How to understand the organization of model weights of the PyTorch version?

[terU3760]

On the Readme of this project, the links to both TensorFlow and PyTorch version of pretrained model weights are provided. But After an inspection of both of them, it seems PyTorch version is much messy than TensorFlow version. How to understand the organization of model weights of PyTorch version?

[cocoon2wong]

Hi @terU3760!
Thanks for your questions!

Both TensorFlow or PyTorch versions of our model weights are organized in the same way. You can download them, unzip them, and put them to any place, then use the same command to verify these models.

One standard weights folder contains 4 parts:

• Core network weights (TensorFlow (*.data + *.index + checkpoint) or PyTorch (a single *.pt) weights files);
• Config file args.json;
• Manual checkpoint file best_ade_epoch.txt;
• Tensorboard log file.

To validate these models, whether TensorFlow or PyTorch, please use

python main.py --sc ${SOME_WEIGHTS_FOLDER}

Here, ${SOME_WEIGHTS_FOLDER} is the name of the weights folder to be validated.

Please provide more descriptions about your confusions so I can further address your concerns.

Thank you!