Originally drafted for the computer-vision course of Prof. Roland Kwitt at the University of Salzburg.
The implementations are rather shallow and define a fixed architecture (rather than specifying the architecture from parameters). That makes them less effective and flexible than the implementations one would use in practice. However, such simple DIY implementations are a great way to really understand the architectures. The fact that you can turn a regular ConvNet into a ResNet or DenseNet with just a few modifications in the forward-method makes you appreciate these simple but powerful improvements of the standard ConvNet architecture.