This repo contains Tensorflow v1 and Pytorch implementations of Deep Norms, Wide Norms, and Neural Metrics, from our ICLR 2020 paper, which may be found at:

• Openreview
• Arxiv

@inproceedings{pitis2020inductive,
  title={An Inductive Bias for Distances: Neural Nets that Respect the Triangle Inequality},
  author={Pitis, Silviu and Chan, Harris and Jamali, Kiarash and Ba, Jimmy},
  booktitle={Proceedings of the Eighth International Conference on Learning Representations},
  year={2020}
}

To create a conda/virtualenv and install dependencies for our code, assuming that ${DEEPNORM_ROOT} the directory that you cloned our repository into:

conda create -n deepnorm
conda activate deepnorm
cd ${DEEPNORM_ROOT}
pip install -r requirements.txt

The main implementation files are:

• metrics_tf1.py: defines the model components in Tensorflow v1 code
• metrics_pytorch.py: reimplementations of the model components in Pytorch (untested)
• custom_metric_loss_ops.py: drop in replacement for Tensorflow v1's contrib.metric_loss_ops that computes pairwise distances using our models

To run Pytorch examples, use pytorch_examples.ipynb.

Some additional files to reproduce some of our experiments (Please note: this is TFv1 research code, there are no guarantees that this will "just work".):

• Make3dAndTaxi.ipynb: creates 3d and Taxi graphs in form of .edgelist
• PrepXYDPickles.ipynb: creates the dataset (but not embeddings)
• data.py: makes landmark embeddings given a .edgelist, and loads datasets used in experiments
• experiment.py: runs experiment given an embedding pickle and a dataset pickle

1. Use Make3dAndTaxi.ipynb to make .edgelist
2. Use PrepXYDPickles.ipynb given .edgelist
3. Use make_lm_embeddings from data.py given .edgelist
4. Use experiment.py to run experiment using pickes created in previous two steps

1. Run python experiment_2d_norm.py. This trains several architectures on different 2D norms and saves the results in ./2D_metrics folder.
2. Launch the notebook 2D Norm Result Plotting.ipynb and run the cells to update the plots for the results.