A poorly-named MNIST handwritten digit recognizer written in java with no dependencies.

git clone https://github.com/bhoener/JavaTorch/
cd JavaTorch

java .\gui.java

This will open a gui window that allows you to draw your own digits and have the network recognize them.

java .\train\train.java <run_name>

Where <run_name> is the name of the run, e.g. bob.

This will run the training script, which should achieve ~0.01 final MSE loss in a few minutes at the default settings, then display a few examples and their corresponding predictions, followed by a validation accuracy, which should be ~0.93. You can also train a slightly less accurate model (0.91 val accuracy) in just a few seconds with the following config:

{
    "hidden_dim": 20,
    "lr": 0.8,
    "lr_final": 0.02,
    "num_steps": 2000,
    "grad_accum_steps": 16
}

Feel free to mess around with these hyperparams in ./train/train.java and try to get a more accurate model.