DualPrompt

Official code for the ICLR 2026 paper Unlocking the Power of Co-Occurrence in CLIP: A DualPrompt-Driven Method for Training-Free Zero-Shot Multi-Label Classification.

Overview

DualPrompt is a training-free zero-shot multi-label classification method built on CLIP. It improves label prediction by injecting class co-occurrence information into text prompts through a dual-prompt design:

• Discriminative prompts for class-specific recognition
• Correlation-aware prompts for modeling label co-occurrence

Features

• Training-free zero-shot multi-label classification with CLIP
• Dual-prompt design: discriminative prompt + correlation-aware prompt
• Multiple co-occurrence sources:
  • data: estimated from a subset of training labels
  • chatgpt: external semantic co-occurrence priors stored in JSON files
  • prior: full training-set co-occurrence prior
• Metrics include mAP, CP, CR, CF1, OP, OR, OF1

Installation

Create a Python environment and install dependencies:

pip install -r requirements.txt

Quick Start

Example on COCO2014 with CLIP ViT-B/16:

python run.py --data_name coco2014 --data_dir ./data --model_name ViT-B/16 --coo_derive data --sample_ratio 0.01 --coo_mode both --coo_max_n 5 --lam 0.5 \

Main Arguments

• --data_name: dataset name, one of coco2014, vg256, objects365
• --data_dir: root path containing dataset folders
• --model_name: CLIP backbone
• --output: output root directory
• --batch_size: batch size for evaluation
• --workers: number of dataloader workers
• --lam: weight of the correlation-aware branch in the final dual-logit fusion
• --coo_derive: co-occurrence source, one of data, chatgpt, prior
• --coo_mode: co-occurrence construction mode, one of top, bottom, both
• --coo_thre: threshold for keeping co-occurring classes
• --coo_max_n: maximum number of co-occurring classes per target class
• --sample_ratio: ratio of training labels used to estimate co-occurrence
• --seed: random seed

Outputs

Results are saved under the directory specified by --output. The final release version keeps only the dual-logit outputs and summary metrics.

Typical files include:

• config.json: full runtime configuration
• log.txt: evaluation log
• coo_dict.json: generated co-occurrence dictionary when --coo_derive data
• dual_logit.npy: final logits of the dual-prompt method
• dual_prob.npy: final normalized prediction scores used for evaluation
• APs.npy: per-class average precision values
• summary_metrics.json: summary metrics including mAP, CP, CR, CF1, OP, OR, and OF1
• mAP.npy, CF1.npy, OF1.npy: key scalar metrics saved separately

Citation

@inproceedings{
xie2026unlocking,
title={Unlocking the Power of Co-Occurrence in CLIP: A DualPrompt-Driven Method for Training-Free Zero-Shot Multi-Label Classification},
author={Ming-Kun Xie and Zhiqiang Kou and Zhongnian Li and Gang Niu and Masashi Sugiyama},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=QGXVZ0OPLy}
}