Question about the weights "detectron2://DeepLab/R-52.pkl"

[mvdelt]

Hi,
Both panoptic_deeplab_R_52_os16_mg124_poly_90k_bs32_crop_512_1024_dsconv.yaml and
panoptic_deeplab_R_52_os16_mg124_poly_200k_bs64_crop_640_640_coco_dsconv.yaml
are using the weights "detectron2://DeepLab/R-52.pkl".

Where this weights came from?
Is this a pre-trained weights or just a randomly initialized weights?

Thanks.

[bowenc0221]

detectron2://DeepLab/R-52.pkl is ImageNet pre-trained weights for a modified ResNet-50 backbone.

[mvdelt]

Thank you for your answer.
So, https://dl.fbaipublicfiles.com/detectron2/PanopticDeepLab/Cityscapes-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_90k_bs32_crop_512_1024_dsconv/model_final_23d03a.pkl is a weights trained with cityscapes dataset based on detectron2://DeepLab/R-52.pkl?

I'm trying to train panoptic-deeplab with my custom dataset(x-ray images), so I need to choose weights for baseline.
I guess training from detectron2://DeepLab/R-52.pkl would be better.. right?

[bowenc0221]

Yes, model_final_23d03a.pkl is the cityscapes model.
And I would suggest you start with the detectron2://DeepLab/R-52.pkl weight for your custom dataset.