Skip to content

Latest commit

 

History

History
107 lines (78 loc) · 4.8 KB

README.md

File metadata and controls

107 lines (78 loc) · 4.8 KB
Raw

nn-viz

A visualizer for JavaScript neural networks, deep learning and machine learning models

Demo

Visualization Demo

Built on

Contributing

This project is in its very earliest stages. Your contributions are very welcome especially in guiding the direction of the project!

To build a community welcome to all, NN-Viz follows the Contributor Covenant Code of Conduct.

And finally, a big thank you to all of you for supporting!

Concept

nn-viz is a Vue.js project that displays the structure of a network. The actual network is being trained in a seperate thread in a web worker.

Examples

For examples on how to create such a worker and how to communicate with the visualization thread, see src/workers.

Configuration

For configuration options to pass to the worker, see the config object in App.vue (This is subject to change. We will most likely move this object to a stand-alone .json or even .js file)

To create a new worker, just add an entry here. A new button will appear in the control section of the visualizer. If you click it, the new worker will be set as active and commands will be sent to this particular web worker.

The web worker itself needs to have the same name and be created inside the src/workers directory.

API

The AND worker already exists as an example for this project but we will use it as an example on how to add a new worker and as part of the API examples regardless.

App.vue, config object:

AND: { archtitecture: [2, 5, 5, 1] }
// This will create a new button "AND"

Incoming messages to web worker

initialize

This event is fired when the worker has been selected, it might look like this:

{ event: 'initialize', config: { archtitecture: [2, 5, 5, 1] } }

goes

This event is fired by the 'GOES' button in the control section. It commands the worker to start training the network. Arguments are the config options you set via the control section.

It might look like this:

{ event: 'goes', goestimes: 100000, cutoff: 0.001, updateInterval: 1000, learningRate: 0.01 }

pause

The pause event commands the worker to pause its training. This pause is started once the current training interval has ended (see updateInterval goes config). No config other than the name of the event is passed.

resume

The resume event commands the worker to continue the training. No config other than the name of the event is passed.

load

This commands the worker to load a certain network. The network is a JSON formatted export. To load it, use Network.fromJSON. The network sent to be loaded is the one last saved from the control section.

Outgoing messages

The web worker updates the visualization via events. The update event takes the following arguments:

  • network: the JSON formatted network export
  • epoch: the currently completed training epoch
  • results: the actual activation values of the network compared to the input and ideal output
  • score: the score/error of the network. For simplicity, the property is just called score

network

To export the network to JSON, simply call network.toJSON. To show activation values as part of the visualization, you need to manually add those.

epoch

The epoch is the counter of how many times an example has been run past the network. Simply pass an integer.

results

Results can be easily obtained via utils.evaluate. This will calculate the mean squared error and the results and deliver them as part of the evaluation with the keyword results

score

Score can be easily obtained via utils.evaluate. This will calculate the mean squared error and the results and deliver the score as part of the evaluation with the keyword score

Example:

const evaluation = utils.evaluate(this.network, examples);
self.postMessage({
  event: 'update',
  epoch: i,
  network: network.toJSON(),
  results: evaluation.results,
  score: evaluation.score
})

Coming up

We encourage anyone who wants to debug any JS machine library to give it a try and give feedback. We have plans to continue developing it and extending its features, some of our planned updates:

  • A way to single step through the network (activation by activation)
  • Providing stand alone visualization components for use in other projects
  • A clearer definition of the format the visualizers take to make it easier to plug in networks of different libraries
  • More options for visualizers
  • Visualization of lineage in the evolutionary modes of carrot