Question about the config of `num_levels` of seg_head

[whystopped]

In the config of seg_head, num_levels is set to 4

UniAD/projects/configs/stage2_e2e/base_e2e.py

Line 266 in 2f38ff1

num_levels=_num_levels_,

UniAD/projects/configs/stage2_e2e/base_e2e.py

Line 286 in 2f38ff1

num_levels=_num_levels_,

but the input of seg_head is bev_embed, which has no multi-scale
so, maybe num_levels should be set to 1?

[YTEP-ZHI]

@whystopped. Thanks ! We will check this part of codes.

[ilnehc]

@whystopped Thanks for reporting the issue. The parameter should be 1. But 4 basically adds more reference points in practice which does not lead to error warnings of the network structure. We tested to set it to 1 and found that this modification would greatly improve the segmentation performance and slightly improve other tasks. I will update the checkpoints and logs in the future.

[Lewis-Lu]

Checked out the same error as this ticket when trying to deploy the segmentation part.

num_levels=4 results in cpu-version MS_Deform_Atten error. Should set num_levels=1

in mmcv.ops.multi_scale_deformable_attn_pytorch: output = (torch.stack(sampling_value_list, dim=-2).flatten(-2) *

the stack would result in incorrect shape (1 level) to that of the attention weight (4 levels).

Can I have the updated checkpoints for this modification ? Thanks in advance.

Best,
Lewis

[ilnehc]

@Lewis-Lu Updated in https://github.com/OpenDriveLab/UniAD/releases/tag/v1.0.1.