Visualisation tools for Tenbilac

[kuntzer]

I wrote a script that produces Tensorflow Playground-like plots for Tenbilac. I will add to tenbilac.plot.py this in a new branch (#16).

[mtewes]

Nice!

[kuntzer]

Done. This is now part of the autoplots with, if availble, the correct labelling of the inputs and outputs. Will merge soon.

[kuntzer]

Done!

[mtewes]

Just thinking (not urgent -- once all the rest is working nicely): could be interesting to

• revert y, so that first feature is first
• write the iteration number on the plot (title), why not add the training time as done in sumevo etc. ... to make an animation for talks of the training.

[mtewes]

Further points:

• I think it should take a Net (and not a Training) object. It's role is very close to net.report(). If training-related stuff is desired to be visible on the plot, this can be passed to the title or other kwargs...
• Use "+" and "x" symbols to mark mult and sum neurons ?

[mtewes]

• somehow make the biases more prominent (maybe as circles with varyign thickness around the nodes) ?
• add act function ?

A netviz animation (this one does not yet start at iteration 0, and simultaneously optimizes the Mult and the Sum layer: https://uni-bonn.sciebo.de/index.php/s/DqbiMgTByiEkFcy

[kuntzer]

A summary of the above points and a few others:

• [x] Debug the display for the name of the cost function
• [x] Use "+" and "x" symbols to mark mult and sum neurons ? (only if not only +)
• [x] somehow make the biases more prominent (maybe as circles with varyign thickness around the nodes)?
• [x] Add act fct name somewhere
• [x] Should be taking a net rather than a Training
• [x] revert y, so that first feature is first write the iteration number on the plot (title)
• [x] why not add the training time as done in sumevo etc. to make an animation for talks of the training.
• [x] Handle multnet better (maybe in different colors?)

[mtewes]

Nice -- I really like this plot, I guess we'll also use it in papers!

I would say that

• optimization algo
• cost function name and value, structure of training data (cases, reas, minibatches etc)
• iteration,
• training time ... could be passed as free text in an optional text display, and should maybe not be hard-coded in the plot itself (to keep the later simple and flexible).

With this option, I guess that multnets don't even need special handling. They are just normal nets which happen to have a mult layer specially initialized.

[kuntzer]

working on this in branch viznet

[kuntzer]

Should be taking a net rather than a Training

mmmh, that would be a big change: a net does not know the error function or the in-/outputs names. The plot would be less interesting. I could imagine a very adaptive plot that could take a train or a net, but the polymorphism in Python is a bit messy.

The plot looks like this now: if the net contains only summation layers: if the net contains both modes: The title is an argument that is by default "default". This will generate the string shown in the plots, but any other string can be passed to title.

Other suggestions?

[mtewes]

But this plot is really about a net, and not about a training! A yes, a net knows and nicely handles its inames and onames! Information about the cost function, iteration number, should be passed into a "title" or "textlines" argument. Note that the first part of the current title is in fact what is shown in the plot.

[kuntzer]

Okay, done. The changes were easy enough

[kuntzer]

Let's close this. (at least for now)