search

PJLab-ADG / LoGoNet

[CVPR2023] LoGoNet: Towards Accurate 3D Object Detection with Local-to-Global Cross-Modal Fusion
246 stars 16 forks source link

The pretrain weight issues #31

Open Orbis36 opened 11 months ago

Orbis36 commented 11 months ago

Hi there,

Thanks for your work first. Basically, I noticed that you freeze the FPN and Swin-T when training the network. However, I think the weights you mentioned didn't contain the weights for the FPN part. So how can you get it?

Orbis36 commented 11 months ago

I think maybe you gave the wrong link in that issue. Considering your codes from 51 to 61 in file centerpoint_MM_waymo.py, You should use the weight pre-trained on COCO instead of ImageNet.

Z-Lee-corder commented 11 months ago

I think maybe you gave the wrong link in that issue. Considering your codes from 51 to 61 in file centerpoint_MM_waymo.py, You should use the weight pre-trained on COCO instead of ImageNet.

I also found this question. But I also found another question: If we want to use pretrained image network parameters, we need uncomment "IMGPRETRAINED_MODEL: '../output_s2/det_model_cfgs/pretrained_ckpt/swin_tiny_patch4_window7_224.pth" in yaml file. Although a part of parameters of image backbone have been predefined and frozen, but they are not imported into image network. in line 62 of "anchor_MM_kitti.py": self.img_backbone.load_state_dict(new_ckpt, strict=False)

When I debug to this position, I found "new_ckpt" is empty. That is to say, the pretrained image network parameters were not imported into the network. The frozen image parameters that are still randomly defined parameters at the beginning of training.

I don't know if my understanding is correct?