Question about mmseg pre-trained weight loading

[endaoguansanlu]

Hello, can mmseg load two pre-training weights, for example, a two-branch model, one branch is segformer, one branch is resnet, I will load the two pre-training weights to this two-branch model and then perform it on cityscapes segmentation.Looking forward to your reply, thank you!

[HAOCHENYE]

Hi, you can load different pre-trainied weights as the tutial shows:
https://mmengine.readthedocs.io/zh_CN/latest/advanced_tutorials/initialize.html#id3

We highly recommend you to experience MMEngine, which is next-generation training framework of Openmmlab. Since we haven't made many changes to the initialization module yet, the content of the tutorial also applies to MMCV.

Since SegFormer and ResNet are also the subclass of BaseModules, you can define your model like this:

model=dict(
    branch1=dict(type='SegFormer', init_cfg=dict(type='Pretrained', checkpoint='path/to/checkpoint1'))
    branch2=dict(type='ResNet', init_cfg=dict(type='Pretrained', checkpoint='path/to/checkpoint2'))
)