What does the "auxiliary_head" mean?

[baibaidedongdong]

I have saw the config file ,I find it has a "auxiliary_head" in every models, but I dont know what it meas, why not only use the "decode_head" ?

[xvjiarui]

Hi @baibaidedongdong
It is the C1_deepsup setting in here. It is a deep supervision trick to improve the accuracy.

[rubeea]

Hi @baibaidedongdong
It is the C1_deepsup setting in here. It is a deep supervision trick to improve the accuracy.

I tried to train my model without the aux head and there isn't any decrease in any of the evaluation parameters (dice, iou, accuracy) apparently as compared to the model trained with the aux head (mAcc parameter screesnhot attached for reference). Moreover, I wanted to ask that these parameters are calculated on the validation dataset or the train dataset with workflow= [('train')] and the following code:

train_segmentor(model, datasets, cfg, distributed=False, validate=True, 
                meta=meta)

I believe that with validate=True, the evaluation hook is registered on the validation dataset and the parameters such as mIoU or mDice are calculated on the validation set. Is that correct?

Thanks in advance

blue: without aux head
orange: with aux head (deep supervision trick)