README

Summary

Project name: Number Recognition

Video showcase: https://www.youtube.com/watch?v=Emt7cQfvIz8

This project is about predicting drawn number in HTML canvas (pixel art style) using machine learning. Can be seen as digit recognition project.

Stacks used:

Backend:

• Django
• Django-REST framework
• Scikit-learn

Dev dependencies:

• Jupyter notebook (ipynb)

Frontend:

• HTML5
• SASS
• Vanilla JS
• A tiny traces of Foundation by ZURB

How do I get set up?

• Use Python 3 (3.5)
• I'm using anaconda as python package manager
• There is environment.yaml in the root folder containing all python dependencies, but if you want to install only necessary stuffs, just install packages listed in Backend section above
• Here is how to install the requirement file in conda: conda env create -f path/to/environment.yaml
• After everything setup, run python manage.py runserver 0.0.0.0:8000 in terminal from the root folder, this will run the dev server in your localhost:8000 or 127.0.0.1:8000 and will be accessible by your LAN
• Then you are setup! you can start using it now

---

Usage

In http://root_url (e.g. http://localhost:8000), there are 2 functions, Predict Data and Train Data.

Predict Data

Predict Data will POST pixel data as features to http://root_url/api/ml-model/predict, and the server will use the machine learning model to try predict the number and gives the response with predicted number.

The machine learning model needs to be trained first in order for this application to works, if it has not been trained, it will return -1 value as the predicted number

To train the model, make a GET request to http://root_url/api/ml-model/train-model, there is optional parameter for predicted model to use which training data.

• By default it will use scikit-learn digit recognition training dataset for training
• To specify other source (at the moment, the only other option is by using csv file which are collected from Train Data option), make a request with keyword parameter "train-data-source" (e.g. http://root_url/api/ml-model/train-model?train-data-source=csv)

The response from server should return indication that the model has been successfully trained, once it is trained, it should be able to predict the number and returning the correct number (if the model is trained properly)

Train Data

• Train Data will POST pixel data and digit data as label and features to the http://root_url/api/ml-model/training-data. The server will then store it in list as well as append it to csv file.
• To see stored label and features pair in in-memory-python-list, send a GET request to http://root_url/api/ml-model/training-data. You can see the stored training data in the csv file by directly going to the num_recognition_training_data.csv at the root level of project

---

Notes: ./num_recognition_app/urls.py handle all routing for the number recognition app

---

Development Notes

• The num_recognition_training_data.csv file in the root level of project contains 800 training data set. The data came from 10x10 Canvas pixel with 2x2 brush, and they are not a good dataset (there are quite some inconsistencies) so it will be unusable for the current state of application.

• So instead I just used scikit-learn digit recognition dataset, and changing the canvas to 8x8 with 1x1 brush. But since many dataset from scikit-learn is gearing towards stretched vertical size (the height of drawn number is always full height of canvas, but the width is probably just half width of canvas, so the aspect ratio will always be 1:>1 (w:h)), in order to get the correct result, you need to draw it the same as how the training dataset is, draw the number with half width x full height of the canvas.

• Right now, it is only using SVM to predict number separation

• There is Training.ipynb file containing traces of progress of training using the num_recongition_training_data.csv file, which wasn't very successful to get high percentage of precision, recall and f-score. (Have tried SVM and neural networks from scikit-learn)

• sklearn_digit_test.ipynb contains experimenting around scikit-learn digit recognition digit dataset to see how to make it suitable for this application

---