Brahmasifier: Deep Learning for Strong Gravitational Lens Classification

Author: Kunal Bhatia

Challenge: LSST Strong Lensing Data Challenge (2025)
Performance: 98.4% classification accuracy on validation set

Scientific Context

Strong gravitational lensing occurs when a massive foreground object (lens) deflects light from a background source, producing multiple images, arcs, or Einstein rings. With next-generation surveys like the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) expected to discover ~10⁵ strong lenses, automated classification methods are essential.

This repository contains deep learning architectures developed for the LSST Strong Lensing Data Challenge, achieving state-of-the-art performance on simulated LSST-like data.

Dataset

The challenge dataset consists of 200,000 objects in four categories:

• 50k SLSim lenses injected into LSST DP0 1-year coadds
• 50k SLSim non-lenses in DP0 data
• 50k Degraded HSC lenses (SIMCT + real SuGOHI sample)
• 50k Degraded HSC non-lenses

Image specifications:

• Dimensions: 41×41 pixels (0.2″/pixel)
• Bands: grizy (5 channels)
• Format: Multi-extension FITS files per band

Data generation details in Strong_Lens_Challenge_Data_Release.pdf.

Architecture

The classifier uses a Vision Transformer (ViT) or ConvNeXt backbone adapted for multi-band astronomical imaging:

Core modifications:

• Input layer: 5-channel (grizy) instead of 3-channel RGB
• Image size: 224×224 (upsampled from 41×41) for ViT, 41×41 for ConvNeXt
• Output: Single sigmoid unit for binary classification
• Pre-training: ImageNet-1K weights transferred to 5-channel input

Training strategy:

• Loss: Binary cross-entropy with logits
• Optimizer: AdamW (lr=1e-4, weight_decay=1e-2)
• Scheduler: Cosine annealing over all steps
• Augmentation: Random horizontal/vertical flips + rotations (90°, 180°, 270°)
• Precision: Mixed FP16
• Validation: Stratified 10% split

Performance

Metric	Validation Set	Test Set
Accuracy	98.4%	98.2%
AUC-ROC	0.997	0.996
Precision	98.1%	-
Recall	98.7%	-

Inference speed: ~500 images/second on NVIDIA A100 (batch size 512)

Repository Structure

Brahmasifier/
├── train.py                # Training script with distributed training
├── predict.py              # Inference and ensemble prediction
├── requirements.txt        # Python dependencies
├── docs/                 
└── README.md

Installation

git clone https://github.com/kunalb541/Brahmasifier.git
cd Brahmasifier
pip install -r requirements.txt

Dependencies:

• Python 3.10+
• PyTorch 2.0+
• PyTorch Lightning 2.0+
• timm (PyTorch Image Models)
• astropy
• torchvision
• torchmetrics

Training

Multi-GPU (DDP)

salloc --partition=gpu_mi300 --nodes=1 --gres=gpu:4 --exclusive --time=72:00:00

cd
cd sl
source sl_env/bin/activate

tar -cf - data/ | pv | tar -xf - -C /tmp/

export PYTHONWARNINGS="ignore"
export PYTHONUNBUFFERED=1
export OMP_NUM_THREADS=2
export MKL_NUM_THREADS=2
export HSA_FORCE_FINE_GRAIN_PCIE=1
export GPU_MAX_HW_QUEUES=8
export MASTER_ADDR=$(scontrol show hostnames $SLURM_NODELIST | head -n 1)
export MASTER_PORT=29500

srun -u --nodes=1 --ntasks-per-node=1 --gres=gpu:4 \
  torchrun \
    --nnodes=1 \
    --nproc-per-node=4 \
    --rdzv-backend=c10d \
    --rdzv-endpoint="${MASTER_ADDR}:${MASTER_PORT}" \
    --rdzv-id="train_$(date +%s)" \
    t.py \
      --mode train \
      --model-name convnext_base \
      --train-roots "/tmp/data/hsc_lenses" "/tmp/data/slsim_lenses" \
      --train-roots-neg "/tmp/data/hsc_nonlenses" "/tmp/data/slsim_nonlenses" \
      --index-cache "/tmp/index.pkl" \
      --val-split 0.20 \
      --epochs 100 \
      --batch-size 512 \
      --num-workers 4 \
      --cache-ratio 1 

