The question Why does VGG16 double number of features after each maxpooling layer? was put on hold after I had already answered it. As such, I think the OP cannot accept my answer (correct?).
In hindsight, I should have seen that coming, because I think that Machine Learning questions are more at risk of being off-topic than other questions. However, unlike the question I mentioned last time (Backpropagation in multi-layer perceptron (MLP) doesn't converge), which I voted to close, this time there's not a line of code in the OP's question: it's purely a question on a certain CNN architecture (VGG16).
If I edit the question removing the reference to Jeremy Howard's fast.ai course, and I add a reference to the paper where VGG16 had been presented, do you think the question could be opened again? Is it acceptable if I edit the question, instead than the OP? I'm not begging for the question to be reopened, just exploring a possibility. If you think it has to be closed, I won't do anything.