Web app classsificator based on the Quick, Draw! Dataset.
Has Node.js and Keras version of the MobileNetv2 underhood. Generates predicts across 340 categories.
To clone and run this application, you'll need Git and Node.js (which comes with npm) installed on your computer. From your command line:
# Clone this repository
$ git clone https://github.com/Arcady1/Doodle-Recognition-Web.git
# Go into the repository
$ cd Doodle-Recognition-Web
# Install dependencies
$ npm install
# Run the app
$ npm build
$ npm start
npm dependencies:
"browserify": "^16.5.2",
"@tensorflow/tfjs": "^2.0.1",
"express": "^4.17.1",
"mathjs": "^7.1.0"
See more in this Jupiter Notebook with MobileNetv2 training pipeline.
Model type: MobileNetV2
Weights initialization strategy: random noise
Main hyperparameters:
- batch_size = 256
- alpha = 1
- input_size = (64, 64, 1)
The training took 6 hours on Tesla P100 (Google Collab).
The model folder also includes:
- A notebook with Imagent version of MobieNetv2 and input size (64, 64, 3)
- Keras models and weights converters from .h5 format to TensorFlow.js Layers format
MobileNetV2: Inverted Residuals and Linear Bottlenecks, arxiv article - Original article with the MobileNetv2 description
TensorFlow JS documentation - This article describe how to convert pre-trained Keras model to TensoFlow JS model
The Quick, Draw! Dataset - Dataset
MIT