Assaf Ben-Kish, Moran Yanuka, Morris Alper, Raja Giryes, Hadar Averbuch-Elor

Hallucinated details are prevalent in the outputs of modern image captioning models. Prior work has largely focused on detecting or mitigating hallucinations by using closed-vocabulary object lists, which simplify the problem but fail to capture most types of hallucinations that occur in practice. By leveraging recent progress in generative foundation models, we propose a unified framework for quantifying and mitigating open-vocabulary hallucinations.

First, We introduce OpenCHAIR, a benchmark for evaluating open-vocabulary hallucinations which surpasses the existing benchmark CHAIR both in diversity and accuracy:

Additionally, we introduce MOCHa, a reinforcement learning-based approach that adjusts captioning models to output detailed, valid captions while avoiding such hallucinations:

git clone https://github.com/assafbk/mocha_code.git
cd mocha_code

To set up our environment, please run:

conda env create -f environment.yml
conda activate mocha
python -m spacy download en_core_web_sm

To perform evaluation over the OpenCHAIR benchmark:

1. Create a csv file with a single column titled 'generated_caption'. The following rows should contain the model's captions for OpenCHAIR's images. An example csv file can be found in:
   OpenCHAIR/example_gen_file.csv

   Additionally, we provide a script for generating such a file for the MOCHa-optimized BLIP-Base model, by running:

   python OpenCHAIR/generate_captions.py \
       --model-ckpt moranyanuka/blip-image-captioning-base-mocha \
       --prompt "a photography of " \
       --batch-size 100
       --num-beams 5

   Additional information:

   • model-ckpt: The huggingface ckeckpoint of the model to be evaluated. Note that the script currently only supports BLIP-Base and BLIP-Large based models.
   • prompt: The prompt appended for the generation

2. Download the Concreteness Rating Dataset (xlsx format) from here.

3. Run the evaluation script:

   python OpenCHAIR/evaluate.py \
       --concreteness-dataset-path <path-to-concreteness-dataset> \
       --generations-file-path <path-to-generated-captions-file>

More configuration options can be found in OpenCHAIR/evaluate.py

The OpenCHAIR dataset can also be accessed from 🤗 Here, and can be loaded as follows:

from datasets import load_dataset
dataset = load_dataset("moranyanuka/OpenCHAIR")['test']

• selecting a large --batch-size can significantly improve the evaluation script's runtime.

We currently support BLIP-Large on the MS-COCO Dataset (will add support for other models and datasets in the near future).

To run the training script:

python vlm_rlhf.py

The configuration file is vlm_rlhf_config.json. Important configurations:

• reward_model_weights: List of weights for all rewards. First is the NLI weight and the second is the BERTScore weight (equivalent to alpha and 1-alpha in the paper). This field tunes the pareto frontier of the fidelity-adequacy curve. Initialized to [0.5,0.5].
• beta: The weight for the kl-penalty reward. Initialized to 0.06.
• num_of_images_per_batch: Number of images per PPO batch. Initialized to 10.
• num_of_samples_per_image: Number of captions to generate per image. Initialized to 10.
  In a single batch there are <num_of_images_per_batch> x <num_of_samples_per_image> captions.
• model_device, ref_model_device, reward_model_device: Cuda device for each model.

All training metrics, including caption samples (for train and verification images) are displayed in the wandb webpage.

Additional configurations:

• output_dir: Where to save model checkpoints. Initialized to <project_dir>/output.
• cache_dir: Huggingface cache dir for all models. Initialized to <project_dir>/hf_cache.
• activate_logging: Enables wandb logging. Initialized to True.
• sampling_temperature: Sampling temperature for the model. Initialized to 1.2.
• save_steps: Model saving interval. Initialized to 200 (Note: best model is always saved regardless of this value).
• eval_steps: Model evaluation interval. Initialized to 10.
• max_step: Maximal amount of training steps. Initialized to 3000.

Check out vlm_rlhf_config.json for more configurations.

We will publish the checkpoints of additional models in the near future.

• If more than one GPU is available, we recommend setting model_device to the first GPU, and ref_model_device and reward_model_device to the second GPU. (Motivation - the former requires grads hence uses the GPU memory more extensively).
• If CUDA is running out of memory, try reducing the batch size (num_of_images_per_batch or num_of_samples_per_image or both).
• To track the learning progress, keep an eye on the generated captions of the verification images (wandb -> Tables -> runs.summary["validation_data"])
• Additionally, it is helpful to look after validation_reward_mean and kl_dist (wandb -> Charts). The kl_dist should not be too large (in BLIP-Large, empirically, no more than 5). In parallel, we want to see validation_reward_mean increase under the small kl_dist constraint. kl_dist is controlled by beta (decreases when we increase beta).

If you find this useful for your research, please cite the following:

@misc{benkish2024mitigating,
      title={Mitigating Open-Vocabulary Caption Hallucinations}, 
      author={Assaf Ben-Kish and Moran Yanuka and Morris Alper and Raja Giryes and Hadar Averbuch-Elor},
      year={2024},
      eprint={2312.03631},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}