shape-learning

This small project is a essentially a "Hello World!" of image-based machine learning, except I went out of my way to generate my own images in bulk for fun. I used python3.9 for this exercise.

Quick Setup

Setting up the environment

In the root directory, create a virtualenv called .venv with the following:

$ python3.9 -m venv .venv

From there, activate the virtual environment using:

$ source .venv/bin/activate

After the environment is activated, the following will install the dependencies needed

$ pip install -r requirements.txt

Generating the images

Let's say we want to create 1000 images to train with, and 200 to test against. We can do the following in IDLE (or however else you might want to do it)

>>> from generate_shapes import create_all_shapes
>>> create_all_shapes(1000, 'training')
>>> create_all_shapes(200, 'testing')

The code above will

1. place 1000 rectangle images in shapes/training/0,
2. place 1000 ellipse images in shapes/training/1,
3. place 1000 triangle images in shapes/training/2
4. place 200 rectangle images in shapes/testing/0,
5. place 200 ellipse images in shapes/testing/1,
6. place 200 triangle images in shapes/testing/2

Now that the data is populated, shapes are classified in the program as

rectangle: 0
ellipse: 1
triangle: 2

From there, you can run the notebook (shapes_ml.ipynb) file using the jupyter notebook command at the root directory. You may notice that the accuracy isn't the best. How might this be improved? How much more accurate is the model if using 2000, 3000, or more training images per shape?

Some notes about the image generation

Currently, the images have the following attributes

1. 24x24 using 8-bit pixels, black and white
2. 1 shape per image, in a random location with random height and width
3. The shape stays fully inside the image