search

SamsungLabs / imvoxelnet

[WACV2022] ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection
MIT License
283 stars 29 forks source link

Question about backbone? #36

Closed rockywind closed 2 years ago

rockywind commented 2 years ago

Hi, @filaPro When I wanted to change the backbone to resnet18, I modified the code. But it ran error!

File "/home/CN/zizhang.wu/anaconda3/envs/imvoxelnet_small/lib/python3.7/site-packages/mmcv/utils/registry.py", line 182, in build_from_cfg raise type(e)(f'{obj_cls.name}: {e}') AssertionError: ImVoxelNet: ResNet: Not implemented yet.

My code is that.


    type='ImVoxelNet',
    # pretrained='torchvision://resnet50',
    pretrained='torchvision://resnet34',
    backbone=dict(
        type='ResNet',
        depth=34, #50,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        norm_cfg=dict(type='BN', requires_grad=False),
        norm_eval=True,
        style='pytorch',
        dcn=dict(type='DCNv2', deform_groups=1, fallback_on_stride=False),
        stage_with_dcn=(False, False, True, True)),```
filaPro commented 2 years ago

Hi @rockywind ,

Deformable convolutions are not supprorted for ResNet18 and ResNet34 in mmdetection. At least it is true about the version from our Dockerfile. So you can simply remove dcn and stages_with_dcn from backbone. Btw, you should also update in_channels in neck.

rockywind commented 2 years ago

OK @filaPro Thank you very much!