I think this idea is exactly right. Many ML papers that include pseudocode basically just write out a minimal PyTorch training loop - down to the loss.backward(

I think this idea is exactly right. Many ML papers that include pseudocode basically just write out a minimal PyTorch training loop - down to the loss.backward(). It has, along with Python, become something of a linga franca for research in the field.