This PR changes the sensitivity analysis experiment so that (a) it differentiates the network output w.r.t. the controls, rather than the loss function and (b) this gradient then gets weighted by average values of the input parameters. This means that we get a better sense of "the sensitivity of the network w.r.t. its inputs", as opposed to just some gradient of the loss. However, it would be good to use representative values for each of the inputs, rather than just the means.
The PR also changes how the preprocessing function is applied. It is now a layer within the network definition, rather than being manually applied to the input data.
This PR changes the sensitivity analysis experiment so that (a) it differentiates the network output w.r.t. the controls, rather than the loss function and (b) this gradient then gets weighted by average values of the input parameters. This means that we get a better sense of "the sensitivity of the network w.r.t. its inputs", as opposed to just some gradient of the loss. However, it would be good to use representative values for each of the inputs, rather than just the means.
The PR also changes how the preprocessing function is applied. It is now a layer within the network definition, rather than being manually applied to the input data.