Key hyperparameters:

• --model-name: Backbone architecture (convnext_tiny, vit_small_patch16_224, etc.)
• --cache-ratio 1.0: Pre-load entire dataset into RAM (recommended for fast training)
• --num-workers: DataLoader workers (default: 8)
• --prefetch-factor: Batches to prefetch per worker (default: 4)

Inference

Single Model

python predict.py \
  --run-dirs ./runs/convnext_tiny \
  --test-root /path/to/test/data \
  --out-csv submission.csv \
  --batch-size 512

Ensemble (Multiple Models)

python predict.py \
  --run-dirs ./runs/convnext_tiny ./runs/vit_small \
  --test-root /path/to/test/data \
  --out-csv submission.csv

RUN_DIRS="./runs/convnext_tiny" "./runs/convnext_base" "./runs/convnext_small"

python -u predict.py \
    --run-dirs $RUN_DIRS \
    --test-root "./tmp/data/test_dataset" \
    --batch-size 512 \
    --num-workers 8 \
    --out-csv "./recreated_probabilities.csv"

Ensemble strategy:

• Test-time augmentation (original, horizontal flip, vertical flip)
• Average logits across all augmentations per model
• Average ensemble logits across all models
• Final prediction: σ(logits_ensemble)

Data Format

Input FITS files must follow the challenge naming convention:

DX_IYYYYYYY_b.fits

• X: 1 (SLSim) or 2 (HSC)
• I: L (lens) or N (non-lens)
• YYYYYYY: Object index
• b: Band (g, r, i, z, y)

Each FITS file contains:

• Primary HDU: Empty
• Secondary HDU: BinTableHDU with columns:
  • Object ID
  • band_X: 41×41 image array

Technical Details

Data Preprocessing

1. Load 5 bands from separate FITS files
2. Stack into 5-channel tensor (C=5, H=41, W=41)
3. Per-image standardization: (img - mean) / std
4. Resize to 224×224 for ViT (bilinear interpolation)
5. NaN/Inf handling: Replace with 0.0

Caching Strategy

With --cache-ratio 1.0, images are:

1. Loaded from FITS once per rank
2. Pre-resized to target resolution
3. Stored in memory for entire training

Memory requirements:

• 200k images × 5 bands × 224² × 4 bytes ≈ 100 GB RAM (for ViT)
• 200k images × 5 bands × 41² × 4 bytes ≈ 7 GB RAM (for ConvNeXt)

Distributed Training

• Strategy: PyTorch Distributed Data Parallel (DDP)
• Synchronization: All-reduce for gradients and metrics
• Index cache: Rank 0 builds index, others wait via file lock
• Data cache: Each rank loads its assigned subset

Ablation Studies

Configuration	Val AUC	Training Time
ConvNeXt-Tiny (41×41)	0.995	2.5h (4×A100)
ViT-Small (224×224)	0.997	6.1h (4×A100)
ConvNeXt + TTA	0.996	-
ViT + TTA	0.998	-
Ensemble (3 models)	0.999	-

Citation

If you use this code, please cite:

@misc{batra2025brahmasifier,
  author = {Bhatia, Kunal},
  title = {Brahmasifier: Deep Learning for Strong Gravitational Lens Classification},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/kunalb541/Brahmasifier}
}

Acknowledgments

• LSST Strong Lensing Science Collaboration for organizing the challenge
• SLSim team for simulation pipeline development
• University of Heidelberg for computational resources

License

MIT License - see LICENSE file for details

References

See sl.bib for complete bibliography