Neural network implementation in C++ for handwritten digit recognition using the MNIST database. Use the program to train, test and visualize results of a neural network trained using MNIST data. The program allows to set the layer depth and sizes in the input arguments. The algorithm implements a cross-entropy cost function with regularization on a sigmoid neuron network. For convenience, the MNIST database is included in the repository. For more details, see the option available in this program.
Usage (examples): train example: ./Neural_network -train 3 {784,30,10} -param {10,0.01,0.01} -reduceLearning {2,20} test example: ./Neural_network -test 1 -display
Packages used:
- Armadillo (linear algebra)
- OpenCV (visualization)
| Options available | |
|---|---|
| -train | Train a new neural network. This mode requires the training set and labels. See training options below for more details. |
| -test | Test a trained network. This mode requires a trained network stored in Results_Network and the test set. After '-test' refer to the folder containing the results by the trailing number in the folder name, e.g. '-test 1' to test the performance of the network stored in 'Network_Results/Results_1'. See test options below for more details. |
| Train options | |
|---|---|
| -layers | Set the total amount of layers and layer sizes used in the network, including the input and output layer. After '-layers', the total number of layers is required. Thereafter, the layer size should be given in curly brackets, e.g. 'layers 3 {784,30,10}'. |
| -param | Set learning hyperparameters. Parameters which are to be set are: batch size before learning step, learning rate, and the regularization parameter, respectively. In case no regularization is to be used, the parameter is to be set to zero, e.g, '-param {1000,0.1,0}'. |
| -reduceLearning | Used to reduce the learning parameter by {factor x, per y epoch}, e.g. -reduceLearning {2,20}. |
| Test options | |
|---|---|
| -display | Opens a window to visualize the test images in a random sequence. Visualization can be stopped by pressing q. |
[](http://codereview.stackexchange.com/q/241074/222703)