base on Official repository for the paper PLLaVA <div align="center"> <h2><a href="https://pllava.github.io/">PLLaVA: Parameter-free LLaVA Extension from Images to Videos for Video Dense Captioning</a></h2> [Lin Xu](https://scholar.google.com/citations?user=_Gu69coAAAAJ), [Yilin Zhao](https://ermu2001.github.io/me.io/), [Daquan Zhou](https://scholar.google.com/citations?user=DdCAbWwAAAAJ), [Zhijie Lin](https://scholar.google.com/citations?user=xXMj6_EAAAAJ), [See-Kiong Ng](https://scholar.google.com/citations?user=_wsommYAAAAJ), [Jiashi Feng](https://scholar.google.com.sg/citations?user=Q8iay0gAAAAJ&hl=en) </div> <!-- [](https://arxiv.org/abs/2311.17005) --> **Project Page: [PLLaVA](https://pllava.github.io/)** [](https://arxiv.org/abs/2404.16994) [](https://www.youtube.com/watch?v=nAEje8tu18U) [](https://huggingface.co/ermu2001/pllava-34b) [](https://paperswithcode.com/sota/zeroshot-video-question-answer-on-activitynet?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/zeroshot-video-question-answer-on-msrvtt-qa?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/zeroshot-video-question-answer-on-msvd-qa?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/video-question-answering-on-mvbench?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/zeroshot-video-question-answer-on-tgif-qa?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/video-based-generative-performance-4?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/video-based-generative-performance-3?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/video-based-generative-performance?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/video-based-generative-performance-2?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/video-based-generative-performance-1?p=pllava-parameter-free-llava-extension-from-1) [](https://paperswithcode.com/sota/video-based-generative-performance-5?p=pllava-parameter-free-llava-extension-from-1) ![]() <div align="center"> <a href="https://pllava.github.io"> <img src="assert/logo.png"> </a> </div> <div align="center"> <video src="https://github.com/magic-research/PLLaVA/assets/55656210/a6619702-12d3-489d-bfcc-0ef7105544b2" width="100%"> </div> ## Overview Welcome to PLLAVA! The primary purpose of this repository is to support research and the development of prototype models. It is designed to facilitate ease of experimentation and enable a clear overview of results. Please note that this repo is currently undergoing development and reconstruction. It's important to mention that we have not optimized the response speed of the application or the frontend logic. Our goal is to maintain simplicity, clarity, and ease of development, making it accessible for both researchers and students. If you have suggestions or want to enhance the application's performance, please feel free to contact us or contribute to the project. We've briefly introduce our work in section [PLLAVA](#%EF%B8%8F-pllava). For more details, feel free to read our paper. Check out section [Usage](#hammer-usage) to start using this repo. If you felt our works interesting, please star us, your support is all we want. If you find our work helpful, feel free to [cite](#page_facing_up-citation) us directly. ## :fire: Updates - **2024/4/24**: Release: - We are releasing our code/models/datasets. ## 🏖️ PLLAVA <div align="center"> <a href="https://www.youtube.com/embed/nAEje8tu18U?si=GXxjgP93j77FzDbw"> <img src="assert/teaser.jpg"> </a> </div> ### Abstract Vision-language pre-training (VLP) has significantly elevated performance across a range of vision-language applications. Yet, the pre-training process for video-related tasks demands an exceptionally high degree of computational and data resources. This paper investigates a straightforward, highly efficient, and resource-light approach to adapting an existing image-language pre-training model for video data. Our preliminary experiments reveal that directly fine-tuning pre-trained image-language models with multiple frames on video datasets leads to performance saturation or even a drop in caption-related tasks. Besides, it is also vulnerable to prompts and tends to provide short descriptions. We conducted a deep analysis and observed that the performance saturation and the vulnerability might be related to the dominant patches that exist in some single video patches. We then propose a simple pooling strategy to smooth the feature distribution along the temporal dimension and thus reduce the dominant impacts from some extreme tokens. The new model is termed Pooling LLaVA, or PLLaVA in