Neural network in C++ from scratch for MNIST dataset classification

The build requires c++14

Building should be pretty straightforward, given all dependencies are present on the correct location (which makes the build, actually, pretty not straightforward to setup)

I recommend checking the CMake files to check where the cmake searches for the dependencies. The libraries used by the project include:

• boost
• nlahmann::json
• xtensor
• xtensor-blas

It is best to install them via conda.

Do not forget to set (if not already) your CONDA_PREFIX environment variable to the path where your conda is installed (usually this will be set automatically after conda installation, e.g. $HOME/anaconda3 or $HOME/miniconda3), cmake will look for it.

All of which are open sourced, well maintained and well document.

Run the following build command from the project root dir:

cmake -DCMAKE_BUILD_TYPE=Release .  # to configure
cmake --build . --target all -- -j 4

Once the build has completed, there should be build directory in the project root. Also, there should have been two links called mnist and mnist-evaluate. The mnist is the allmighty training and evaluation script. It should take care of downloading the data set, training the model and evaluating it on the test set.

Most training parameters can be customized via environment variables. I recommend checking the source code of model.h header file for most of them.

Some of the parameters include:

• LEARNING_RATE=3.0
• TRAIN_EPOCHS=30
• BATCH_SIZE=10
• LOSS="mse"

You can also set LOSS to "categorical_cross_entropy", the performance is similar (altho "mse" appears to be slightly better, it might be the case of hyper param tuning) in the long run, however x-entropy seems to converge faster (strikes better performance earlier). In case of cross entropy, I also recommend setting higher BATCH_SIZE. Something like 30` should do.

and some additional logging parameters

• LOG_STEP_COUNT_STEPS = 30000

It is also possible to continue training, as the model automatically creates checkpoints after each epoch. This is possible by setting the CONTINUE_TRAINING env variable.

To run the training script, execute from the project root:

./mnist

Evaluation happens after the training, but it is still possible to run evaluation on a pre-trained model, which should have been exported into the export/ dir after the training along with its checkpoints.

With the default setting, the model should read over 95% accuracy. That's not exactly state of the art, but it is quite impressive considering the simplicity of the model.

./mnist-evaluate

UPDATE: There has been recently added RUN script, which builds the project automatically and allows to set environment easily and runs the training and evaluation.

TODO: This has not been implemented yet. Sorry for the inconvenience.

My greatest thanks goes to the QuantStack project for their work on xtensors and their incredible support which I needed quite often during creation of this project.

Some parts of backpropagation algorithm were inspired by the great blog Neural Networks and Deep Learning by Michael Nielsen. The blog has been invaluable to me, thank you very much for this!

An awesome explanation of cross entropy and softmax along with their derivatives was provided by Sefik Ilkin Serengilin this blog

Author: Marek Cermak [email protected]