Awesome

Data Preparation:

Prepare the data as follows.

TextVID: download TextVID at TextVid only. download TextVID and preprocessed features at TextVid and features

NeXTQA, STAR and TVQA: The prepocessed feautures are available at here.

EgoScehma: Download raw videos from EgoSchema. We provide prepocessed feature here

MVBench: Download raw videos from Hugging Face.

MSRVTT: Download raw videos from MSRVTT.

./data
   |─ nextqa
   |   |─ train.csv
   |   |─ val.csv
   |   └─ clipvitl14.pth
   |─ star
   |   :
   |─ tvqa
   |   :
   └─ egos
       :

Model Preparation:

Prepare the model as follows.

LLMs: Download the pretrained Llama models from Llama2 and Llama3.

TOPA Checkpoints: Download our pretrained models

./pretrained
   └─ llama2
   |    |─ 7B
   |    |   |─ consolidated.00.pth
   |    |   └─ params.json
   |    |─ 13B
   |    |   :
   |    |   :
   |    └─ tokenizer.model
   └─ llama3
        |─ 8B
        |   |─ consolidated.00.pth
        |   └─ params.json
        └─ tokenizer.model

./vqa_checkpoint
   └─ checkpoint_pretrain
        |─ llama2_7b
        |─ llama2_13b
        └─ llama3_8b

Training & Evaluation

Text-only Pre-alignment

./scripts/pretrain/llama2_7b.sh

Zero-shot inference

./scripts/eval/zeroshot_eval_egos.sh
./scripts/eval/zeroshot_eval_nextqa.sh
./scripts/eval/zeroshot_eval_star.sh
./scripts/eval/zeroshot_eval_tvqa.sh

Evaluate on MVBench

mvbench.ipynb

Evaluate on video captioning benchmarks

MSRVTT.ipynb

VATEX.ipynb

Acknowledgements

This repo is built upon Flipped-VQA and benefits from LLaMA-Adapter, DeCap, MVBench, Llama2 and Llama3.

Citations

@article{li2024topa,
  title={TOPA: Extend Large Language Models for Video Understanding via Text-Only Pre-Alignment},
  author={Li, Wei and Fan, Hehe and Wong, Yongkang and Kankanhalli, Mohan and Yang, Yi},
  journal={arXiv preprint arXiv:2405.13911},
  year={2024}
}