Dear all,
This communication is especially important for those of you willing to compete the homework.
Just a small correction about yesterday's lab:
The multilayer perceptron we developed yesterday should have a log softmax activation function before the output instead of the usual softmax.
This because PyTorch NLLLoss expects the logarithms of softmax instead of the standard simplex.
The network will train anyway with softmax, but it probably won't converge optimally and get stuck on poor solutions.
I'll push the correction on GitHub ASAP.
Sorry for the inconvenience.