Neural Network written in C (STD C11). The main goal is to understand artificial neural networks by building one from scratch and training over the MNIST dataset. As a separate goal, I tried to create the cnet interface as flexible as possible, with the possibility to work with several layers, activations and losses.
This is not a serious project, it hasn't been fully tested, nor it will be mantained. The objective was to understand anns and that's all. GENANN is a serious project that can be used, seriously.
The mnist folder contains code to test the cnet library over the mnist dataset.
It contains two main scripts: train and test. The train file contains code to train the model, the output is be saved into the mnist out folder., including the saved model and the history file (containing loss and accuracy).
The test file contains code to test the saved model over the test-set and creates different reports. The reports are saved into the mnist/out folder, and contains
- a confusion matrix saved into
mnist/out/conf.datthat can be visualized using this gnuplot script. - a classification report, which includes precision, recall, f1-score and support for each of the digits.
Currently, the model reaches 0.94 accuracy over the mnist test-set within 200 epochs (~5hs). The training is pretty slow as it uses SGD (batches with a single training sample) and does not use any optimization, such as momentum.
| accuracy | loss |
|---|---|
 |
 |
CLASSIFICATION REPORT
precision recall f1-score support
Digit 0 0.951533 0.981633 0.966349 980
Digit 1 0.973753 0.980617 0.977173 1135
Digit 2 0.948744 0.914729 0.931426 1032
Digit 3 0.931480 0.928713 0.930094 1010
Digit 4 0.925224 0.945010 0.935013 982
Digit 5 0.928981 0.909193 0.918980 892
Digit 6 0.933131 0.961378 0.947044 958
Digit 7 0.947628 0.932879 0.940196 1028
Digit 8 0.919140 0.921971 0.920554 974
Digit 9 0.936492 0.920714 0.928536 1009
Final Accuracy: 0.940300 - Samples: 10000
The generated files will be saved in the following folders:
- bin/obj: object files for library
- bin/lib: cnet static library
- bin/exec: all executable files
The Makefile provides the following targets:
- cnet: Builds the cnet static library
- integration-tests: Builds a quick integration test
- mnist-train: Trains a model on the mnist dataset (see the mnist section)
- mnist-test: Uses the saved model to predict over the mnist testset (see the mnist section)
The project builds a static library that provides several functions, these will all start with the cnet_ (general purpose functions) or nn_ (network specific functions) prefix and they can be found in the cnet header. The most important functions are:
- nn_init: intialize a cnet model
- nn_free: free the initialized memory for a cnet model
- nn_add: adds a layer to the model
- nn_predict: predict over a single sample
- nn_train: trains the model over the given hyperparameters, this function also saves the history into a given file. This history can be displayed using the metrics plot script using gnuplot.
- nn_save: save the model into a given file
- nn_load: load the model from a given file
To create this project I used the following resources as guide: