This repo implements a FixedDropout torch layer and a WeightDropLinear linear layer. The latter drops weights an biases of a linear layer during the forward pass, using the given probability, whereas the former implements a dropout layer with variable probability specified during the forward pass.
Both bandwidth manipulation methods are implemented as part of a MLP with a single hidden layer for sinus function prediction. FixedDropout does not work as intended because dropping activation layers results in a non-continuous function prediction.