ncvis 1.0

Noise contrastive data visualization

ncvis

NCVis is an efficient solution for data visualization. It uses HNSW for fast nearest neighbors graph construction and a parallel approach for building the graph embedding.

Installation

Conda [recommended]

You do not need to setup the environment if using conda, all dependencies are installed automatically.

$ conda install -c alartum ncvis 

Pip

Important: be sure to have a compiler with OpenMP support. GCC has it by default, wich is not the case with clang. You may need to install llvm-openmp library beforehand.

Install numpy and cython packages (compile-time dependencies):

$ pip install numpy cython

Install ncvis package:

$ pip install ncvis

Using

import ncvis

vis = ncvis.NCVis()
Y = vis.fit(X)

A more detailed example can be found here.

Experiments

Datasets can be dowloaded by using the download.sh script:

$ bash examples/data/download.sh <dataset>

Replace <dataset> with corresponding entry from the table. You can also download all of them at once:

$ bash examples/data/download.sh

The datasets can be then accessed by using interfaces from the data Python module. Make sure that these packages are installed:

$ pip install -r examples/requirements-pip.txt

or

$ conda install --file examples/requirements-conda.txt

Dataset	<dataset>	Dataset Class
MNIST	mnist	MNIST
Fashion MNIST	fmnist	FMNIST
Iris	iris	Iris
Handwritten Digits	pendigits	PenDigits
COIL-20	coil20	COIL20
COIL-100	coil100	COIL100
Mouse scRNA-seq	scrna	ScRNA
Statlog (Shuttle)	shuttle	Shuttle
Flow Cytometry	flow	not yet
GoogleNews	news	not yet

Each dataset can be used in the following way:

Sample Code	Action
d = data.MNIST()	Load the dataset.
ds.X	Get the samples as numpy array of shape (n_samples, n_dimensions). If samples have more than one dimension they are all flattened.
ds.y	Get the labels of the samples.
len(ds)	Get total number of samples.
ds[0]	Get 0-th pair (sample, label) from the dataset.
ds.shape	Get the original shape of the samples. For example, it equals to (28, 28) for MNIST.