Packing the Digit Generator

You train a variational autoencoder on $28 \times 28$ greyscale digit images, flattened to $784$ numbers each.

Every layer is fully connected. A fully connected layer from $n$ inputs to $m$ outputs has $n \times m$ weights plus $m$ biases, and activation functions have no parameters.

part	layer	shape
encoder	hidden layer	$784 \to 256$
encoder	output layer	$256 \to$ a centre and a spread for each of the $16$ code dimensions
decoder	hidden layer	code $\to 256$
decoder	output layer	$256 \to 784$

Training goes well. You now want a standalone program that produces brand-new digits on demand. You will package only the trained parts that this program actually needs in order to run.

How many trained parameters (weights plus biases) does the packaged generator contain?

Give the exact whole number.

Packing the Digit Generator

Hints