search

cog-imperial / OMLT

Represent trained machine learning models as Pyomo optimization formulations
Other
257 stars 56 forks source link

Add partition formulation to notebook and update mnist notebooks #72

Closed jalving closed 2 years ago

jalving commented 2 years ago

This PR adds the ReluPartitionFormulation to the neural_network_formulations.ipynb notebook. I updated text cells to reflect the addition. The formulation is trivial for a single input, but it does show how one might use it to partition inputs.

I also updated the mnist adversary examples. I noticed they still referenced the old NeuralNetworkFormulation. I updated them to say FullSpaceNNFormulation and said they could technically use the ReluBigMFormulation. Please have a look over the changes.

codecov[bot] commented 2 years ago

Codecov Report

Merging #72 (e9e8530) into main (fc42b5b) will increase coverage by 0.06%. The diff coverage is n/a.

@@            Coverage Diff             @@
##             main      #72      +/-   ##
==========================================
+ Coverage   94.12%   94.18%   +0.06%     
==========================================
  Files          24       24              
  Lines        1242     1239       -3     
  Branches      192      192              
==========================================
- Hits         1169     1167       -2     
+ Misses         43       42       -1     
  Partials       30       30
Impacted Files Coverage Δ
src/omlt/block.py 100.00% <ø> (ø)
src/omlt/neuralnet/activations/linear.py 100.00% <ø> (ø)
src/omlt/neuralnet/activations/smooth.py 100.00% <ø> (ø)
src/omlt/neuralnet/layers/full_space.py 100.00% <ø> (ø)
src/omlt/neuralnet/layers/reduced_space.py 100.00% <ø> (ø)
src/omlt/io/onnx.py 85.71% <0.00%> (ø)
src/omlt/scaling.py 100.00% <0.00%> (ø)
src/omlt/formulation.py 92.53% <0.00%> (ø)
src/omlt/gbt/__init__.py 100.00% <0.00%> (ø)
... and 9 more

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 7787c00...e9e8530. Read the comment docs.

jalving commented 2 years ago

Good catch! There was a typo such that it built the Big-M formulation again. You were correct; they do build different problems with different numbers of variables and constraints.

The latest commit should show this.