Note

This package has been deprecated.

TripletEmbeddings.jl

This is a Julia 1.0 implementation of Triplet Embeddings. Currently, the following algorithms are implemented:

• Stochastic Triplet Embeddings:
  • STE (normal kernel in probabilities)
  • tSTE (t-Student kernel in probabilities)
• Generalized NMDS with the following loss functions:
  • Hinge loss
  • Smooth hinge loss
  • Logarithmic
  • Exponential

These implementations are based on based on Laurens Van der Maaten's implementation in MATLAB. However, this implementation uses multithreading for the calculation of the gradients, making it the fastest implementation.

Usage

Loading the package

Install using:

julia> ]
(v1.0) pkg> add https://github.com/kmundnic/TripletEmbeddings.jl

Using the package

To use multiple threads, on the command line, write (before opening Julia):

export JULIA_NUM_THREADS=n

where n is the number of threads (defaults to 1).

The following code is a usage example:

using Plots
using Random
using TripletEmbeddings # see instructions above

Random.seed!(4)

# Load data dummy data
data = TripletEmbeddings.load_data()

# Generate triplets
triplets = TripletEmbeddings.label(data)

# Create an embedding with the triplets and compute the embedding
dimensions = 1
params = Dict{Symbol, Real}()
params[:σ] = 1/sqrt(2) # Normal distribution variance

te = TripletEmbeddings.STE(triplets, dimensions, params)
@time violations = TripletEmbeddings.fit!(te; max_iter=50, verbose=true)

TripletEmbeddings.scale!(data, te)
plot(data, label="Data")
plot!(TripletEmbeddings.X(te), label="Embedding") # Can also use te.X.X to access the embedding