Consider the MNISTDataset class, which provides access to the MNIST dataset. The constructor of this class takes two parameters: one is the name of the file contains images, while the other is the name of the file that contains the labels. It loads whole files into its memory, which is not a best practice, but for this dataset, this approach works well because the dataset is small. For bigger datasets, we have to implement another scheme of reading data from the disk because usually, for real tasks, we are unable to load all the data into the computer's memory.
We use the OpenCV library to deal with images, so we store all the loaded images in the C++ vector of the cv::Mat type. Labels are stored in a vector of ...
