HeatMap visualization of SLIM learning

[remiadon]

Describe the workflow you want to enable

Would be nice if we can visualize SLIM learning data representation stepwise

Some idea to do this:

• create an async watcher, making deep copies of an object attributes at predefined steps
• create a notebook showing data import + SLIM instantiation + watcher instantiation
• make watcher send data to a matplotlib FuncAnimation or streamz DataFrame
• plot the resulting animation

This way we could get a sexy representation of what MDL pattern mining algorithms are doing

[remiadon]

visualization is an entire field on its own
I'll write a watcher just to keep track of compression ratio and plot them

[remiadon]

OK new thoughts on this

What we actually want is a logger
scikit-learn people already had a discussion about this

The latest trend is here
They are defining a callback API, just like in Keras

Making skmine compatible with this new sklearn-callbacks library should allow plenty of new usages.
But this is not straightforward, we have to keep lot of sklearn-compt attributes inside our objects.

Conclusion;

• writing our own logging handlers is a simple, striaghtfoward way to track metrics
• providing compatibility with the newest trends in sklearn is even better, but requires a lot of work