Some scripts to experiment with image classification models on the cifar10
dataset. The script run_training.py runs a single trial as specified by the
configuration configs/trial_config.py. The script run_experiment.py runs a
sequence of trials as specified by one of the experiment configs in configs/.
When running an individual training run, the output (checkpoints and
tensorboard) is written to ./trials/<trial_name>. For experiments, the output
is at ./experiments/<experiment_name>/. This allows for an easy comparison of
the trials runs within an experiment.
The current models are a two layer MLP, and a small residual net. The existing experiment configurations are to find
- optimizer parameters (SGD) for overfitting quickly,
- regularization parameters (dropout for MLP, weight decay for both models) to find a non-overfitting solution.