Debiasing Large Visual Language Models / Debiasing Multimodal Large Language Models

This is the official repo for Debiasing Large Visual Language Models, including a Post-Hoc debias method and Visual Debias Decoding strategy. These strategies not only prove beneficial in minimizing hallucinations but also contribute to the generation of more helpful and precise illustrations

🔥 Update

• [2024-03-08]: ⭐️ Paper online. Check out Debiasing Multimodal Large Language Models for details.
• [2024-03-11]: 🚀🚀 Codes released.

🎯 Overview

• Our investigation reveals a noteworthy bias in the generated content, where the output is primarily influenced by the underlying Language Models (LLMs) rather than the input image.

• We introduce Post-Hoc debias, where a "calibration" step is implemented for the model's output probabilities using an affine transformation.

• We introduce Visual Debias Decoding (VDD), a simple and training-free method that contrasts output distributions derived from original and image-free visual inputs.

• These strategies not only prove beneficial in minimizing hallucinations but also contribute to the generation of more helpful and precise illustrations.

🕹️ Usage

Environment Setup

conda create -yn vdd python=3.9
conda activate vdd
cd LLaVA-Align
pip install -r requirements.txt

Re-implementation of Our Results

For all experiments presented in our paper, refer to the experiments/scripts directory for detailed commands and scripts. Below, we provide simple implementation examples and guidance.

How to Use Post-Hoc Debiasing in LVLMs

To implement Post-Hoc Debiasing in LVLMs, follow these steps:

1. Obtain Output Distributions:

   • Generate output distributions with naive image and meaningless vision information, such as pure text input (None) or by replacing vision tokens with </unk>.
   • For the POPE benchmark, use the following code to calculate the top-k tokens with their probabilities for each input:

     ./eval/calibrate/llava_calibrate.py

2. Initialize Debiasing Weight:

   • With the obtained naive classification distribution and debiased classification distribution, initialize the debiasing weight $W$ and bias $b$.
   • Adjust the output distribution using affine transformation.

3. Concrete Example - POPE Binary Classification:

   • For the POPE binary classification setting, run the following code to see a concrete example:

     ./eval/eval_pope_calibrate.py

Feel free to modify the parameters or refer to the code for additional details on the implementation of Post-Hoc Debiasing in LVLMs.

How to Use VDD in LVLMs

To help you get started quickly, here's an example using LLaVA on how to replace the conventional sampling method with the VDD/VCD method during generation:

1. Add the following at the beginning of the start-up script:

from vcd_utils.vcd_sample import evolve_vcd_sampling
evolve_vcd_sampling()

The evolve_vcd_sampling function replaces the sampling function in the transformers library. The modified sampling function includes an option for visual contrastive decoding, while keeping the rest unchanged.

2. Slightly modify llava_llama.py:

   a. Add contrastive decoding parameters in the LlavaLlamaForCausalLM class's forward function to avoid exceptions in model.generate.

   b. Add the prepare_inputs_for_generation_cd function.

3. set the hyperparameter in the generate function:

output_ids = model.generate(
    input_ids,
    images=image_tensor.unsqueeze(0).half().cuda(),
    use_dd=args.use_dd,
    use_dd_unk=args.use_dd_unk,
    cd_alpha = args.cd_alpha,
    cd_beta = args.cd_beta,
    do_sample=True)

--use_dd: use pure text input for debias decoding
--use_dd_unk: replace the image tokens with </unk> token, the output logits will be used for debias decoding
--use_dd --use_dd_unk: use both of them

How to Test the Effects of Decoding Configurations on LVLMs

To assess the impact of decoding configurations on LVLMs, follow these steps:

1. Implement For-Loops:

   • Implement for-loops on the temperature, top-p, and top-k configurations in the decoding process.
   • Collect the results obtained for each configuration.

2. Evaluate Results:

   • Evaluate the collected results using a similar approach as described in the preceding sections.

3. Concrete Example:

   • For a concrete example for POPE or LLaVA-Bench, run the following code to see how the implementation works:

     ./eval/sampling/llava_sampling.py

📌 Examples

Figure 13. Qualitative examples showcasing the impact of VDD on LLaVA-v1.5-13B. VDD is demonstrated to be less hallucinated.

Figure 14. Another Qualitative examples showcasing the impact of VDD on LLaVA-v1.5-13B. VDD is demonstrated to be more helpful and precies.

📑 Citation

If you find our project useful, we hope you can star our repo and cite our paper as follows:

@misc{zhang2024debiasing,
      title={Debiasing Large Visual Language Models}, 
      author={Yi-Fan Zhang and Weichen Yu and Qingsong Wen and Xue Wang and Zhang Zhang and Liang Wang and Rong Jin and Tieniu Tan},
      year={2024},
      eprint={2403.05262},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

📝 Related Projects

• Contrastive Decoding: Open-ended Text Generation as Optimization
• Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond
• LLaVA 1.5: Improved Baselines with Visual Instruction Tuning
• VCD: Visual Contrastive Decoding