Eusipco results

Experiments following the paper "Conditional Generative Models for High-Resolution Range Profiles: Capturing Geometry-Driven Trends in a Large-Scale Maritime Dataset", E. Brient, S. Velasco-Forero, R. Kassab submitted at Eusipco 2026

Cleaned-up DDPM / GAN training code for HRRP radar profiles.
Executable code lives in src/ship_hrrp_gen, configs in configs/.

Quick catch up on the paper

This repo follows the paper Conditional Generative Models for High-Resolution Range Profiles: Capturing Geometry-Driven Trends in a Large-Scale Maritime Dataset (EUSIPCO 2026 submission).

1) Conditioning strategy (what is controlled)

The paper compares conditioning variants and shows that geometry-based conditioning is the key driver:

• unconditional: no metadata;
• aspect-only (asp): heading vs radar LOS;
• dimensions-only (dims): ship length/width;
• aspect + dimensions (scals): best overall compromise.

Main takeaway: adding dimensions strongly improves coarse structure realism; combining aspect + dimensions gives the best global behavior.

2) LOSP/LRP reminder (coarse geometry prior)

The theoretical projection length along line-of-sight is:

$$ \mathrm{LOSP}(L,W,asp)=|L\cos(asp)|+|W\sin(asp)| $$

In practice, we estimate an empirical LRP (Length on Range Profile) from the HRRP envelope. The important point is the trend agreement: generated LRP should follow LOSP across aspect angle.

3) Results at a glance

Quantitative summary (paper-style table)

Generated HRRP examples

As in the paper, it is normal that nearest-match comparisons are not perfect at fine scale (HRRP is highly variable and noisy). The key objective here is to match coarse-scale geometry-driven structure.

LOSP/LRP trend on generated data

This figure is used to verify that generated profiles reproduce the expected LOSP/LRP geometric trend, including variability across ships and aspect angles.

Quick requirements

• Python ≥ 3.9
• PyTorch (CPU or CUDA, matching your GPU if available)
• The generated demo set (data/ship_hrrp.pt).

Installation

cd github_repo
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e .   # installs the ship_hrrp_gen package from src/

Run a training

python -m ship_hrrp_gen.train \
  --config configs/gan_scalars.yaml \
  --seed 42

Useful flags:

• --config: direct YAML path if it exists; otherwise resolved from the workspace config/ directory.
• --data-path: optional dataset path; defaults to data/ship_hrrp.pt when the CSV dataset is absent.
• --iteration: experiment iteration id used in logger naming.
• --seed: random seed for train/validation/test split.
• --fast-dev-run: one short local sanity run on a small subset.
• --skip-eval: skip final checkpoint reload and test metrics.

Quick sanity check

python -m ship_hrrp_gen.train \
  --config configs/gan_scalars.yaml \
  --fast-dev-run \
  --seed 0

Artifacts (checkpoints, figures, TensorBoard logs) are written under results/ following the figure_path in the config.

Layout

• src/ship_hrrp_gen/: models (DDPM, GAN), dataset, utils, training script (train.py).
• configs/: the eight supported YAML configs: ddpm, ddpm_scalars, ddpm_va, ddpm_dims, gan, gan_scalars, gan_va, and gan_dims.
• requirements.txt: Python deps.
• .gitignore: ignores caches, venvs, and training outputs.

Notes

• Intended for a quick demo run on data/ship_hrrp.pt; no CSV dataset or multi-GPU setup is required.
• Final metrics rely on compute_metrics in ship_hrrp_gen.utils. If there is no test split (test_idx empty), use --skip-eval.