ODIN - CIFAR10

Reproduces the ODIN benchmark for OOD detection, from the paper Enhancing the reliability of out-of-distribution image detection in neural networks.

import pandas as pd  # additional dependency, used here for convenience
import torch

from pytorch_ood.benchmark import CIFAR10_ODIN
from pytorch_ood.detector import ODIN, MaxSoftmax
from pytorch_ood.model import WideResNet
from pytorch_ood.utils import fix_random_seed

fix_random_seed(123)

device = "cuda:0"
loader_kwargs = {"batch_size": 64, "num_workers": 12}

model = WideResNet(num_classes=10, pretrained="cifar10-pt").eval().to(device)
trans = WideResNet.transform_for("cifar10-pt")
norm_std = WideResNet.norm_std_for("cifar10-pt")

detectors = {
    "MSP": MaxSoftmax(model),
    "ODIN": ODIN(model, eps=0.002, norm_std=norm_std),
}

results = []
benchmark = CIFAR10_ODIN(root="data", transform=trans)

with torch.no_grad():
    for detector_name, detector in detectors.items():
        print(f"> Evaluating {detector_name}")
        res = benchmark.evaluate(detector, loader_kwargs=loader_kwargs, device=device)
        for r in res:
            r.update({"Detector": detector_name})
        results += res

df = pd.DataFrame(results)
print((df.set_index(["Dataset", "Detector"]) * 100).to_csv(float_format="%.2f"))

This produces a table with the following output:

Dataset	Detector	AUROC	AUPR-IN	AUPR-OUT	FPR95TPR
TinyImageNetCrop	MSP	94.59	93.10	95.77	17.18
TinyImageNetResize	MSP	88.22	86.01	89.25	42.56
LSUNResize	MSP	91.45	89.46	92.65	29.06
LSUNCrop	MSP	96.49	95.69	97.20	12.49
Uniform	MSP	87.22	29.65	98.60	32.50
Gaussian	MSP	90.00	34.68	98.96	24.75
TinyImageNetCrop	ODIN	96.78	96.46	97.10	14.17
TinyImageNetResize	ODIN	91.44	91.31	91.45	38.83
LSUNResize	ODIN	94.66	94.48	94.80	26.25
LSUNCrop	ODIN	98.10	98.11	98.16	9.37
Uniform	ODIN	95.94	73.27	99.57	17.16
Gaussian	ODIN	97.02	80.50	99.67	12.89