• Make a Server for Digit Recognition
• Make a client App for Digit Recognition

• Keras
• Nodejs
• Express
• Keras-js
• multer
• pngjs
• Other common python libraries : numpy, matplotlib etc
• Other common npm libraries : bodyparser, fs, path etc

ToDo

• Tensorflow
• Keras
• Nodejs
  Run npm install in Server/digit_reco_server to install :
  • keras-js
  • pngjs
  • multer
  • Other common nodejs libraries like express etc
• h5py - To enable storing models on disk
• matplotlib
• numpy

The server is built on Nodejs. The Multi-layer perceptron model is built using Keras.

cd Server/digit_reco_server/keras_model
python data_gen.py
python encoder.py models_generated/model.hdf5

Doing this will :

1. Download the MNIST dataset
2. Construct a simple multi layer perceptron model trained on MNIST dataset
3. Store the model in models_generated folder in a format that can be used by keras-js for prediction

cd Server/digit_reco_server
npm start

Properties of image to be uploaded :

• Dimensions : 28 X 28
• Format : png (with data in RGBA)

Images can be generated from MNIST dataset using Utilities/image_gen.py :

cd Utilities
python image_gen.py

Images are saved in Utilities/Images folder. These images can be uploaded to the server when launched. Any image uploaded must have same properties as those generated by image_gen.py

Upon uploading, the server returns the predicted digit in the image.

Route used : /imgupload

This route should be used by any client application that uses this server.

• Method : Post
• Request body : The PNG image
• URL : /imgupload

The predicted digit

To change or update the model used for prediction, ideally only the following files must be changed :

• data_gen.py and model_gen.py in Server/digit_reco_server/keras_model
• The function interpret_result() in Server/digit_reco_server/predictor/use_keras_model.js
  
  After making the required changes, the model must be compiled and saved as described above

ToDo