jpcbertoldo/fitymi

Fake It Till You Make It: Near-Distribution Novelty Detection by Score-Based Generative Models

Fake It Till You Make It: Near-Distribution Novelty Detection by Score-Based Generative Models

Official PyTorch implementation of "Fake It Till You Make It: Near-Distribution Novelty Detection by Score-Based Generative Models" by Hossein Mirzaei, Mohammadreza Salehi, Sajjad Shahabi, Efstratios Gavves, Cees G. M. Snoek, Mohammad Sabokrou, Mohammad Hossein Rohban

1. Requirements

Environment

The current version requires the following python and CUDA versions:

• python 3.7+
• CUDA 11.1+

Additionally, the list of the packages used for this implementation is available in the requirements.txt file. To install them, use the following command:

pip install -r requirements.txt

Datasets

To replicate the results of the experiments, please download the following generated anomaly datasets:

• CIFAR-10 Generated Dataset
• CIFAR-100 Generated Dataset

Each of these datasets is used as counterfeit anomalies during the model's training. Each dataset contains generated samples for every class of the dataset. These samples are concatenated together with respect to their class number. E.g., the first 5000 images of the cifar10_training_gen_data.npy are generated samples based on the first class of the CIFAR-10 dataset (Airplane), the second 5000 images are based on the second label, and so on. (Note that for the CIFAR-100 dataset number of samples for each class is 2500).

Currently, only the generated samples for these two datasets are available. For other datasets, please refer to this implementation of an SDE-based generative model, and follow the guidelines provided in the paper to generate anomaly samples.

2. Training and Evaluation

One-Class Novelty Detectiion

python main.py --dataset <DATASET> --label <NORMAL_CLASS> --output_dir <RESULTS_DIR> --normal_data_path <NORMAL_DATA_DIR>\
    --gen_data_path <GEN_DATA_DIR> --pretrained_path <MODEL_DIR> --train_batch_size 16 --eval_batch_size 16 --nnd --download_dataset

The option --label indicates the normal class. Use the --gen_data_path option to set the path to generated datasets provided in the Datasets section. The --pretrained_path option specifies the path to the pre-trained model. Please refer to this implementation of the ViT model to see the list of the available models and how to access them. Finally, the --nnd option should be used to evaluate the model on the NND setting described in the paper. This option is only available for the CIFAR-10 dataset.

An example of training and evaluation of the model on the first class of both datasets is available in this notebook.

For high-resolution datasets, it is recommended to increase the learning rate to 5e-3 and the epoch count to 150.

3. Results

The model's performance on the available datasets is provided in the table below:

	CIFAR-10 (ND)	CIFAR-10 (NND)	CIFAR-100
AUROC	99.1	90.0	98.1

To see the results on other datasets, please refer to our paper.

4. CIFAR-10-FSDE

In this section, the dataset for the CIFAR-10-FSDE benchmark is provided. This dataset is a subsample of the generated data on the CIFAR-10 dataset. It can be used as a measure to evaluate anomaly detection and out-of-distribution methods in the near-distribution setting.

To download the dataset, please use the link provided below:

• CIFAR-10-FSDE

This dataset contains the test samples for every class of the CIFAR-10 dataset (1000 samples per class). These samples are concatenated together in the manner described in the Datasets section.

To see the performance of various anomaly detection and out-of-distribution methods, please refer to our paper.

5. Citation

If you find this useful for your research, please cite the following paper:

@article{mirzaei2022fitymi,
  title={Fake It Till You Make It: Near-Distribution Novelty Detection by Score-Based Generative Models},
  author={Hossein Mirzaei and Mohammadreza Salehi and Sajjad Shahabi and Efstratios Gavves and Cees G. M. Snoek and Mohammad Sabokrou and Mohammad Hossein Rohban},
  year={2022},
  eprint={2205.14297},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}