Unexpected key(s) in state_dict: "heads.sample_x_indexs", "heads.prior_feat_ys", "heads.prior_ys", "heads.criterion.weight".

[sahamitul]

Hi,

I run:
python torch2onnx.py <config_file> --load_from <pth_file>
and get crash:-

pretrained model: https://download.pytorch.org/models/resnet101-5d3b4d8f.pth
Traceback (most recent call last):
File "torch2onnx.py", line 49, in
main()
File "torch2onnx.py", line 28, in main
net.load_state_dict(new_state_dict)
File "/home/msaha/anaconda3/envs/clrnet/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1223, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for Detector:
Unexpected key(s) in state_dict: "heads.sample_x_indexs", "heads.prior_feat_ys", "heads.prior_ys", "heads.criterion.weight".

Any ideas? @xuanandsix
Thanks!

[xuanandsix]

You did not successfully replace the clr_head.py file.
cp clr_head.py in this code to CLRNet/clrnet/models/heads/ in official code.

[sahamitul]

Thanks @xuanandsix.

It seems to be the clr_head.py from this repo is being indeed loaded.
I am trying to debug now - then why I still get this crash, as you say I should not.
Meanwhile if you have more thoughts, let us know here.

[xuanandsix]

1、By 'pip list' check if the path to the 'clrnet' package is correct.
2、Or choose to replace the clr_head.py file first before running package install (i.e. 'python setup.py build develop').
If you need training source code, it is recommended to maintain training and deployment with two environments.

A log has been upload here https://github.com/xuanandsix/CLRNet-onnxruntime-and-tensorrt-demo/blob/main/my_log/test_onnx.log
My deployment process from scratch. You can view each step by search: ***@ai02

[sahamitul]

Hi @xuanandsix, I notice:

In your example "python torch2onnx.py configs/clrnet/clr_resnet18_tusimple.py --load_from tusimple_r18.pth",
when I print the keys in "state_dict" from "https://github.com/xuanandsix/CLRNet-onnxruntime-and-tensorrt-demo/blob/main/torch2onnx.py#L22",
I do not see these:-
module.heads.sample_x_indexs
module.heads.prior_feat_ys
module.heads.prior_ys
module.heads.criterion.weight

But I have these when I load my trained model.

So maybe this is why you are not getting my crash that I list in this issue. What do you think?

[Yutong-gannis]

@sahamitul Have you solved this problem?

[mengxia1994]

same issues，any advices?

[Yutong-gannis]

@mengxia1994 maybe you can try other backbone. I met this problem when use resnet-18, but it worked when I use dla 34.

[LukeAI]

I found that exporting seems to work ok if I use the pretraining weights provided by https://github.com/Turoad/clrnet but if I train my own network, I get this error, with both dla34-culane and resnet18-tusimple. Not sure what's different

[LukeAI]

OK I got this working literally just by deleting the offending dict keys:

      del new_state_dict["heads.sample_x_indexs"]                                  
      del new_state_dict["heads.prior_feat_ys"]                                    
      del new_state_dict["heads.prior_ys"]                                         
      del new_state_dict["heads.criterion.weight

inserted just before

CLRNet-onnxruntime-and-tensorrt-demo/torch2onnx.py

Line 28 in 07a47e6

net.load_state_dict(new_state_dict)

Edit. probably better to do as @AshwinAKannan suggests with strict=False

    net.load_state_dict(new_state_dict, strict=False)

[ashwin-999]

I passed 'strict=False'

model = load_state_dict(path, strict=False)