FEVoRI

This is implementation of Face-Enhanced Viper of Role Interactions (FEVoRI) and Context Query Reduction (ConQueR) based on ViperGPT

Please follow the Installation section in ViperGPT step by step

Please download TiM dataset from here and place QA annotation under specific folder path

To reduce the cost on repeatly generating same trope identification code when running on TiM. You can follow the sample code generation configuration in configs/code_generation to generate trope identification functions under different settings first: For example, the below command generate code with FEVoRI+ConQueR method. The config naming convention is {modality}_{face Identification}_{coder}_{ICL Example}

CONFIG_NAMES=code_generation/v+d_df_gpt4_complex python main_batch.py

After code generation, there should be new .csv code file under cached_code\v+d_df_gpt4_complex.

Then you can run the same trope identification function on different movies

CUDA_VISIBLE_DEVICES=0 CONFIG_NAMES=ablations/v+d_df_gpt4_complex python main_batch.py

The result file should be placed under results/, and you can use eval.py to evaluate the performance

python eval.py path/to/result.csv

If you use this code, please consider citing the paper as:

@article{su2024investigating,
  title={Investigating Video Reasoning Capability of Large Language Models with Tropes in Movies},
  author={Su, Hung-Ting and Chao, Chun-Tong and Hsu, Ya-Ching and Lin, Xudong and Niu, Yulei and Lee, Hung-Yi and Hsu, Winston H},
  journal={arXiv preprint arXiv:2406.10923},
  year={2024}
}