Perturbing the simple autoencoder

So we have a toy autoencoder that works, if not very usefully. Toys are for playing with; what happens if we tweak it a bit - change the parameters?

The first change is a refactoring - move the data provision and normalisation code into a module. That will clean up the code somewhat, and should make no substantial difference to the results.

• Refactoring

Then we can start experimenting in earnest:

• Less training on the data
• Less data to train on

• Fewer neurons in the hidden layer (8 rather than 32)
• Fewer neurons in the hidden layer (2 rather than 32)
• More neurons in the hidden layer (64 rather than 32)

• With L1 regularization

• With linear activations (rather than tanh)
• With elu activations
• With relu activations
• With sigmoid activations
• With softmax activations
• With softplus activations
• With leaky relu activations

• With video diagnostics

• Trained on 2m air temperature (rather than MSLP)
• Trained on precipitation