning LitePT into a single-stage network with 8 blocks, to help detection of small objects. Our object detection codebase is built on top of the OpenPCDet framework, and I plan to integrate that into current codebase though this will take some time. If you want to access the object detection codebase at this stage, feel free to let me know your email address and I will share that with you.

[beyoursel]

Thank you for your reply ! Now I figured out my problem when loading the litept pretrained model in mmdet3d, which maybe is a bug of Spconv. I can load the model checkpoint after reshaping the shape of the weights.

[beyoursel]

Hi, thanks for sharing this great work, could you send me the code for detection. my email address is taole19980101@gmail.com