Attention Visualizer

Use it! haldenl.github.io/attention-visualizer

A tool for visualizing attention in sequence-to-sequence summarization models. Built with D3 + React.

Publication: Visualizing Attention in Sequence-to-Sequence Summarization Models, (poster) VAST 2018

The example data presented was generated using a pre-trained model by See et al (2017): Get To The Point: Summarization with Pointer-Generator Networks.

Loading your own data

You can view the behavior of your own models by opening the Data Sources panel in the lower right, and adding a url to your own data. Data should follow the following format:

{
  "attentionRecords": [{
    "inputIndex": 0,
    "outputIndex": 0,
    "weight": 1
  }, 
  {
    "inputIndex": 1,
    "outputIndex": 0,
    "weight": 0
  }],

  "inputTokens": [{
    "index": 0,
    "token": "word"
  }, 
  {
    "index": 1,
    "token": "example"
  }],

  "outputTokens": [{
    "index": 0,
    "token": "word"
  }, 
  {
    "index": 1,
    "token": "example"
  }]
}

Running locally

Use the command yarn to install dependencies, then yarn start to run a local instance.