SMART reproduction (ICML 2026, IkPdV0stxX)
Collection
Artifacts for the Trackio reproduction of 'Sample Margin-Aware Recalibration of Temperature Scaling'. • 3 items • Updated
ResNet-50 / CIFAR-100 (Mukhoti et al. recipe) — SMART reproduction backbone
Trained from scratch for the reproduction of Sample Margin-Aware Recalibration of Temperature Scaling
(ICML 2026, OpenReview IkPdV0stxX), because the checkpoints the paper cites (Mukhoti et al. 2020) are
not mirrored on the Hub.
Recipe (paper Sec. 4.1): 350 epochs, SGD momentum 0.9, weight decay 5e-4, LR 0.1 -> 0.01 -> 0.001 at 150/100/100 epochs, random crop(32, pad 4) + horizontal flip, batch 128, 45,000 train / 5,000 val / 10,000 test.
| value | |
|---|---|
| test top-1 | 77.68% |
| val top-1 | 77.54% |
| uncalibrated ECE (15 bins) | 10.02% |
| training time | 2 h 9 min, 1x RTX 4000 Ada |
Files: resnet50_cifar100.pt (state dict for official_repo_anon/models/resnet_cifar.py:ResNet50_CIFAR),
logits_val.npz / logits_test.npz (logits + labels), split_indices.npz, train_history.json.
Logbook: https://huggingface.co/spaces/vimarsh/repro-sample-margin-aware-recalibration-of-temperature-scaling
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support