Why is model state dictionary different in pkl file vs pyth from pytorchvideo?

I am trying to implement the slowfast model on live feed. I was able to do so using the demo code and feel that it is not straighforward to use in a custom application. While searching, I found [https://pytorchvideo.org/docs/tutorial_torchhub_detection_inference](url) for `slowfast_r50_detection` and that seems a good source. I was playing with the model and checkpoints to use them offline and noted down below observation

I counted states from the output of `list(checkpoint['model_state'].keys())` where
`checkpoint = torch.load('file mentioned in first column of below table')`
Input checkpoint fileCommentOriginal PyTorchVideo weights (pyth file)Working | 662 statesOriginal pkl file (from the screenshot)It shows missing keys error because of nomenclature issue | 1313 statesOriginal pkl Converted (as per #552)Tensor shape mismatch issue | 1313 statesCustom training weightsIt will show missing keys error because of nomenclature issue | 662 statesCustom training weights convertedNeed to change detection head in model as per number of classes | 662 states

I am referring to the nomenclature convention issue mentioned in #552 Issue 1.

The pkl file is the pytorch checkpoint (not the caffe2) from below link in the model zoo.

The pyth checkpoint from pytorchvideo is from this link: [https://dl.fbaipublicfiles.com/pytorchvideo/model_zoo/ava/SLOWFAST_8x8_R50_DETECTION.pyth
](url)

Why is there a lot more weights in the pkl file compared to the pyth file. Even weirder, if I train the model on a custom dataset, it has the same number of states (662) as the original pyth. Could somebody help answer it? Thanks in advance.