Following the VCD, use

conda create -yn imccd python=3.9
conda activate imccd
cd IMCCD
pip install -r requirements.txt

For evaluation, modify and run the bash in experiments/cd_scripts, use the

cd experiments
sh cd_scripts/llava1.5_pope.bash

or directly use the python command:

cd experiments
CUDA_VISIBLE_DEVICES=0 python ./eval/object_hallucination_vqa_llava.py \
--model-path ${model_path} \
--question-file ./data/POPE/${dataset_name}/${dataset_name}_pope_${type}.json \
--image-folder ${image_folder} \
--answers-file ./output/${dataset_name}/${type}/llava15_${dataset_name}_pope_${type}_answers_${mask_mode}_imccd_seed${seed}_${sampling}.jsonl \
--cd_alpha $cd_alpha \
--cd_beta $cd_beta \
--use_cd \
--use_mask \
--mask_mode  $mask_mode \
--noise_step $noise_step \
--seed ${seed} \
--sampling ${sampling}

The usage of some important parameters:

• answers-file: the output answer files
• use_cd ： use contrastive decoding.
• cd_alpha： the value of alpha. Valid if and only if use_cd is valid.
• cd_beta： the value of beta. Valid if and only if use_cd is valid.
• use_mask: use imccd. Valid if and only if use_cd is valid.
• mask_mode: change the variants of our method. Valid if and only if use_cd and use_mask are valid. The variants are as follows:
  • ved .
  • imccd .
• noise_step steps of adding noise in VCD. Valid if and only if use_cd is valid.
• sampling: sample or greedy search.

After obtaining the answer files, the evaluation results are estimated by modifying experiments/test_scripts/test_pope.sh or directly using the following code:

python ./experiments/eval/eval_pope.py \
--gen_files ${answer_file_name} \
--gt_files experiments/data/POPE/${dataset_name}/${dataset_name}_pope_${type}.json

Similarly, the evalutaion on CHAIR and mme dataset can be conducted by the instructions in experiments/cd_scripts nd experiments/test_scripts.

The code is heavily borrowed from VCD and thanks for their contribution.