Software Track

software-track/
├── src/
│   └── bitbybit/
│       ├── kernel/          # hash‑kernel abstractions
│       │   ├── _base.py     # _HashKernel (provided)
│       │   ├── random.py    # RandomProjKernel (implement)
│       │   └── learned.py   # LearnedProjKernel (implement)
│       ├── nn/              # hash‑backed torch.nn layers (provided)
│       ├── utils/           # helpers (provided)
│       └── patch.py         # swaps torch.nn ⇄ bitbybit.nn
├── train.py                 # minimal training template
├── publish.py               # leaderboard uploader
├── config.py                # hashing layerwise config
└── requirements.txt

Installation

Do this step before installing other packages:

sudo apt-get update
sudo apt-get install -y \
    python3-dev \
    libjpeg-dev \
    zlib1g-dev \
    libpng-dev \
    libtiff-dev \
    libfreetype6-dev \
    liblcms2-dev \
    libwebp-dev \
    libharfbuzz-dev \
    libfribidi-dev

Then do this:

uv venv --python 3.12
source .venv/bin/activate
uv pip install -r pyproject.toml

Baseline Models

Weights are fetched from https://github.com/chenyaofo/pytorch-cifar-models using the keys

"cifar10_resnet20"
"cifar100_resnet20"

Task Checklist

1. Implement kernels: complete RandomProjKernel and LearnedProjKernel.

2. Train: extend train.py to fine‑tune projections (and optionally the backbone).

3. Submit: save checkpoints to

   software-track/submission_checkpoints/<model>.pth
   

   then run

   python publish.py \
     --team-name <team-name> \
     --key <pre-shared-key>  

Submissions are unlimited within the 24‑hour window; the server keeps your best score.

Scoring

Evaluation lives in bitbybit.utils.score.calculate_submission_score.

Run Scripts

export PYTHONPATH="$PWD/src"