LTS

This repository contains the code for the paper Logit Scaling for OOD detection.

Setup

# create conda env and install dependencies
$ conda env create -f environment.yml
$ conda activate lts

# set environmental variables
$ export DATASETS=<your_path_to_datasets_folder>
$ export MODELS=<your_path_to_checkpoints_folder>

# download datasets and checkpoints
$ bash scripts/download.sh

Please download ImageNet dataset manually to $DATASET dir by following this instructions.

Run

To reproduce results from the paper, run the following commands:

$ python ood_eval.py --config config/cifar10_config.yml --use-gpu --use-tqdm
$ python ood_eval.py --config config/cifar100_config.yml --use-gpu --use-tqdm
$ python ood_eval.py --config config/resnet50_config.yml --use-gpu --use-tqdm

Usage

import torch
import torch.nn as nn

from lts import lts

class Net(nn.Module):
    ...
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.features(x)
        x = self.avgpool(x)
        
        # plug-in LTS to get better OOD detection. Function calculates OOD scaling factor s
        s = lts(x)
        
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x, s

net = Net()
for i, data in enumerate(testloader):
    inputs, labels = data
    logits, s = net(inputs)
    
    # get ood predictions
    ood_prediction = get_score(logits * s)

Citations

If you use our codebase, please cite our work:

@article{djurisic2024logit,
  title={Logit Scaling for Out-of-Distribution Detection},
  author={Djurisic, Andrija and Liu, Rosanne and Nikolic, Mladen},
  journal={arXiv preprint arXiv:2409.01175},
  year={2024}
}