This repository provides a comprehensive implementation of Neural Network in C++ for the task of Handwriting recognition i.e. identify the the handwritten digits in the images.
The repository is organized as follows:
src/Layers: Contains the implementation of differnet layers in our network.
- "activations.cpp" : Implementations of forward and backward pass of 2 activation functions namely - ReLU and Softmax
- "fullyconnected.cpp" : Implementation of forward and backward pass of fully connected layers of the network
src/Optimization: Contains implementation of network optimization methods
- "loss.cpp" : Implementation of forward and backward pass of cross-entropy loss
- "sgd.cpp" : Implementation of forward and backward pass of Stochastic Gradient Descent with Momentum
src/: Includes implementation of network optimization methods
- "eval.cpp" : Implementation of evaluation metric "Accuracy" to measure the model performance
- "mnist_data_reader.cpp" : Implementation of module for reading the data i.e. reading images and labels
- "network.cpp" : Contains the code to run the forward and backward pass of the full network architecture, which implements Fully connected and activation layers
-
src/mnist.cpp: It has the main function which triggers the project. -
src/test_image_reader.cpp: Function to test if the images are read correctly. -
src/test_label_reader.cpp: Function to test if the labels are read correctly.
We can test the implementation by running the following shell scripts with our own (and not publicly available) datasets:
build.sh: Contains all necessary code to prepare/build your executable(s)read_dataset_images.sh: Checks if you can successfully read in, convert and write an MNIST dataset image to file.read_dataset_labels.sh: Checks if you can successfully read in and write an MNIST dataset label to file.mnist.sh: Triggers the training and testing of your neural network implementation.