KeyError: "CascadeRCNN: 'backbone.layers.0.blocks.0.attn.relative_position_bias_table'"

[luooofan]

Thanks for your work!

I occured the error when I run the code.

I run the command:
python tools/train.py configs/swin/mydef_cascade_mask_rcnn_swin_small_patch4_window7_mstrain_480-800_giou_4conv1f_adamw_3x_coco.py --cfg-options model.pretrained=./models/cascade_mask_rcnn_swin_small_patch4_window7.pth

How to solve it?

[weiyx16]

I think the reason is the pretrained model you used is the one finetuned on COCO det, not the one from ImageNet pretrain. You'd better use the corresponding model held on this page.

[luooofan]

I've tried to print some infomation at there thanks to your reminding.

...
relative_position_bias_table_keys = [k for k in state_dict.keys() if "relative_position_bias_table" in k]
# insert code
print(relative_position_bias_table_keys)
print([k for k in model.state_dict().keys() if "relative_position_bias_table" in k])
...

And I get these:

[luooofan]

I think the reason is the pretrained model you used is the one finetuned on COCO det, not the one from ImageNet pretrain. You'd better use the corresponding model held on this page.

Maybe this is the keypoint, I'll try it.

[luooofan]

It works. Thanks.

[weiyx16]

Good to know!

[JHLEE17]

Thank you. @weiyx16
It works for me too.
but, I wonder why using COCO pre-trained weights is not available.

because I'm trying to train custom detection dataset on Swin Transformer-Object Detection.
It was success that training custom dataset using ImageNet pre-trained weights but,
it wasn't success that training using COCO pre-trained weights.

[hosea7456]

Thank you. @weiyx16
It works for me too.
but, I wonder why using COCO pre-trained weights is not available.

because I'm trying to train custom detection dataset on Swin Transformer-Object Detection.
It was success that training custom dataset using ImageNet pre-trained weights but,
it wasn't success that training using COCO pre-trained weights.

I had the same problem.

[yinggo]

Thank you. @weiyx16
It works for me too.
but, I wonder why using COCO pre-trained weights is not available.

because I'm trying to train custom detection dataset on Swin Transformer-Object Detection.
It was success that training custom dataset using ImageNet pre-trained weights but,
it wasn't success that training using COCO pre-trained weights.

Does anybody know that? I want to finetune the COCO pre-trained weights to my own dataset.

[bfialkoff]

What was the conclusion? Which is the correct base model from this page

[hans0809]

What was the conclusion? Which is the correct base model from this page

Hi, did you solve this ? I also encountered it today and did't solve yet.

[swimmant]

使用config文件中load_from关键字进行加载预训练模型，不要使用参数--cfg-options model.pretrained=<PRETRAIN_MODEL> [model.backbone.use_checkpoint=True] [other optional arguments] 这种方式加载模型

[zccsir]

适用于：调用训练好的模型继续优化
configs/base/default_runtime.py
load_from = r"E:\workspace\Python\Pytorch\Swin-Transformer-Object-Detection\mask_rcnn_swin_tiny_patch4_window7.pth"