Solve the identify medium problem

[Joshuaalbert]

This problem is a necessary step in the larger project. This is required to enable the bridge in the H-network.
Given an image and a graph representation of a medium, determine if the image is from the graph.
Treat this as the task of learning a distance measure between graphs.

• Ensure the encoding graph network of both the image and 3d medium is properly normalised. You do not want activation blowing up, nor being zero. This can be accomplished with good choice of units and residual connections.
• Make sure your compare function is symmetrised so that compare(x,y)=compare(y,x).
• Train and save 3 models.
• For a bunch of test batches visualise a) 3d medium, b) image, c) graph representation of the 3d medium, d) graph representation of the image

debug process

1. Start at the data, the thing you know some truths about and work forwards towards the error.
2. When there are nan's it likely means there is log(negative value) somewhere, or very large or very small values flowing through your network, or sometimes a zero*inf.
3. When things are identically zero, look at the magnitude of inputs and outputs of layers (you can print them). All inputs should be O(1). Or rather the scale should not be very large or small.

collaborate

1. Find a problem, share it, find edge cases independently and asynchronously that give the problem, inform each other of edge cases you find, solve together, repeat.