If you're looking for a great deep learning framework, then this isn't it. This project started out as a bit of an experiment in using the Eigen linear algebra library and trying to write a non-trivial C++ program, which is a language I don't use a lot in my day-to-day work.
Once compiled this program will load MNIST digit data (or any files which use a similar format), perform one-hot encoding of the labels and then train a single hidden layer ANN against these data using back-prop.
It'll display the error rate, and then attempt to classify the first 20 entries of the test set.
- Eigen
- zlib
- g++
- CMake 3
Currently this works with g++ on Linux and WSL. I've tried getting it to work with Visual Studio 2017 /
Windows 10, but zlib and endian coversions behave funny. clang should be fine, but I haven't tried it yet.
In the directory where you have cloned the repo, create a sub-directory build. Inside this
sub-directory run cmake
$ mkdir build
$ cd build
$ cmake ..
This should produce an executable named train
To run this, you will need the MNIST data set. Assuming you have downloaded this into ~/data, run
the training process as follows:
./train ~/data/t10k-images-idx3-ubyte.gz ~/data/t10k-labels-idx1-ubyte.gz
- Add a softmax output layer.
- Give more control over the learning rate, number of layers, etc.
Make it easy to choose between ReLU and sigmoid activation functions.- Annealing, adaptive learning rate, all the things that actually make neural nets viable.
- General refactor to allow the components to be tested.