esi-fastlight is a compact conditional continuous-normalizing-flow (CNF) framework that synthesises Proximity-Focusing RICH (pfRICH) detector hitmaps entirely in PyTorch.
- Generate physics-realistic hitmaps in microseconds per event – no Geant4 installation or GPU ray tracing required
- Stress-test tracking / clustering algorithms with unlimited synthetic events
- Train downstream ML models on diverse, physics-consistent data
# clone
git clone https://github.com/BNLNPPS/esi-fastlight
cd esi-fastlight/
# minimal dependencies (Python ≥3.9)
pip install torch>=1.13 torchdiffeq matplotlib numpy
# train the unconditional CNF (≈ 8 min on an RTX 4090, ~1 h on laptop GPU)
python3 train_cnf_amp_fast_compat.pyDuring training, PNGs in ./progress/ compare real (Geant4 / Opticks) vs generated hitmaps every 20 epochs.
| Component | Details |
|---|---|
| Model | Compact CNF (Neural ODE) implemented in pure PyTorch + torchdiffeq |
| Conditioning | None yet – prototype learns unconditional X-Y response; a conditional version is provided for multiplicity-aware sampling |
| Architecture | 3-layer MLP, widths 128 → 128 → 64, SiLU activations (≈ 100 k parameters) |
| Prior | Mixture of thin ring (Cherenkov rim) + tight Gaussian centre + wide Gaussian noise |
| Training objective | Maximum log-likelihood (negative log-probability of mapped samples) |
| Optimisation | AdamW (lr 3 × 10⁻⁴), gradient clipping 0.5, AMP mixed precision (torch.cuda.amp) |
| Data | Parsed (x, y) hit positions from opticks_hits_output.txt; split 70 / 15 / 15 % into train / val / test on event granularity (reach out to author for data) |
| Speed-ups | Fixed-step RK4 (≈ 20 NFEs) instead of adjoint Dopri5; smaller network; AMP; optional GPU |
python3 **`TrainingCNF.py`**The script SynthEvents.py samples tens-of-thousands of full events with realistic hit multiplicities while keeping GPU RAM low.
# prerequisite: a trained conditional checkpoint + counts.npy multiplicity histogram
python3 SynthEvents.py \
--ckpt cnf_condN_iso.pt \
--num-events 50000 \
--output events.npz \
--hits-per-batch 250000 \
--chunks 3What the script does
- Streams latent samples in manageable batches through a Sampler ODE (no log-probability tracking).
- Writes hits to a memory-mapped binary file and stores an index array → events.npz wrapper.
- Saves 10 quick-look PNGs in ./synth/ to visually verify the synthetic data.
Blue – real hitmaps from validation data • Orange – synthetic events from the CNF.
Only spatial hit distributions are modelled; time-of-hit conditioning will arrive in a future release.