This repository contains a educational implementation of a generic deep neural network in Julia. Its purpose is to facilitate the understanding of how deep neural networks work, both for the feed-forward and the back-propagation. It comes from a more specific version initially developed for the Neurocomputation course at UAM, which imposed the structure of the library for the neural network (check it here). Due to this structural imposition for educational purposes, we could not make the library as efficient as possible (via matrix multiplications instead of for loops). However, it is still very efficient thanks to Julia, since it is 10 times faster than the Python implementation that the rest of the students did.

You can download and install Julia in this link.

Once Julia has been installed, you have to install the ArgParse package via the following command:

julia -e 'using Pkg; Pkg.add("ArgParse")'

Once you have installed Julia and its ArgParse package, you are ready to train the neural network.

To display the available arguments that can be provided to customize the training, you can execute the following command:

julia Backpropagation.jl -h

As you can see, there are three modes of execution:

• Mode 1: the program reads the data in the input_file and divides it in train and test split according to the specified percentage.
• Mode 2: the program reads the data in the input_file, considering all the data both as the train and test sets.
• Mode 3: the program reads the train data in the input_file and the test data in the input_test_file.

To further illustrate the usage of the neural network, we show below a sample command:

julia Backpropagation.jl --input_file data/problem_2.txt --output_name problem_2_example --learning_rate 0.01 --epochs 500 --mode 1 --net_config '[15, 20, 15]' --percentage 0.7

Finally, we provide a Python script (plot_stats.py) to plot the evolution of the MSE and the Accuracy in the train and test set over the training epochs.