Question about auxiliary_head

[ChenJiangxi]

Thank you for your wonderful framework!
I have a question about the auxiliary decoder. For a auxiliary FCNdecoder to HRNet-w48, if I set in_index = [0,1,2], and in_channels=[48,96,192], does that mean the FCNdecoder get the output from stage3 of HRNet-w48?

Thanks in advance!

[MengzhangLI]

What is current config file, could you give me a link?

[ChenJiangxi]

There is my config file. Thanks for your reply.
https://github.com/ChenJiangxi/segmentation-on-singlegreen/blob/main/fcn_hr48_512x1024_80k_singlegreen_aux.py

[ChenJiangxi]

And I have some problems on other issue #1040
I'm confused about this answer. The config file for FCN decoder and HRNet is:
in_channels = [48, 96, 192, 384]
so I think the decode_head could only support four feature maps as input. Am I right?

[MengzhangLI]

No, you can set

in_channels=[18, 36, 72],
in_index=(0, 1, 2),
channels=sum([18, 36, 72]),

And theorically it could work (but it means you need to train new model on your own and I think it is necessary).

the function of it is concating all feature maps (after resizing to similar height and width) from backbone (HRNet-w48 in this case) in here:

https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py#L173

whose channels = #Chennel_featuremap1 + #Chennel_featuremap2 + #Chennel_featuremap3 + #Chennel_featuremap4.

See here channels=sum([18, 36, 72, 144]),.

[ChenJiangxi]

I think my set is correct accordingly, which is suitable for HRNet-w48. And yours is suitable for HRNet-w18
in_channels=[48,96,192], in_index=(0, 1, 2), channels=sum([48,96,192]),
oh god! it's embarrassing, I fogot that in_index means Input feature index. so the decoder head just get the feature map from stage4 of hrnet.
Thank you for your late night reply! Sorry to bother you.

[MengzhangLI]

I totally do not suggest you modify the number of feature maps for input of decoder/auxiliary heads.

As for deeplabv3_r50-d8, its decoder in_index=3 and auxiliary in_index=2, which is index of feature maps of backbone (i.e., its backbone has output [OUTPUT0, OUTPUT1, OUTPUT2, OUTPUT3], OUTPUT2 is input of auxiliary head and OUTPUT3 is input of decoder head).