lightning-uq-box/epinet_cifar10_logits

Cached CIFAR-10 test logits for the epinet tutorial

Test-set logits for six uncertainty-quantification agents on CIFAR-10, cached so that the joint-prediction figures in the lightning-uq-box epinet tutorial can be rebuilt on CPU in seconds, without a GPU and without retraining 32 ResNets.

Consumed by docs/tutorials/classification/epinet.ipynb.

Contents

File	Description
test_logits.pt	torch.save dict of logit tensors (below)
metrics.csv	Scalar metrics from the same run (run.py evaluate)
flop_counts.json	Measured inference FLOPs per image, per agent
parameter_counts.json	Parameter counts for the base net and the epinet head
run_metadata_cache_logits_0.json	Torch/CUDA/GPU versions and loader config

test_logits.pt

Load with torch.load(path, weights_only=True). Keys:

Key	Shape	dtype	Meaning
labels	(10000,)	int64	CIFAR-10 test labels
ensemble	(30, 10000, 10)	float16	30 independently trained ResNet-18 members
epinet	(50, 10000, 10)	float16	50 epistemic-index samples, one frozen base net
dropout	(50, 10000, 10)	float16	50 MC-dropout draws, one base net

Layout is [num_samples, num_test_inputs, num_classes]. Cast to float32 before computing log-losses. Ensembles of size k are the first k rows of ensemble, so the members are nested rather than independently resampled.

Stored as float16 to keep the file at 26 MB; the logits are only ever consumed through a softmax, where the precision loss is immaterial.

How this was generated

A ResNet-18 base network trained on CIFAR-10 for 50 epochs (batch size 128, 50000/10000 split, cosine schedule), then an epinet head fitted for 10 epochs on top of the frozen base, plus 30 ensemble members and an MC-dropout baseline. From the lightning-uq-box repo:

python experiments/epinet_cifar10/cifar/run.py base         --members 30 --epochs 50 --out results/epinet_cifar_full
python experiments/epinet_cifar10/cifar/run.py dropout      --epochs 50 --out results/epinet_cifar_full
python experiments/epinet_cifar10/cifar/run.py epinet       --epochs 10 --out results/epinet_cifar_full
python experiments/epinet_cifar10/cifar/run.py evaluate     --members 30 --out results/epinet_cifar_full
python experiments/epinet_cifar10/cifar/run.py cache_logits --members 30 --out results/epinet_cifar_full

Produced on a single NVIDIA A100-SXM4-40GB, torch 2.13.0+cu130, CUDA 13.0. No test-driven checkpoint selection was used.

Headline numbers

Classification error and log-losses on the 10000-image test set (from metrics.csv, dyadic joint log-loss at tau=10):

Agent	Test error	Marginal log-loss	Joint log-loss
single net	0.0577	0.2119	0.2579
MC dropout	0.0566	0.2127	0.2012
epinet	0.0582	0.2123	0.1623
ensemble (3)	0.0479	0.1562	0.1065
ensemble (10)	0.0443	0.1361	0.0884
ensemble (30)	0.0424	0.1308	0.0714

The epinet matches a single network on marginal log-loss while approaching a 3-member ensemble on the joint metric, at 1.01x the inference FLOPs.

License

Apache-2.0, matching lightning-uq-box. Derived from CIFAR-10 (Krizhevsky, 2009); the underlying images are not redistributed here, only model outputs on them.