The final project and supporting work for the CS404 Artificial Intelligence paper. This project incorporates a neural network from scratch based on Kinsley, H., & Kukieła, D. (2019). Neural Networks from Scratch in Python. The implementation can be found in the following file:
- PneumoniaImagesANNScratch.ipynb
The purpose of this project was to explore how capable was a properly designed classic Artificial Neural Network in image classification compared to Convolutional Neural Networks. Our approach consists of the following:
- Input layer - 2304 Nodes
- 1st Hidden layer - 1024 Nodes
- 2nd Hidden layer - 512 Nodes
- 3rd Hidden layer - 256 Nodes
- Output layer - 1 Node
The number of nodes within the network is purely dictated by the resolution of the image, we found that a 48x48 image resolution was adequate (with the images being gray scaled), as higher resolutions resulted in longer runtimes without significant gains in accuracy. Moreover, we used the TanH activation functions throughout the network except for the output layer, where Sigmoid was used.
Random sampling without replacement was exclusively used for selecting the data from the original data, then for selecting the batches. While one single batch cannot contain two of the same images, multiple batches can have overlapping images. The model is set to train for exactly 40 Epochs. We achieved the best results with this configuration. Lastly, a fixed 46 batch size was used.
- Best results - 91% Accuracy, 0.31 Loss
- Average results - ~83% Accuracy, 0.37 Loss
In order to run and test this project, the following is needed:
- Latest version of Python3
- Latest version of JupyterLab and Jupyter Notebook
- The images from the dataset that can be found here.
- Latest version of Git (Optional)
- All packages at the start of the file
You can either use the following command in your terminal or just download it directly from GitHub.
git clone https://github.com/EdinZiga/PneumoniaImagesANN
Make sure to remember where you clone it as you will need it in the following steps.
The dataset that can be found here. Extract the file into the cloned repository.
Navigate to the repository within your terminal, and open it using Jupyter Notebook.
C:\...\PneumoniaImagesAnn>jupyter notebook
Make sure that you have all the necessary packages installed before running to ensure no errors appear. Packages used can be found below:
import os
import numpy as np
import pydot
import graphviz
import random
import matplotlib.pyplot as plt
import pandas as pd
from PIL import Image
from tqdm import tqdm
from sklearn.metrics import confusion_matrix, roc_curve, auc
from sklearn.metrics import ConfusionMatrixDisplayNavigate to the load images section. There are 4 directories that have to be set manually. Make sure they align with the location of the downloaded images.
directory = r"D:\PythonProjects\PneumoninaANN\chest_xray\train\NORMAL"
directory = r"D:\PythonProjects\PneumoninaANN\chest_xray\train\PNEUMONIA"
directory = r"D:\PythonProjects\PneumoninaANN\chest_xray\test\NORMAL"
directory = r"D:\PythonProjects\PneumoninaANN\chest_xray\test\PNEUMONIA"The network from scratch takes a bit to run, depending on your setup, that might be a while. Prepare some reading material.
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change. The best way to contribute is to run this on your local machine and see if you can get better results than us. Different network tunings are also welcome.
GNUv3 Feel free to use for whatever you wish, give credit where credit is due.