Missing key(s) in state_dict: "model.33.anchor_grid"

[Argentum471]

Hello @zhiqwang
I got this error when trying to load my custom yolov5n6 model. Is there a way to load yolov5-v6.0 models?

model = YOLOv5.load_from_yolov5(model_path, score_thresh=0.25, version="r6.0")

[zhiqwang]

Hi @123556666 , Sorry for missing this ticket. Could you check the version of yolort you're using?

In version 0.5.2, we've supported loading custom yolov5n6 model and auto check whether it is P6 or not.

https://github.com/zhiqwang/yolov5-rt-stack/blob/4cba0437389e2cccbcc299ab196c922945c93d45/yolort/utils/update_module_state.py#L70-L72

If it still reports an unsupported error, you can send me your model file if it's possible and I'll reproduce this bug and figure out how to fix it.

[Argentum471]

Sorry for long reply.
I am using version 0.5.2. I got model.33.anchor_grid error with loading my custom yolov5n6 model a bit earlier at
checkpoint_yolov5 = load_yolov5_model(checkpoint_path).
Pretrained yolov5n6 model loading, working and exporting fine. The error happens because of an additional anchor_grid attribute in Detect layer after training.

There was also a similar problem with custom yolov5 ckpt. I am using ultralytics/yolov5 v6.0 for model training.

[zhiqwang]

Seems to be caused by incompatibility with the latest version of YOLOv5 from the error log information. I'll try to fix this bug in the next two days. And I received the model checkpoint you sent, thank you again.

EDITED: I create an issue #236 for this problem, let's move the discussion there.

[zhiqwang]

Hi @123556666

We update the upstream YOLOv5 in #235 , please pull the main branch of yolort and check whether this modification will fix your problem.

[Okoyl]

I encounter the same issue.
issing key(s) in state_dict: "model.33.anchor_grid"
when loading a yolov5n6 trained model with the latest ultralytics/yolov5 commit.

[zhiqwang]

Hi @Okoyl , could you try to reinstall yolort from source like the following

pip uninstall yolort
pip install 'git+https://github.com/zhiqwang/yolov5-rt-stack.git'

we plan to distribute 0.6.0 at the end of the month.