AI model from scratch in C++ for image classification (MNIST dataset)

• OpenCV
• Doxygen (only to generate the documentation)

cmake -S . -B build -G "Unix Makefiles" Then in build/ : make

Run the executable file in build/bin/

PDF

Accuracy: 87 %

• nbEpochs = 11
• batchSize = 64
• maxNbInstancesPerClass
  • training set: 6000
  • test set: 1000
• Network
  • Input: 28*28
  • Layers
    • Dense (512) LeakyRelu
    • Dense (10) Softmax
  • Learning rate: 0.03

1. Improve performance: add an option to use GPU with CUDA to increase the training speed. Use in the Tensor class the following code: https://developer.nvidia.com/blog/programming-tensor-cores-cuda-9/
2. Conv2D and Max-pooling layers
3. Confusion matrix
4. Accuracy graph
5. More loss functions

1. Run in the project root folder: doxygen
2. Open the doc folder and either open the LaTeX version or the HTML one (the index.html file)

Robin Meneust