How to properly migrate a trained model from MMDetection v2 to MMDetection v3?

[Ddlozada21]

Hi! I'm trying to migrate a trained model from MMDetection 2.x to MMDetection 3.x and I'm facing some issues.

I have:

A config file originally written for MMDetection v2 (with model, data, pipeline, etc.)

A .pth checkpoint file trained using that config

My goal is to continue training (fine-tuning) the model in MMDetection v3 using this .pth

I’ve already tried adapting the config to the v3 format following the migration guide, but when I load the model using the Runner with load_from, I get errors such as:

TypeError: conv2d() received an invalid combination of arguments - got (list, Parameter, NoneType, ...)
Also, I receive warnings like:

unexpected key in source state_dict: fc.weight, fc.bias

It seems like the architecture has changed enough that loading the weights directly is no longer fully compatible.

What is the recommended way to migrate a model trained with MMDetection 2.x (checkpoint + config) to MMDetection 3.x so I can continue training or inference?

Should I:

• Only extract backbone weights from the .pth and re-use them with init_cfg?

• Manually adapt the state_dict?

• Is there a conversion script available?

Thanks in advance for the help!