About differences between configs on nuScenes and Argoverse2

[curiosity654]

Hi, thank you for your great work. As the FSF model on AV2 isn't released, I'm trying to train on AV2 myself. I found that the pretrained FSD model on nuScenes and AV2 are very different in terms of size (900+MB on nuScenes and 100+MB on AV2). Looking into the configs I found that the Sparse UNet backbone setting are very different, with nuScenes's backbone channel number are generally 2 times larger than on AV2. Is there any insight behind the design or just tuned by experience? Thank you very much for any feedback.

config of AV2:

backbone=dict(
        type='SimpleSparseUNet',
        in_channels=64,
        sparse_shape=[32, 2048, 2048],
        order=('conv', 'norm', 'act'),
        norm_cfg=dict(type='naiveSyncBN1d', eps=1e-3, momentum=0.01),
        base_channels=64,
        output_channels=128,
        encoder_channels=((64, ), (64, 64, 64), (64, 64, 64), (128, 128, 128)),
        encoder_paddings=((1, ), (1, 1, 1), (1, 1, 1), ((0, 1, 1), 1, 1)),
        decoder_channels=((128, 128, 64), (64, 64, 64), (64, 64, 64), (64, 64, 64)),
        decoder_paddings=((1, 0), (1, 0), (0, 0), (0, 1)),
    ),

config of nuScenes:

backbone=dict(
        type='SimpleSparseUNet',
        in_channels=64,
        sparse_shape=sparse_shape,
        order=('conv', 'norm', 'act'),
        norm_cfg=dict(type='naiveSyncBN1d', eps=1e-3, momentum=0.01),
        base_channels=64,
        output_channels=128, 
        encoder_channels=((128, ), (128, 128, 128), (128, 128, 128), (256, 256, 256), (512, 512, 512)),
        encoder_paddings=((1, ), (1, 1, 1), (1, 1, 1), ((0, 1, 1), 1, 1), (1, 1, 1)),
        decoder_channels=((512, 512, 256), (256, 256, 128), (128, 128, 128), (128, 128, 128), (128, 128, 128)),
        decoder_paddings=((1, 1), (1, 0), (1, 0), (0, 0), (0, 1)), 
    ),

[liyingyanUCAS]

Hello, this difference in model size is primarily due to a design choice aimed at reducing training time on the AV2 dataset. The AV2 dataset is larger and more computationally intensive, so the model's complexity and size were reduced to make training more efficient.