EQL-Div - Python 3 port of original theano code

Paper: Learning equations for extrapolation and control by S. S. Sahoo, C. H. Lampert, and G. Martius in ICML 2018

This is the source code using Python Theano. We are also working on a tensorflow implementation.

You need

• Python 3.6
• Theano
• graphviz

Files to use:

• mlfg_final.py
• use createjobs.py to perform hyper-param scan (here for a formula called F0) for more complicated problems, use a larger number epochs, e.g. 10000 or 20000 it will create lots of files. Do this on the cluster or somehow run all the ..sh files on your machine
  • use finish...sh to collect results
  • you also see a typical command line call. The -extrapol flags are used to specify additional data files for testing. The first one is used here for the interpolation test set.
• use Evaluation.ipynb to perform model selection and look at the result

Dataset generation

• see ICML-Datasets.ipynb

For visualisation file extensions to look out for:

• .trainloss for MSE + L1 loss
• .L1 for L1 loss i.e sum of magnitude of all the weights
• .MSE for mean squared error loss
• .extrapoltrainloss mean square error when the magnitude of predicted value goes beyond 10