I want to re-implement this paper [https://github.com/cvondrick/videogan] in PyTorch, but I don’t have the GPU to re-train 5,000 hours of video. Can I convert their pre-trained weights for PyTOrch somehow?
Yes, it’s possible to do it by hand.
You could serialize the models in a common format (for example h5), write the model definition in pytorch, load the weights and assign them to the modules.
This code may help: