Tiny-ImageNet

This project seeks to find an optimal model for classification on the Tiny-ImageNet dataset

Directory Structure

configs/                # YAML files defining tuning experiments
slurm/                  # SLURM job scripts for tuning
notebooks/              # Jupyter notebook code
utilities/              # Utility code
scripts/                # Python standalone scripts
pyproject.toml          # Builds Python modules
requirements.txt        # Defines non-torch requirements

Constraints

To increase the difficulty of this challenge, the final model is capped at 15M total parameters and only 50% of the training data is to be considered. Additionally, no pretrained weights were accepted.

Results

Constrained

The final model with the proper constraints achieved a final testing accuracy of 57.66%. This model was the product of a myriad of tuning processes outlined in the final report. The final run notebook can be viewed at notebooks/final.ipynb

Unconstrained

For exploration of the final model, the model was trained again with the full dataset. The final accuracy of this model was 64.38%. The final run notebook can be viewed at notebooks/unconstrained_final.ipynb.