Skip to content
/ DiVE Public

An interactive 3D web viewer of up to million points on one screen that represent data. Provides interaction for viewing high-dimensional data that has been previously embedded in 3D or 2D. Based on graphosaurus.js and three.js. For a Linux release of a complete embedding+visualization pipeline please visit https://github.com/sonjageorgievska/Em…

License

Notifications You must be signed in to change notification settings

NLeSC/DiVE

Repository files navigation

Please cite the software if you are using it in your scientific publication:

DOI Codacy Badge

DiVE - Interactive Visualization of Embedded Data

DiVE is an interactive 3D web viewer of up to million points on one screen that represent data. It is meant to provide interaction for viewing high-dimensional data that has been previously embedded in 3D. For embedding (non-linear dimensionality reduction, or manifold learning) we recommend LargeVis (a new algorithm by Microsoft Research, ) or tSNE .

For an online demo click here. You can also use this site to upload your datasets complying to the data format described below.

New: a pipeline for Linux consisting of LargeVis and DiVE has been released here: https://github.com/sonjageorgievska/Embed-Dive .

Installation - for users

The simplest way is to download the code and open index.html with your browser. Try it by uploading datasets from the data folder. The application can work completely offline.

To use it with a local http server:

  1. Open your command line interpreter (CLI)
  2. Clone this repository
  3. Go to the main folder of DiVE in your CLI (where index.html is)
  4. Install node.js server together with the node package manager npm from (https://www.npmjs.com/get-npm).
  5. Type npm install connect serve-static
  6. Type node server.js
  7. Open your browser and type http://localhost:8082/index.html

Build - for developers

  1. Clone this repository
  2. Install node, npm, and grunt-cli
  3. Run npm install to install all the build requirements
  4. Run grunt to build. The resulting compiled JavaScript will be in dist/ and the docs will be in doc/

Data description and functionality

  • Every point has 3 coordinates and a unique ID. (For a best view, the absolute values of the coordinates should be smaller than 1. When using LargeVis with similarities (weights) as input, this can be achieved by re-scaling the similarities to be smaller than 1.)

  • A point also has Properties:

    • Properties is a list of strings which can be empty. Each string which is a number represents the value of a respective numerical property. Each string which is not a number represents the value of a respective categorical property. These values are used in the Coloring section of the UI of the web-page. When the user selects a property, if the property has categorical (non-numerical) values, each point is colored in a color representing the value of the categorical property. If the property is numerical, then after the user has selected a color, every point is colored with a shade of the selected color. The intensity of the color corresponds to the intensity of the selected property for the particular point.
  • A node can also have an image associated to it, see the Data format section for more info.

User interaction

Search

  • A user can search for all points that contain a certain substring in their ids, names or properties, by using the Search section. Then all points that are a match become red, and the rest become grey. One can search also for boolean expressions of regular expressions. An example of a boolean expression is xx AND yy OR NOT zz, where xx, yy, and zz are regular expressions and NOT binds more than AND, which binds more than OR. In this case all points that contain in their metadata the regular espressions xx and yy, or that do not contain zz, will be coloured in red.

  • Show only found nodes will show only the nodes that result from the search.

  • The Resume colors button at the bottom returns the colors of the points to the previous coloring scheme.

Visualization options

  • Centralize : will move data back to center of the screen, zoomed-in
  • See all data : will zoom-out such that all data is visible
  • Scase point size: very useful when the user has zoomed-in enough. When this option is not selected, the points do not get bigger as the camera moves closer to them, so that they can be separated and inspected individually.
  • Show point info in popup : when selected, the information about a point when hovering over it will be displayed in a pop-up message rather than at the top left corner of the screen

Coloring by value of property

As explained in section Data description and functionality .

Data format

  • The data is in a JSON (JavaScript Object Notation) format. (See folder data for examples.) To obtain data.js, first a data structure

      Dictionary<string, Point>
    

is created in any programming language, where the keys are the id’s of the points and Point is an object of the class

	public class Point
	    {
	        public List<double> Coordinates;
	        public List<double> Properties;
	    }

Coordinates and Properties are as discussed in the previous section.

Next, the dictionary is serialized using JavaScriptSerializer and written in data.json (name is flexible). Here is an example of an entry of the serialized dictionary in a data.json file:

		"3951": {
    "Coordinates": [0.99860800383893167, 0.61276015046241838, 0.450976426942296],
    "Properties": ["0", "1", "5", "12688892", "0.998", "5", "True", "0", "False", "5", "1", "True", "1", "518", "0", "-1", "Rhodotorula", "", "Sporidiobolales", "Microbotryomycetes"]
}

Optionally, if data has properties, the dictionary should also contain an entry

		"NamesOfProperties":["name1", "name2",  , "name_n"]

Optionally, if images are associated to the nodes, the node image can be displayed in a pop-up when hovering over the node. If the datafile starts with namedataset_ then the folder with images should be images_namedataset in folder data. (see examples in folder data, sorry for their size). The name of an image should be nodeId.jpg.

If your images have a .png extension then the fingerprints_namedataset folder is an option, although it is currently made for the Sherlock purposes.

From output of LargeVis to input of DiVE

The output of LargeVis is a text file - every line has the id of the point, and 3 coordinates (real numbers). Only the first line is an exception: it contains the number of points and the dimension. Here is an example:

	4271 3
	0 -33.729916 17.692684 17.466749
	1 -32.923210 17.249269 18.111458

It can be processed into an input of the viewer by using the python script "MakeVizDataWithProperMetaData.py" in the folder "scripts_prepareData". It is called with

	python MakeVizDataWithProperMetaData.py -coord coordinatesFile -metadata metaDataFile -dir baseDir -np -namesOfPropertiesFile 
  • coordinatesFile: the output file of LargeVis

  • metaData: file containing meta information about data. Format: [id] [metadata]. Format of metadata: "first_line" "second_line" "third_line" (number of lines is not limited). Example line of metadata: 35 "A dog" "Age:2" "Color brown".

  • baseDir: base directory to store output file

  • namesOfPropertiesFile: A json file containing list of properties names. Ex: ["Height", "Weight", "Place of birth"]. If file is omitted, its name should be "No"

Licence

The software is released under the GPL2 licence. Contact the author if you would like a version with an Apache licence

About

An interactive 3D web viewer of up to million points on one screen that represent data. Provides interaction for viewing high-dimensional data that has been previously embedded in 3D or 2D. Based on graphosaurus.js and three.js. For a Linux release of a complete embedding+visualization pipeline please visit https://github.com/sonjageorgievska/Em…

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published