Simple 3 layer artificial neural network for classification of iris plants.

The motivation for this project was practice for skills learned in basic deep learning course. Apart for implementation of ANN model, some basic data pre-processing techniques, such as normalization and one-hot encoding was used. It also was aimed for improving my mathematical understanding of forward run of Neural Network, Back-propagation, Activation Functions and Hyper-parameters.

Data is from Open source "Iris Data Set" from UCI, Machine Learning Repository. But I used the version available on Kaggle Link. I used this version as data was in *.csv format rather then *.data format on UCI repository. I found it easy for *.csv to use.

Please keep downloaded data in the same folder as python files.

As this project was for learning purpose, I have written code in both in simple Numpy, and Tensorflow framework. Tensorflow code was very low level, with just simple implementation of ANN.

The repository contains,

1. Iris.csv Contains data
2. iris.names Contains information about data set. i.e. its attributes, publications etc
3. main.py Contains models using numpy
4. main_tf.py Contains models using Tensorflow
5. utils.py Contains utility functions e.g. get_data, sigmoid, softmax

1. Model developed using Numpy The model is developed using class definition with name ANN, the class as function, fit: to fit model to data forward: for forward loop predict: to find prediction y given x score: to find error of model for certain data.
2. Model developed using TensorFlow The model is developed using class definition with name ANN_TF, the class is based on simple tensorflow structure.

To run the simple numpy file

python main.py

To run the Tensorflow based file

python main_tf.py

Results from model ANN are better as we can use regularization techniques for this model. but for model ANN_TF the result is not so good. ANN gives accuracy ~= 0.9 and ANN_TF ~= 0.8. The learning rate can be given as input to model to check its response.

I would like to give credit to lazyProgrammer for developing a very comprehensive and understandable course for deep learning. Also UCI for making the Iris Data Set available. Also Azeem Bootwala for introduction to this data set and encouragement.

MIT ©️ Jalil Ahmed.