mnist-digit-classifier

A digit classifier based on the convolutional neural network built using the TensorFlow library and trained on the most popular MNIST dataset.

The architecture of the neural network, i.e. the settings of each convolutional layers such as its strides, activation function etc., is taken from this kaggle article. The model is trained using the stochastic gradient descent, with the epoch iteration of 15, batch size of 16, and sparse categorical cross entropy as its loss function.

The model is pretty accurate, however, when I tested my own handwriting, it was sometimes failing giving me the incorrect result. It has managed to learn important characteristics of individual digits, and when drawing, you must make sure it is clear for it recognises them. Follow the preview below, and it can guarantee you some success.

Technology

Front-end

• JQuery
• HTML/CSS

Back-end

• Flask
• AWS Lambda Serverless (Deployment)
• Tensorflow

Deployment

• Visit my demo page for live action!
• This deployment was made possible by AWS Lambda service with AWS API gateway as the trigger. The source of Lambda function handler can be seen in the path model/lambda_app_model_inference.js and requires @tensorflow/tfjs-node package, which can installed through npm i from the model root directory. The generated directory must exist in the same directory level as the script (in model directory) as it has the trained model files generated from the Python Tensorflow (read below for more information).

Preview

Result

Property	Value
Accuracy	~98.02%
Loss	~6.6%
Validation Accuracy	~99.1%
Validation Loss	~2.85%

Build

Create a python virtual environment by following this article. Once the environment is activated, install the dependencies:

pip install -r requirements.txt

Then start the flask server

python server.py

The actual model is located in its model directory. To generate a new model:

python generator.py