Detection models conversion from pytorch

[tylertroy]

You state in the YOLOR page (https://github.com/leondgarse/keras_cv_attention_models/tree/main/keras_cv_attention_models/yolor) that "Model weights converted from official publication." Similar statements are made for YOLOX and EfficientDet.

I couldn't find in the repo the means for conversion. Am I missing it here and if it's not available, could you share how these weights were converted.

I am also assuming that the provided training scripts do not provide for training the detector models. If I am mistaken, any guidance on this would be greatly appreciated. I am mistaken, it's clear in the README! sorry for the confusion. Nevertheless the above question of model conversion still stands.

Thank you for making available this amazing repo.

[leondgarse]

• The conversion is actually using my keras_reload_from_torch_model. The specific conversion for yolox is in my own notebook Atom_notebook yolox/reload-torch-weights. Yolor is almost similar. It's still for my own usage and not documented.
• Some others also there: Efficientdet conversion and more others pytorch_to_keras.
• Ya, the training script is coco_train_script.py. Just I still don't know the best parameters like optimizer / augment for training, so you need to find out by yourself. :)