(2024CVPR) MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding

Project Page | Paper

The official repository of our paper "MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding".

Model Overview

Demo

You can explore our demo by running demo.ipynb. This demonstration illustrates how our MA-LMM serves as a plug-and-play module that can be integrated into InstructBLIP seamlessly, requiring no fine-tuning for zero-shot evaluation.

Requirements

You can install the conda environment by running:

git clone https://github.com/boheumd/MA-LMM.git
cd MA-LMM
pip install -e .

If you are running the code on Apple Silicon, you need to use eva-decord instead of decord. Here is the modification in the requirements.txt file you should do:

contexttimer
eva-decord
einops>=0.4.1
fairscale==0.4.4
...

Before running pip install -e ., ensure you have the correct requirements.

Dataset

For the long-term video understanding task, we conduct experiments including (LVU) and two standard video summarization datasets (Breakfast, COIN).

For the video question answering task, we conduct experiments including MSRVTT, MSVD, and ActivityNet. For the video captioning task, we also conduct experiments on Youcook2 dataset.

You can download videos for each dataset through the script provided here (lavis/datasets/download_scripts). For LVU/Breakfast/COIN datasets, please download the original videos through the official link provided above.

Then extract video frames of each video with fps=10. Example preprocess code is provided here extract_frames.py. Since different FFMPEG versions are used, the actual extracted frame lengths can be slightly inconsistent. You may need to update the actual frame_length for each video in the annotation file.

 ├── data
     └── activitynet
         ├── annotation
         ├── frames
         ├── videos
     └── breakfast
         ├── annotation
         ├── frames
         ├── videos
     └── coin
         ├── annotation
         ├── frames
         ├── videos
     └── lvu
         ├── annotation
         ├── frames
         ├── videos
     └── msrvtt
         ├── annotation
         ├── frames
         ├── videos
     └── msvd
         ├── annotation
         ├── frames
         ├── videos
     └── youcook2
         ├── annotation
         ├── frames
         ├── videos

Running

Download Pre-trained LLM

We use Vicuna-v1.1 as our pre-trained LLM weights, you can download from this link as arrange in this format.

├── llm
     ├── vicuna-7b
     ├── vicuna-13b

Finetuning on Downstreaming Tasks

Our model leverages pre-trained weights from InstructBlip, which was only pre-trained on image-text pairs. Our training process occurred on four A100 GPUs. If you would like to fine-tune the model for various video datasets, please run the following command:

bash run_scripts/${dataset}/train.sh

LVU dataset

    # Please choose the task from the following list
    # ['director', 'genre', 'relationship', 'scene', 'way_speaking', 'writer', 'year']
    datasets.lvu_cls.task ${task}

Testing

We also provided finetuned checkpoints for each video dataset. Please download the saved_model.tar and unzip it. For the test script corresponding to each dataset, provide the path to the extracted checkpoint to execute the evaluation.

bash run_scripts/${dataset}/test.sh ${checkpoint_path}

Zero-shot Evaluation

Our model can also leverage pre-trained weights from InstructBlip without any finetuning to conduct zero-shot evaluation on video datasets.

bash run_scripts/${dataset}/test.sh

Hyper-parameters

One important hyper-parameters memory_bank_length, please change that in the training script on different datasets.

    # pre-defined length of the memory bank
    model.memory_bank_length ${value}
    # value=0 means without using the memory bank

Memory Bank Compression Code

The core algorithm for the memory bank compression algorithm is here.

Citation

If you find our code or our paper useful for your research, please [★star] this repo and [cite] the following paper:

@inproceedings{he2024malmm,
  title = {MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding},
  author    = {He, Bo and Li, Hengduo and Jang, Young Kyun and Jia, Menglin and Cao, Xuefei and Shah, Ashish and Shrivastava, Abhinav and Lim, Ser-Nam},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2024}
}

Acknowledgement

We referenced the repo below for the code

• LAVIS