FA: Forced Prompt Learning of Vision-Language Models for Out-of-Distribution Detection

The paper has been accepted by ICCV2025
Arxiv: https://arxiv.org/abs/2507.04511

Requirements

Installation

Create a conda environment and install dependencies (you do not need to install the toolbox Dassl.pytorch first):

pip install -r requirements.txt

Get Started

Configs

The running configurations can be modified in configs/my_config.yaml.
The ID dataset configurations can be modified in my_dataset/.

Train

CUDA_VISIBLE_DEVICES=3 python main.py --config configs/my_config.yaml --is_train 1

:blue_book: Models Zoo

• ID: ImageNet-1K, backbone: ViT-B/16

K	Seed	iNaturalist		SUN		Places		Textures		Avg		Model/log
		FPR95	AUROC	FPR95	AUROC	FPR95	AUROC	FPR95	AUROC	FPR95	AUROC
2	1	11.83	97.18	27.03	93.71	30.23	92.45	29.29	92.94	24.60	94.07	here
	3	15.11	96.44	27.72	93.17	31.36	92.07	30.53	92.97	26.18	93.66	here
	5	12.53	96.92	26.68	93.67	31.19	92.16	29.82	92.53	25.06	93.82	here
3	5	15.66	96.23	26.57	93.66	29.51	92.82	28.48	93.14	25.05	93.96	here
4	5	16.44	96.30	27.96	93.54	30.99	92.60	26.37	93.95	25.44	94.10	here

Inference

CUDA_VISIBLE_DEVICES=2 python main.py --config configs/my_config.yaml --is_train 0

The specific paths for different OOD datasets can be modified in the inference section of the main.py code.