Assignment 2 Report: AI Art

Assignment 2 Report: AI Art

Basically: Recreate Images From Triangles

What is a chromosome's representation?

class Chromosome:
    def __init__(self, painting: Painting):
        self.painting = painting
        self.fitness = score(self.painting)

    def set_fitness(self, fitness):
        self.fitness = fitness

    def get_score(self):
        return self.fitness

    def calculate_fitness(self):
        self.fitness = score(self.painting)

My Chomosome's representation is a class that contains a "painting", which is a data structure consisting of a predefined base (background) created by a fixed number of triangles (genes) and a fitness score. The fitness score is calculated right after instantiation.

The size of the painting is the same size as the target image

Population Size and Selection technique

Population Size: Started with 100 chromosomes but I find the algorithm gets stuck to at local maximum after roughly 80 generations. After increasing the population to 250, the gain still exists after roughly 200 generations. I decided to optimize the mutation rate and keep a 100 chromosome population.

Selection method: At first, I select the top 2 with the lowest fitness score (I'm minimizing the difference) to repopulate the population but got stuck again in a local maximum, the gain stopped showing after 15 generations. Afterward, I select the best and one random chromosome in the population (they both survive the elimination round)

The Fitness Function

The fitness function is the histogram difference from the target image.

How crossover/mutation is done

Crossover: Two "painting" is chosen, one best, one random, the child inherit half of the gene (triangle) from each of the chosen parents.

Mutation: A number of the triangles (genes) in a chromosome will be reset to their original position, change color, and change position based on 2 input arguments.

Results

Test Images are the ones furthest to the right.

From left to right: 1, 700, 1400 generations. Roughly 2 hours to completely "evolve" 1400 generations for each picture, I stopped at 1400 generations.

Further Comment

Girl With Pearl, 287 generations

I tried experimenting with different survival/mutation rates, with survival rate less than 20% and 61% mutation rate, the solutions quickly converge to local maximum and there was no gain after roughly 280 generations.

I found success with 60% survival rate and mutate all 40% of the offsprings, in the later stage of the program, gains appear within roughly every 12 generations.

Running the code

1. To run the code in this repository clone it, set up a virtual environment and install the required packages from requirements.txt

git clone https://github.com/pierrepicaud/recreate_images_from_triangle.git
cd recreate_images_from_triangle
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Put the image you want to simulate in the ".\img" folder, and change the target_image_path in main.py

3. Run main.py using

python main.py

The paths to the target image and output directory are hard-coded, but can easily be changed. the lines are in the main routine.

    target_image_path = "./image/lamp.png"
    checkpoint_path = "./output/"

Credit

I took the Triangle class and the Painting class from here and implemented the genetic algorithm myself, which is the whole point of this assignment.