EvDiff

Paper release for Conditional Diffusion Models for Energy-Efficient Driving.

This repository contains the executable notebooks used for preprocessing, training, ablation experiments, evaluation, and manuscript figure generation for the EvDiff conditional diffusion model. Large generated artifacts are intentionally excluded from version control.

Environment

Install the Python dependencies:

pip install -r requirements.txt

The experiments were run with PyTorch on a single NVIDIA T4 GPU. Training settings used in the paper: 100 epochs, batch size 32, 1000 diffusion steps, and fixed seeds documented in the ESI.

Paths

The notebooks support portable paths through environment variables:

export EVDIFF_EXP_ROOT=/path/to/evdiff
export EVDIFF_BERLIN_ROOT=/path/to/rucker_raw_dataset
export EVDIFF_DATA_ROOT=/path/to/evdiff/artifacts/original_berlin_rucker
export EVDIFF_DIRECT_TENSOR_DIR=/path/to/normalized_tensors_new
export EVDIFF_FIG_DIR=/path/to/figures/paper_draft

For a fresh run, start with preprocessing, then run the proposed model or ablation notebooks, then run the figure-generation notebook.

Repository Policy

Do not commit generated experiment outputs:

• artifacts/
• checkpoints/
• mlflow/
• predictions/
• figures/
• model checkpoint files such as *.pt

These files can be regenerated from the notebooks and external dataset.