bvanjoi / ai_mouse_movements

Send POST requests for mouse movements generated by a GAN!

API for Generating Mouse Movements with Neural Networks

Send POST requests to automatically move your mouse with a neural network!

(Note: The api now only generates coordinates and no longer moves the mouse for you.)

Table of Contents

• API for Generating Mouse Movements with Neural Networks
  • Table of Contents
  • pymousegan
    • Getting Started
    • Dependencies
    • Training Pipeline
      • Preprocessing
      • GAN
  • [JS API]
    • Getting Started [Client]
    • Model Format Conversion
      • From tf.keras to .json
    • Dependencies
    • How does it work?
    • How fast is the api?

---

pymousegan

This is the Python library containing the code for creating neural networks.

The training is done in the Colaboratory notebook. pymousegan contains the models and training pipeline for the GAN.

Example notebooks are located at python/notebooks

Getting Started

git clone https://github.com/jchen42703/ai_mouse_movements.git
cd python
pip install .

Dependencies

• numpy
• tensorflow
• pandas
• matplotlib

Training Pipeline

Preprocessing

1. Translated so that the starting coordinate is (0, 0).
2. Scaled so that the destination coordinates is (1, 1).
3. Reflection across all axes done during training.

GAN

The model used in the current version is a BidirectionalLSTMDecoderGenerator from an AdditiveBasicGAN with a BidirectionalLSTMDiscriminator (with minibatch discrimination) and BidirectionalLSTMDecoderGenerator. The full example is located at https://github.com/jchen42703/ai_mouse_movements/python/README.md.

Here are the model summaries:

---

[JS API]

Getting Started [Client]

cd js
npm install .
nodemon index.js

1. Install dependencies with npm install
2. nodemon index.js or node index.js to run the server on PORT=3000.
3. Send a POST request (json) to http://localhost:3000/, such as:

{
    "start": [1, 1],
    "destination": [82 ,55]
}

Model Format Conversion

From tf.keras to .json

pip install tensorflowjs
tensorflowjs_converter --input_format=keras model/weights.h5 model/tfjs_model

Dependencies

• @tensorflow/tfjs
• @tensorflow/tfjs-node
• express
• nodemon for convenience

How does it work?

1. POST request to https://localhost:3000/
2. express handles the POST request and calls the prediction function loadAndPredict.
3. The function returns a promise and when it resolves, the output is a list of coords and lags:

• [x, y, lag]
  • The lag is the time in ms that the mouse stays at that coordinate

{
  "coords": [
    [
      1,
      1,
      24.451885223388672
    ],
    [
      1.789207100868225,
      1.6034066677093506,
      23.39274024963379
    ],
    [
      2.462282180786133,
      2.276571035385132,
      24.84036636352539
    ],
    [
      2.7074904441833496,
      2.716768264770508,
      26.283510208129883
    ],
    [
      2.862687110900879,
      3.18359637260437,
      27.842201232910156
    ],
    ...
  ]
}

How fast is the api?

On average, it runs from 390ms to 430ms

• For cold starts: 500ms - 600ms