Train a network on MNIST

If you haven’t already, start by following the Setup Instructions.

A straightforward training run on MNIST can be run with:

python train.py --experiment_name mnist_example --dataset mnist --num_gpus_to_use 1 --max_epochs 20 --model.type ResNet9

The rest of the hyperparameters will be filled in via the default argparse arguments at the start of the train file: - the optimiser used will be Adam - the learning rate will be scheduled with CosineAnnealing - batch size will be 256 - … and many other choices to be aware of

If you don’t want to rely on all of these defaults, and don’t want to have to type out all of the arguments, you can run your experiment from a configuration file. For example, the above configuration could look like the following, put in a .yaml file in the folder experiment_files:

experiment_name: mnist_example
dataset: mnist
num_gpus_to_use: 1

# hyperparameters
max_epochs: 20

# model choices
model.type: ResNet9

And then run: python train.py -config experiment_files/mnist_config.yaml. You can also use JSON.