Out-of-Distribution Detection with a Single Unconditional Diffusion Model
Authors: Alvin Heng, alexandre thiery, Harold Soh
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that with a single model, Diff Path is competitive with prior work using individual models on a variety of OOD tasks involving different distributions. |
| Researcher Affiliation | Academia | 1Department of Computer Science, National University of Singapore 2Department of Statistics and Data Science, National University of Singapore 3Smart Systems Institute, National University of Singapore {alvinh, harold}@comp.nus.edu.sg |
| Pseudocode | Yes | We provide pseudocode in Algorithm 1. Algorithm 1 OOD detection with Diff Path |
| Open Source Code | Yes | Our code is publicly available at https://github.com/clear-nus/diffpath. |
| Open Datasets | Yes | All experiments are conducted as of pairwise OOD detection tasks using CIFAR10 (C10), SVHN, and Celeb A as inlier datasets, and CIFAR100 (C100) and Textures as additional outlier datasets. |
| Dataset Splits | No | The paper refers to an "ID train set Xtrain" in Algorithm 1, and discusses fitting models to this data and then calculating OOD scores for test samples. However, it does not explicitly provide details about a specific "validation dataset split" (e.g., percentages or counts) distinct from the training and test sets for model selection or hyperparameter tuning. |
| Hardware Specification | Yes | On a single Nvidia A5000 GPU, Diff Path takes approximately 0.25s and 0.94s per integration step on 32 32 and 64 64 images respectively with a batch size of 256. We train all baselines using 1-3 A5000 GPUs. |
| Software Dependencies | No | The paper mentions using the "sklearn library" for KDE and GMM fitting and specific GitHub repositories for diffusion baselines, but it does not provide specific version numbers for these software dependencies (e.g., "sklearn 0.24.2"). |
| Experiment Setup | Yes | We train our own Celeb A model and utilize the Image Net checkpoint trained using Improved DDPM s Lhybrid objective (Eq. 16 of Nichol and Dhariwal [40]) from the official repository2. Both models use a cosine noise schedule with a total of 4000 diffusion steps. For Diff Path-1D, we fit a KDE using a Gaussian kernel with a bandwith of 5. For Diff Path-6D, we fit a GMM with hyperparameters obtained by sweeping over a predefined number of mixture components (e.g., 50, 100) and covariance type (e.g., diagonal, full, tied). |