Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models [Paper]
Diagram of Self-Introspective Decoding.
Abstract: Hallucination remains a significant challenge in Large Vision-Language Models (LVLMs). To alleviate this issue, some methods, known as contrastive decoding, induce hallucinations by manually disturbing the raw vision or instruction inputs and then mitigate them by contrasting the outputs of the original and disturbed LVLMs. However, these holistic input disturbances sometimes induce potential noise and also double the inference cost. To tackle these issues, we propose a simple yet effective method named Self-Introspective Decoding (SID). Our empirical investigations reveal that pre-trained LVLMs can introspectively assess the importance of vision tokens based on preceding vision and text (both instruction and generated) tokens. Leveraging this insight, we develop the Context and Textaware Token Selection (CT2S) strategy, which preserves only the least important vision tokens after the early decoder layers, thereby adaptively amplify vision-and-text association hallucinations during auto-regressive decoding. This strategy ensures that multimodal knowledge absorbed in the early decoder layers induces multimodal contextual rather than aimless hallucinations, and significantly reduces computation burdens. Subsequently, the original token logits subtract the amplified fine-grained hallucinations, effectively alleviating hallucinations without compromising the LVLMs’ general ability. Extensive experiments illustrate that SID generates less-hallucination and higher-quality texts across various metrics, without much additional computation cost.
As we design the LVLMs decoding strategy, it is convenient to use SID by installing our modified transformers package.
conda env create -f environment.yml
conda activate SID
python -m pip install -e transformers
After setup the environment, you can directly use our code base to imply three LVLMs Decoding-based Hallucination Alleviation methods: Vision Contrastive Decoding (VCD), Instruction Contrastive Decoding (ICD), OPERA, and our SID:
python pope_eval.py --pope-type coco_adversarial --model llava-1.5 --use-cd --use-fast-v --sample --sample-greedy #SID_greedy
python pope_eval.py --pope-type coco_adversarial --model llava-1.5 --use-vcd --sample --sample-greedy #VCD_greedy
python pope_eval.py --pope-type coco_adversarial --model llava-1.5 --use-icd --sample --sample-greedy #ICD_greedy
python pope_eval.py --pope-type coco_adversarial --model llava-1.5 --beam 5 #Beam Search
python pope_eval.py --pope-type coco_adversarial --model llava-1.5 --beam 5 --opera #OPERA
The CHAIR metric utilizes the same configuration.
We provide extensive evaluation metrics including GPT-4V eval_utils/gpt4v_eval.py , GPT4 shr_eval.py, POPE pope_eval.py, CHAIR eval_utils/chair_eval.py
The following evaluation requires for MSCOCO 2014 / AOKVQA / GPA / Visual Genome dataset. Please download here dataset/download_cqa.py, dataset/download_ha_dpo.py, dataset/download_visual_genome_v1.2.py and extract it in the data path.
Besides, it needs you to prepare the following checkpoints of 7B base models:
- Download LLaVA-1.5 merged 7B model and specify it at
eval_configs/llava-1.5_eval.yaml. - Download Vicuna 7B v1.1 model and specify it at
minigpt4/configs/models/blip2_instruct_vicuna7b.yaml. - Download Shikra merged 7B model and specify it at
eval_configs/shikra_eval.yaml.
| Argument | Example | Description |
|---|---|---|
--model |
llava-1.5 |
Specify the LVLM model. |
--data-path |
/path/to/dataset |
Path to the dataset file or folder. |
--pope-type |
coco_adversarial |
Type for POPE evaluation. |
--sample |
store_true |
Use the modified decoding strategy. |
--sample-greedy |
store_true |
Use CD with sampling and greedy decoding. |
--beam |
5 |
Beam search number. |
--opera |
store_true |
Use OPERA. |
Some codes are based on the LVLMs codebase of OPERA, VCD, and HA-DPO . Thanks for their excellent works!