meet the issue KeyError: 'backbone.layers.0.blocks.0.attn.relative_position_bias_table' about swin-transformer-semantic-segmentation HOT 6 CLOSED

@panda-2014 I also met the same problem. This is because of the inconsistency of the state_dict.keys(). Strip the prefix of state_dict can solve the problem. I've solved the problem by adding the following codes in mmsegmentation/mmcv_custom/checkpoint.py:line 322
if list(state_dict.keys())[0].startswith('backbone.'): state_dict = {k[9:]: v for k, v in state_dict.items()}

from swin-transformer-semantic-segmentation.

Hi, "upernet_swin_base_patch4_window7_512x512.pth" is a model finetuned on ADE20K, not an ImageNet pre-trained model.
You would better use pre-trained models held in this page.

from swin-transformer-semantic-segmentation.

@panda-2014 Hello, could you tell me the run code when training on one gpu on windows(RTX 3090)? I need it as an example to train the model.

from swin-transformer-semantic-segmentation.

@panda-2014 Hello, could you tell me the run code when training on one gpu on windows(RTX 3090)? I need it as an example to train the model.

I run on linux, and the I have change to run on rtx8000x4 GPU, and occur some new issues. I still have no idea about the issue :(

from swin-transformer-semantic-segmentation.

Do you have solved this problem?I also met the same problem
the model and pre-trained weight is right，I checked and tried the others

from swin-transformer-semantic-segmentation.

512 pretrained model can be found here: https://github.com/open-mmlab/mmsegmentation/tree/master/configs/swin

from swin-transformer-semantic-segmentation.