short. With the proposed pooling strategy, we achieve new state-of-the-art performance on all evaluated datasets. Notably, on the recent popular Video ChatGPT benchmark, PLLaVA achieves a score of 3.48 out of 5 on average of five evaluated dimensions, which is the new state-of-the-art score on the leaderboard and is 0.31 higher than the previous SOTA results from GPT4V (IG-VLM). On the latest multi-choice benchmark MVBench, PLLaVA achieves 58.1% accuracy on average across 20 sub-tasks, which is the new state-of-the-art result and is 14.5% higher than GPT4V (IG-VLM). <div align="center"><img src="assert/module.png"></div> ### SEARCHING FOR OPTIMAL POOLING STRATEGY There are two dimensions for the pooling strategy: the spatial dimension and the temporal dimension. We empirically found that reducing the spatial dimension with a larger temporal dimension could lead to better model performance, compared to reducing the temporal dimension directly. <div align="center"><img src="assert/zeroshot.png"></div> ### STATE-OF-THE-ART PERFORMANCE We compare the performance of PLLAVA with recent popular methods over both question-answer and captioning datasets. The results are shown below. <div align="center"><img src="assert/performance.png"></div> ## :hammer: Usage This section provides guidance on how to run, train, and evaluate our models. ### Install First, you will need to set up the environment and download some pre-trained weights. This repo is built up using [transformers](https://github.com/huggingface/transformers) for model construction along with [accelerate](https://github.com/huggingface/accelerate) for distributed training. Follow the instructions to install the needed environment. 1. Above all, the following environment set up is for python 3.10. If you choose to use conda for environment setup, we recommend creating the virtual environment with: ```bash conda create -n pllava python=3.10 ``` 1. Firstly, install [pytorch](https://pytorch.org/) from the official website. The code runs on torch 2.2.1, cu118 or cu122. Select the version that suits your drive version. ``` torch 2.2.1+cu118 torchaudio 2.2.1+cu118 torchvision 0.17.1+cu118 ``` If your driver version is higher than cu121, you could probably try installing with the following scripts: ```bash pip install -r requirements.txt ``` Otherwise, you would need to install a torch for your server first, then install the other packages: ```bash pip install -r requirements.torch.txt # decide your own requirements, (this is for cu11), or install torch directly following the official website. pip install -r requirements.no_torch.txt # install the following ``` 1. Prepare the model. We prefer to have huggingface models explicitly downloaded to a MODELS directory. However, if you are familiar with huggingface-hub usage, feel free to organize the model yourself. ``` python python_scripts/hf.py ``` Here are some detailed information of the obtained models: | Model | Link | Initialized From | | ------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------- | | pllava-7b | [](https://huggingface.co/ermu2001/pllava-7b) | [llava-hf/llava-v1.6-vicuna-7b-hf · Hugging Face](https://huggingface.co/llava-hf/llava-v1.6-vicuna-7b-hf) | | pllava-13b | [](https://huggingface.co/ermu2001/pllava-13b) | [llava-hf/llava-v1.6-vicuna-13b-hf · Hugging Face](https://huggingface.co/llava-hf/llava-v1.6-vicuna-13b-hf) | | pllava-34b | [](https://huggingface.co/ermu2001/pllava-34b) | [llava-hf/llava-v1.6-34b-hf · Hugging Face](https://huggingface.co/llava-hf/llava-v1.6-34b-hf) | The model directory should look like this, where you would only need the corresponding model's weights and directory. ``` $ tree MODELS MODELS |-- pllava-13b | |-- added_tokens.json | |-- config.json | |-- generation_config.json | |-- model-00001-of-00006.safetensors | |-- model-00002-of-00006.safetensors | |-- model-00003-of-00006.safetensors | |-- model-00004-of-00006.safetensors | |-- model-00005-of-00006.safetensors | |-- model-00006-of-00006.safetensors | |-- model.safetensors.index.json | |-- preprocessor_config.json | |-- processor_config.json | |-- special_tokens_map.json | |-- tokenizer.json | |-- tokenizer.model | `-- tokenizer_config.json |-- pllava-34b | |-- added_tokens.json | |-- config.json | |-- generation_config.json | |-- model-00001-of-00015.safetensors | |-- model-00002-of-00015.safetensors | |-- model-00003-of-00015.safetensors | |-- model-00004-of-00015.safetensors | |-- model-00005-of-00015.safetensors | |-- model-00006-of-00015.safetensors | |-- model-00007-of-00015.safetensors | |-- model-00008-of-00015.safetensors | |-- model-00009-of-00015.safetensors | |-- model-00010-of-00015.safetensors | |-- model-00011-of-00015.safetensors | |-- model-00012-of-00015.safetensors | |-- model-00013-of-00015.safetensors | |-- model-00014-of-00015.safetensors | |-- model-00015-of-00015.safetensors | |-- model.safetensors-deprecated | |-- model.safetensors.index.json | |-- preprocessor_config.json | |-- processor_config.json | |-- special_tokens_map.json | |-- tokenizer.json | |-- tokenizer.model | `-- tokenizer_config.json |-- pllava-7b |-- added_tokens.json |-- config.json |-- generation_config.json |-- model-00001-of-00003.safetensors |-- model-00002-of-00003.safetensors |-- model-00003-of-00003.safetensors |-- model.safetensors.index.json |-- preprocessor_config.json |-- processor_config.json |-- special_tokens_map.json |-- tokenizer.json |-- tokenizer.model `-- tokenizer_config.json ``` With the above steps, you should be able to proceed on with the following usages. ### Run Application To run our models, make sure you have downloaded a model pretrained weights from the huggingface spaces. Then, run the following scripts with the corresponding path input. Since we are only training with lora and the projector, the model to be run are determined with: - **model_dir**: model directory, one with config.json as compatible with transformers. This refers to the base model's directory, for example "llava-hf/llava-v1.6-vicuna-7b-hf"/"ermu2001/pllava-7b"/"MODELS/pllava-7b". (default to: MODELS/plave-7b) - **weights_dir**: your weights directory. could be the same as model_dir, but if you have a weights directory for the lora weights, you should set this weights_dir to that directory to load the lora weights. This directory should be local. Also, it would need to contain a config.json file within. (default to: ${model_dir}). ```bash model_dir="model directory" weights_dir="weights directory" bash scripts/demo.sh ${model_dir} ${weights_dir} ``` Now check out the application demo and try play with PLLAVA! ### Train Follow the following steps to reproduce our results or train your own variant: #### 1. Data Preparation To train our model from a starting Image-aligned Vision LLM, you would need to download the data first. Our data set up is mainly based on the original Videochat2's training data. Check out [Instruction Data](./DATA.md) to prepare the instruction training data. Ideally, setting up a root data directory and alter the code [here](./tasks/train/instruction_data.py#L6) would accomodate the data for training most smoothly. #### 2. Start Training Now you're only a few step away from starting the training. Follow the instructions: ##### Setup Accelerator Customize a accelerate training config. For example, a simple config using multiple gpus with no distribution strategy (only torch DDP) would look like: ```yaml compute_environment: LOCAL_MACHINE debug: false distributed_type: MULTI_GPU downcast_bf16: 'no' gpu_ids: all machine_rank: 0 main_training_function: main mixed_precision: bf16 num_machines: 1 num_processes: 8 rdzv_backend: static same_network: true tpu_env: [] tpu_use_cluster: false tpu_use_sudo: false use_cpu: false ``` Check out out the [Accelerate](https://huggingface.co/docs/accelerate/index) documents for more details. ##### Overwatch the training configuration Next, you should go over a basic training configuration of the training process in [here](tasks/train/config_pllava_nframe.py). Then passing this file as the first arg to the [training script](tasks/train/train_pllava_nframe_accel.py) would utilize every arguments in the file. You can customize some of the hyper parameters for your own training process by passing them in the format of "key" "value" pair in the following arguments. A example training scripts could be find [here](scripts/train_pllava.sh). We recommand customize a [configuration](tasks/train/config_pllava_nframe.py) to set up a customized training! With the above steps, you would be able to start the training process. The output would be well organized in the output directory, each a qualified model directory to pass in to demo as weights_dir, since we are only saveing the lora weights and projector weights to avoide redundancy. ### Evaluation This section mainly introduce how to reproduce the evaluation or evaluate your own model. #### Set up Evaluation Data Make sure you set up the "DATAS" directory as in [DATA.md](DATA.md), then you would be able to run the inference with fortune! The evaluation data directory of DATAS would look like: ``` DATAS/: