Many pre-trained models exist within the torchvision library; can these models be used or is it considered as an external data source?
If it’s publicly available, I believe yes it is.
Yes, pretrained models which are packaged in libraries like pytorch and tensorflow are allowed.
What if the pretrained models are not a part of either library, but trained on a standard publically available dataset, eg. ImageNet. And similarly many of the transformer models, BERT, GPT2, etc.
And who decides what a popular dataset is.
Yes, the broader sentiment is that as long as the pretrained models are generally available for all participants you can very well use them.
It would also be important to announce all such pretrained models that you are using, on the forums.