Skip to content

Getting Started

Goal: set up the project and run a first optimization pass end-to-end.

Prerequisites

  • Python toolchain managed with uv and a GPU-enabled JAX installation.
  • Access to a small dataset bundle for smoke testing.

Install the package

Choose an environment and install with uv: 

# (optional) create a fresh venv via uv
uv venv .venv
source .venv/bin/activate

# install runtime dependencies
uv sync

# install in editable mode if you plan to hack on the code
uv pip install -e .
If you only need the documentation toolchain, install the docs extras: 
uv sync --group docs

Quick usage

Run the optimization script on a small dataset to verify the loop: 

UV_CACHE_DIR=.uv-cache-local uv run --project PCT python PCT/bin/optimize_network.py \
  --dataset datasets/june_training_data/train_redp_1 \
  --epochs 2 --batch-size 64 --lr 1e-3 --model singletrack-augmented
This will write artifacts under results/singletrack-augmented/run_* (checkpoints, loss curves, analysis plots).

Steps

  1. Prepare data: place the chosen dataset bundle in datasets/ and confirm paths in the CLI args or config.
  2. Kick off a minimal run (above) to confirm the loop completes on your hardware.
  3. Inspect outputs: review logs, checkpoints, and metrics artifacts in the run directory.

Next up

  • Move on to the training loop walkthrough to understand the control flow and instrumentation in more detail.