Salmon

Salmon is a tool for efficiently generating ordinal embeddings. It relies on "active" machine learning algorithms to choose the most informative queries for humans to answer.

Documentation

This documentation is available at these locations:

• Primary source: https://docs.stsievert.com/salmon/
• Secondary source: as a raw PDF (and as a slower loading PDF).
• Secondary source: as zipped HTML directory, which requires unzipping the directory then opening up index.html.

Please file an issue if you can not access the documentation.

Running Salmon offline

Visit the documentation at https://docs.stsievert.com/salmon/offline.html. Briefly, this should work:

$ cd path/to/salmon
$ conda env create -f salmon.lock.yml
$ conda activate salmon
(salmon) $ pip install -e .

The documentation online mentions more about how to generate an embedding offline: https://docs.stsievert.com/salmon/offline.html#generate-embeddings

With this, it's also possible to create a script that uses and imports Salmon:

from salmon.triplets.samplers import TSTE
import numpy as np

n, d = 85, 2
sampler = TSTE(n=n, d=d)

em_init = np.array([[i, -i] for i in range(n)])
sampler.opt.initialize(embedding=em_init)

queries, scores, meta = sampler.get_queries(num=10_000)

This script allows the data scientist to score queries for an embedding they specify.