Train and test code

[WenLinLliu]

I couldn't find the training and testing code in the repository. Moreover, the testing code doesn't include the model inference process; it merely reads the results and performs correction. Will the training code be open-sourced? And during testing, is it only necessary to perform correction and fusion of multiple models?

[kemaloksuz]

Hi, thanks for your interest in our work.

Our calibration (python tools/analysis_tools/model_calibration.py) and MoE scripts (tools/test_ensemble.py) are designed to be able to work independent of the training/testing framework. You can pick a different framework as long as you obtain the detections as a json file, and pass them to these scripts similar to the examples we provide. We chose this for the sake of being more flexible in using any detection framework.

Having said that, we haven't removed mmdetection training and testing features from the repo as mmdetection already provides many detectors. This documentation in the repo, which was already provided in mmdetection, can help you to train your model and test them. As an example from that documentation, you can train a model:

• on a single GPU by python tools/train.py PATH_TO_CONFIG_FILE
• on multiple GPUs, usingbash ./tools/dist_train.sh PATH_TO_CONFIG_FILE GPU_NUM
  after replacing PATH_TO_CONFIG_FILE and GPU_NUM of your choices.

To obtain json files after training a model, you can check this response, which we provided for a different question in a slightly different context.

Please feel free to let us know if you have further questions.