---
language:
- en
license: mit
library_name: pytorch
pipeline_tag: image-classification
tags:
- clip
- prompt-learning
- few-shot-learning
- image-classification
- specpl
---

# SpecPL Prompt-Learning Checkpoints (Base-to-Novel)

## Model Details

This repository provides released checkpoints for:

- **SpecPL: Disentangling Spectral Granularity for Prompt Learning** (ICML 2026)
- Paper: https://arxiv.org/abs/2605.04504
- Code: [Mlrac1e/SpecPL-Prompt-Learning](https://github.com/Mlrac1e/SpecPL-Prompt-Learning)

Included trainer families:

- `CoOpSpecPL`
- `MaPLeSpecPL`
- `MMRLSpecPL`

## Repository Contents

- 33 checkpoints total (11 datasets x 3 trainer families)
- `release_index.csv` with one row per checkpoint, including checkpoint path/hash/size, Base-to-Novel metrics (B/N/HM), and extracted experiment configuration fields
- checkpoint files are organized under canonical paths: `checkpoints/<trainer>/<dataset>/shots_<k>/<cfg>/model.pth.tar`

This release intentionally excludes training/test logs and internal manifest files.

## Intended Use

These checkpoints are intended for:

- research reproducibility
- base-to-novel prompt-learning comparison
- checkpoint reuse with the original SpecPL training/evaluation codebase

## Quick Usage (Official Repo)

Use these checkpoints with the official codebase.

```bash
git clone https://github.com/Mlrac1e/SpecPL-Prompt-Learning.git
cd SpecPL-Prompt-Learning

# Set dataset/cache paths
export DATA_ROOT=path/to/data
export CLIP_ROOT=path/to/clip

# Checkpoint files in this release
ls /path/to/Output_Release_HF/checkpoints
```

Select the checkpoint path from `release_index.csv`, place it at the output location expected by the official scripts, and run the corresponding Base-to-Novel evaluation script from the repository documentation.

## Training And Evaluation Context

- Protocol: base-to-novel generalization
- Shot setting: 16-shot
- Datasets: 11 benchmarks used in the paper/repo

## Limitations

- These are raw training checkpoints, not end-user inference packages.
- Results depend on the original environment/configuration and evaluation scripts.
- Dataset licenses and access conditions follow each dataset's original terms.

## Citation

```bibtex
@inproceedings{zhou2026specpl,
    title     = {SpecPL: Disentangling Spectral Granularity for Prompt Learning},
    author    = {Zhou, Jingtao and Kang, Xirui and Huang, Feiyang and Po, Lai-Man},
    booktitle = {Proceedings of the International Conference on Machine Learning (ICML)},
    year      = {2026}
}

@misc{zhou2026specpldisentanglingspectralgranularity,
    title         = {SpecPL: Disentangling Spectral Granularity for Prompt Learning},
    author        = {Jingtao Zhou and Xirui Kang and Feiyang Huang and Lai-Man Po},
    year          = {2026},
    eprint        = {2605.04504},
    archivePrefix = {arXiv},
    primaryClass  = {cs.CV},
    url           = {https://arxiv.org/abs/2605.04504}
}
```