Using Torchdistill with MMDetection

[ai21mtech13002]

Hi Yoshitomo,

Can I use models trained in MMDetection with TorchDistill to run experiments?

[yoshitomo-matsubara]

Hi @ai21mtech13002

Yes, you can.
If the trained model weights are a form of state_dict, you can specify the file path as ckpt

P.S. Could you also close the previous discussion?

[yoshitomo-matsubara]

Hi @ai21mtech13002

Technically, you can use datasets in non-COCO formats with torchdistill, but then you will need to modify the exaple script, examples/object_detection.py in this case (especially evaluate function should be replaced with a DOTA-format-compatible evaluate function)

[ai21mtech13002]

What other modifications would be required in the config file? I am enclosing a sample annotation file for reference
P0000.txt

P.S. The annotations are for Rotated Bounding Boxes. I also considered converting them to COCO annotations , but COCO does not support Rotated Bounding Box annotations.

[ai21mtech13002]

Also, mmdetection carries out the following preprocessing prior to passing the batch through the model.

These are defined as classes in the mmdetection repository.

How do I account for the same in Torchdistill?

[yoshitomo-matsubara]

Hi @ai21mtech13002

You will register and fine datasets and model configs like you did here, potentially + your custom forward_proc to re-select what you want to feed to the model.

P.S. The annotations are for Rotated Bounding Boxes. I also considered converting them to COCO annotations , but COCO does not support Rotated Bounding Box annotations.

I am not very familiar with Rotated Bounding Box annotations, MMDetection or their pipeline designs either. If the training pipeline just performs dataset pre-processing, you can modify the dataset model class (and introduce your custom collate and/or batcher if necessary) to work with torchdistill.

[ai21mtech13002]

Thanks Yoshitomo. I will look into this.