Roman-Supernova-PIT/transient-real-bogus

Roman SN PIT — Real/Bogus Transient Classifier Ensemble

Binary real-vs-bogus classifier for difference-image cutouts from Nancy Grace Roman Space Telescope supernova simulations, developed for the Roman SN PIT (Roman Supernova Project Infrastructure Team).

Given a 64×64 difference-image cutout centered on a candidate detection, each model predicts the probability that the candidate is a real astrophysical transient (PSF-injected point source) as opposed to a bogus detection (subtraction artifact / noise peak from the peak-finder).

Ensemble structure

6 architecture families × 4 independently-trained members = 24 models. Each family lives in its own subfolder, with its own model card describing that architecture and its 4 members:

Folder	Backbone	Pretrained	Mean val. balanced acc. (4 members)
DenseNet169/	Custom from-scratch DenseNet (growth_rate=32, blocks=6-12-32-32)	No	96.06%
ResNeXt50/	timm resnext50_32x4d	ImageNet	96.72%
RegNetY016/	timm regnety_016	ImageNet	97.21%
EfficientNetB0/	timm efficientnet_b0	ImageNet	97.69%
ConvNeXtTiny/	timm convnext_tiny	ImageNet	96.81%
DeiTTiny/	timm deit_tiny_patch16_224 (img_size=64)	ImageNet	92.92%

Validation accuracy is the best-epoch balanced accuracy on the held-out validation split, as logged in each checkpoint (val_acc key).

Task & data

• Input: single-channel FITS difference-image cutout, resized to 64×64 if needed, normalized with astropy ZScaleInterval then min-max scaled to [0, 1], replicated to 3 channels → tensor of shape (3, 64, 64).
• Label 1 (positive): PSF injected into a real difference image at a known position/SNR (psf_injection_script.py, SNR sampled uniformly in [3, 10]).
• Label 0 (negative): peak-finder detections (≥3σ) on unmodified difference images that do not correspond to an injection (find_peaks_above_k_sigma_training.py).
• Output: single sigmoid unit, i.e. P(real transient) in [0, 1].

Loading a single checkpoint

Each .pth file is a dict: {'epoch': int, 'model_state_dict': ..., 'val_acc': float}. See the per-family README for the exact nn.Module definition needed to load_state_dict, e.g.:

import torch

ckpt = torch.load("EfficientNetB0/EfficientNetB0_Ensemble_Model1_best.pth",
                   map_location="cpu")
model = create_efficientnet(num_classes=1)   # see EfficientNetB0/README.md
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
print(ckpt["epoch"], ckpt["val_acc"])

Ensemble inference

For the full 24-model ensemble, average the sigmoid outputs of all loaded members (optionally restrict to a subset of families):

import torch

probs = torch.stack([m(x) for m in all_24_models])   # each m(x) -> (B,)
p_real = probs.mean(dim=0)

Training details (shared)

• Optimizer: AdamW, cosine-annealing-with-warm-restarts schedule.
• Per-member seeded WeightedRandomSampler for class balance and ensemble diversity across the 4 members of each family.
• No data augmentation — tensors are cached in RAM at dataset load time.
• TRAIN_EPOCHS = 15 (CNN families), DEIT_TRAIN_EPOCHS = 30 (DeiTTiny).
• Early stopping on best validation balanced accuracy; checkpoint saved on every new best.

Training/evaluation code: training_script.py, evaluate_injection_pipeline.py (Roman SN PIT pipeline repository — not included here).

Citation

If you use these models, please cite the Roman SN PIT project. Citation details to be added.