Training neural networks using Flux.jl

using Flux

using CSV 

using DataFrames

using Random

using Statistics

using StatsPlots

# Generate a temporary file path

# Download data in the temporary file path from the UCI website

# https://archive.ics.uci.edu/ml/machine-learning-databases/00236/seeds_dataset.txt

# Read the seeds dataset

# Values are separated by one or more tabulation

# There are no missing values

# There are no column names

# Name the variables (measures of wheat kernels = grains) 

# 1 = area (A)

# 2 = perimeter (P)

# 3 = compactness (C = 4*pi*A/P^2)

# 4 = length of kernel

# 5 = width of kernel

# 6 = asymmetry coefficient

# 7 = length of kernel groove

# 8 = cultivar (1, 2 or 3) : variety of wheat

rename!(seeds,

  [:Column1 => :area, :Column2 => :perimeter,

  :Column3 => :compactness, :Column4 => :kernel_length,

  :Column5 => :kernel_width, :Column6 => :asymmetry,

  :Column7 => :kernel_groove, :Column8 => :cultivar]

  )

# Set seed for replicability

# Number of samples in training set

# Around 70% of data

# Indices of training and testing sets

# Training set: n unique random indices

# Testing set: other indices

# Training sets

# Testing sets

# Build training set for predictors (features)

# Build testing set for predictors (feautures)

# 1. Build training set for the predicted variable (cultivars)

# 2. Transform the cultivar variable into 3 columns (one-hot encoded)

			# Rows are types of cultivar

			# Columns are training samples

			# Sorting labels allows corresponding rows to refer to the same cultivar 

# 1. Build testing set for the predicted variable (cultivars)

# 2. Transform the cultivar variable into 3 columns (one-hot encoded)

			# Rows are types of cultivar

			# Columns are testing samples

			# Sorting labels allows corresponding rows to refer to the same cultivar

# Simple model 

# Fully collected layer of 7 features and 3 possible outputs

# Result: output node with the highest score (softmax)

# Untrained model

# Train the model with a gradient descent optimiser (to find local minimum)

# Low learning rate of 0.01

# Loss function (cross entropy)

# Data iterator to handle training epochs 

# Every element in data_e represent one epoch

# Train model 

# Accuracy

# Confusion matrix

# Predicted in rows, reference in columns

# Most of the values are on the diagonal (which is good)

# Add one hidden layer with 14 nodes

# Sigmoid activation in the input layer

# Define loss function 

# Data iterator to handle training epochs rather than looping

# Every element in data_e represent one epoch

# Train model 

# Accuracy

# Confusion matrix

# Worse than previous model