Training an unsupervised convolutional neural network to learn a general representation of a Lua module

Question

I am trying to train a CNN in keras to learn a general representation of a Lua module, e.g. requires at the beginning, local variables, local functions, interface (returns) and in between some runnable code (labeled "other"). So for each module (source code) I generate an AST which I then encode in a json file. The file contains the order of the node in the AST, the text it represents and the type of node it is (require, variable, function, interface, other). It can contain other metrics but so far I have settled on these three, where only the order and type of node will be converted into a vector to serve as input to the CNN. Now I don't have any labels for these modules (I want to treat one module as one input to the CNN), so I have concluded that I need to use unsupervised learning. In keras this should translate to using autoencoders, where I use the encoder part to learn weights for the representation and then connect a fully-connected layer and generate an output. Now before I specify the output, I want to specify the input more closely. In my mind It should be a 3D vector let's say (x,y,z). x represents the number of nodes of an AST that are taken into consideration, y represents the local neighborhood of said node (for now I have settled on 5 nodes) and z should represent the node itself, the order and type of node. So with that, I would want the output of the network to be in the (almost) same dimension. I want x outputs for every node that was taken as input and a number (ideally between 0-1) to specify how "correct" the node under consideration is in response to the learned representation. My question as a beginner to neural networks is, how feasible is this and are there any points which are simply impossible to do or are wrongly interpreted on my part?