Same MMPose/MMDeploy setup in two locations, one works, one fails.

[korneliaWatson]

Hi,

I'm training MMPose on a custom keypoint detection task. I have repository setup in two locations, let's call them location A (with GPU) and location B (without GPU).

Codebase and mm* dependencies are exactly the same in both locations.

mmcv==2.0.0
mmdeploy==1.2.0   
mmdet==3.1.0
mmengine==0.8.4
mmpose==1.1.0      

However,

• in location A I can successfully train a model.
• in location B I get KeyError: 'CocoPersonFront is not in the mmpose::dataset registry. Please check whether the value of CocoPersonFront is correct or it was registered as expected. More details can be found at https://mmengine.readthedocs.io/en/latest/advanced_tutorials/config.html#import-the-custom-module'.

I compared all modified files and they are exactly the same (I use a script to perform changes so they should be the same). I can't figure out why location B errors.

Another problem I have been experiencing...

As I managed to train a model in location A, I wanted to prepare a separate inference script with the use of MMDeploy, as per example here.

In location A, my inference script works.
In location B, I get the following error:

09/07 15:18:58 - mmengine - INFO - Successfully loaded onnxruntime custom ops from /home/project/venv-mmdeploy-infer/lib/python3.10/site-packages/mmdeploy/lib/libmmdeploy_onnxruntime_ops.so
Traceback (most recent call last):
  File "/home/project/tasks/infer_mmdeploy.py", line 91, in <module>
    main(config, args.img)
  File "/home/project/tasks/infer_mmdeploy.py", line 38, in main
    model_inputs, _ = task_processor.create_input(image_filepath, input_shape)
  File "/home/project/venv-mmdeploy-infer/lib/python3.10/site-packages/mmdeploy/codebase/mmpose/deploy/pose_detection.py", line 261, in create_input
    data.update(meta_data)
TypeError: 'NoneType' object is not iterable

I looked at meta_data variable, and in location A it is present, in location B it is indeed None.

I begin to think that this variable is linked to the training process as the only difference between location A and location B is that I managed to train a model in location A but not in B. Is that the case? If so, it would suggest that it is not possible to decouple mmpose such that inference can be done without processes done for and during model training?