http://d2l.ai/chapter_convolutional-modern/nin.html
Why are the results so bad for MXnet and tensorflow in 7.3.3? This doesn’t happen for Pytorch.