ODIN - CIFAR100

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 CIFAR100_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}

model = WideResNet(num_classes=100, pretrained="cifar100-pt").eval().to(device)
trans = WideResNet.transform_for("cifar100-pt")
norm_std = WideResNet.norm_std_for("cifar100-pt")

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

results = []
benchmark = CIFAR100_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	86.32	84.81	88.23	43.35
TinyImageNetResize	MSP	74.64	70.91	77.29	65.52
LSUNResize	MSP	75.38	71.16	78.50	63.36
LSUNCrop	MSP	85.59	84.36	87.40	47.13
Uniform	MSP	77.92	16.86	97.60	40.44
Gaussian	MSP	84.78	23.12	98.41	30.22
TinyImageNetCrop	ODIN	86.89	84.01	89.02	40.54
TinyImageNetResize	ODIN	80.79	78.44	82.08	60.10
LSUNResize	ODIN	81.25	78.04	83.04	58.13
LSUNCrop	ODIN	86.91	85.69	88.79	42.73
Uniform	ODIN	95.52	59.22	99.55	15.13
Gaussian	ODIN	98.57	85.14	99.86	5.76