This repository hosts the pretrained weights released for our NeurIPS 2026 submission “DPLC: Dirichlet Process Guided Long-tail Clustering”.
DPLC is a Dirichlet Process guided long-tailed clustering method that adaptively discovers the latent sub-cluster structure of imbalanced data, providing a non-parametric signal for dynamic class reweighting and similarity regularization to achieve balanced clustering under severe imbalance ratio.
What is inside We are gradually releasing pretrained backbones. So far, the following weight is available:
| File | Backbone | Pre-training | Dataset | Notes |
|---|---|---|---|---|
byol_cifar10_1000ep.pth |
BYOL | 1000 epochs | CIFAR-10 | Encoder checkpoint to skip Step-1 pre-training and reproduce DPLC results directly. |
Weights for CIFAR-20 and STL-10 will be updated as time permits.
Intended use These checkpoints are meant to be loaded as the encoder / backbone in the DPLC pipeline, so you can skip the ~6-hour BYOL pre-training (Step 1) and go straight to DPLC training and evaluation (Step 2). For reproduction, you can directly pass the checkpoint path to the checkpoint argument for inference.
For full training/eval code, dataset preparation and hyper-parameters, please refer to the GitHub repository[Github]https://github.com/V2tPkDuck/DPLC.