DATAS/VideoQA: DATAS/VideoQA/TGIF_QA: test_a.json test_q.json DATAS/VideoQA/TGIF_QA/videos: tumblr_m4387mGrlc1r6m5e8o1_250.gif ... DATAS/VideoQA/TGIF_QA/videos_mp4: tumblr_m4387mGrlc1r6m5e8o1_250.mp4 ... DATAS/VideoQA/TGIF_QA/video_gif: tumblr_m4387mGrlc1r6m5e8o1_250.gif ... DATAS/VideoQA/MSVD_Zero_Shot_QA: test_a.json test_q.json DATAS/VideoQA/MSVD_Zero_Shot_QA/videos: -4wsuPCjDBc_5_15.avi DATAS/VideoQA/MSVD_Zero_Shot_QA/msvd_qa: DATAS/VideoQA/ActivityNet: test_a.json test_q.json DATAS/VideoQA/ActivityNet/all_test: v_--tFD65KaK4.mp4 ... DATAS/VideoQA/MSRVTT_Zero_Shot_QA: test_a.json test_q.json DATAS/VideoQA/MSRVTT_Zero_Shot_QA/videos: DATAS/VideoQA/MSRVTT_Zero_Shot_QA/videos/all: video0.mp4 ... DATAS/MVBench: ... DATAS/Recaption/Inter4K: annotations.json DATAS/Recaption/Inter4K/60fps: DATAS/Recaption/Inter4K/60fps/UHD: 1.mp4 ... ``` #### Start Evaluate Once you have construted the evaluation data, you can start the evaluation as in [here](scripts/eval.sh). This script is for evaluating 7B/13B models. As pllava-34b model uses a slightly different prompting, it is evaluated with this [script](scripts/eval_yiprompt.sh). ``` bash scripts/eval.sh ``` Same as running the demo, you would need to determine the model_dir and weights_dir to evaluate the model. Feel free to comment out some commands and produce partial evaluation. #### Overwatch the Results The evaluation results would be shown to you with our results gallery demo: ```bash bash scripts/gallery.sh ``` Feel free to use the compare version to compare differnt models' results or use the single gallery version to check out one model's results. They are basically the same. Check out this [script](scripts/gallery.sh) for more details #### For Captioning and Recaptioning Follow instructions at [DATA.md](DATA.md#extending-reacptioning) and you can extend the recaptioning data with a few steps. Feel free to point out high quality dataset of videos, we would proceed on doing captioning on those datasets. # :page_facing_up: Citation If you find this project useful in your research, please consider cite: ```BibTeX @misc{xu2024pllava, title={PLLaVA : Parameter-free LLaVA Extension from Images to Videos for Video Dense Captioning}, author={Lin Xu and Yilin Zhao and Daquan Zhou and Zhijie Lin and See Kiong Ng and Jiashi Feng}, year={2024}, eprint={2404.16994}, archivePrefix={arXiv}, primaryClass={cs.CV} } ``` # :dizzy: Acknowledgement This code base is mainly built upon [Videochat2](https://github.com/OpenGVLab/Ask-Anything/tree/main/video_chat2). SALUTE. We would also like to recognize and commend the following open source projects, thank you for your great contribution to the open source community: - [LLaVA](https://github.com/haotian-liu/LLaVA): Fantastic Open Source Image LLM Model. - [VideoChatGPT](https://github.com/mbzuai-oryx/Video-ChatGPT/tree/main): Great Evaluation Benchmarking Framework. - [VideoLlava](https://github.com/PKU-YuanGroup/Video-LLaVA/tree/main/videollava)：Video LLM repo with helpful resources. ", Assign "at most 3 tags" to the expected json: {"id":"9789","tags":[]} "only from the tags list I provide: [{"id":77,"name":"3d"},{"id":89,"name":"agent"},{"id":17,"name":"ai"},{"id":54,"name":"algorithm"},{"id":24,"name":"api"},{"id":44,"name":"authentication"},{"id":3,"name":"aws"},{"id":27,"name":"backend"},{"id":60,"name":"benchmark"},{"id":72,"name":"best-practices"},{"id":39,"name":"bitcoin"},{"id":37,"name":"blockchain"},{"id":1,"name":"blog"},{"id":45,"name":"bundler"},{"id":58,"name":"cache"},{"id":21,"name":"chat"},{"id":49,"name":"cicd"},{"id":4,"name":"cli"},{"id":64,"name":"cloud-native"},{"id":48,"name":"cms"},{"id":61,"name":"compiler"},{"id":68,"name":"containerization"},{"id":92,"name":"crm"},{"id":34,"name":"data"},{"id":47,"name":"database"},{"id":8,"name":"declarative-gui "},{"id":9,"name":"deploy-tool"},{"id":53,"name":"desktop-app"},{"id":6,"name":"dev-exp-lib"},{"id":59,"name":"dev-tool"},{"id":13,"name":"ecommerce"},{"id":26,"name":"editor"},{"id":66,"name":"emulator"},{"id":62,"name":"filesystem"},{"id":80,"name":"finance"},{"id":15,"name":"firmware"},{"id":73,"name":"for-fun"},{"id":2,"name":"framework"},{"id":11,"name":"frontend"},{"id":22,"name":"game"},{"id":81,"name":"game-engine "},{"id":23,"name":"graphql"},{"id":84,"name":"gui"},{"id":91,"name":"http"},{"id":5,"name":"http-client"},{"id":51,"name":"iac"},{"id":30,"name":"ide"},{"id":78,"name":"iot"},{"id":40,"name":"json"},{"id":83,"name":"julian"},{"id":38,"name":"k8s"},{"id":31,"name":"language"},{"id":10,"name":"learning-resource"},{"id":33,"name":"lib"},{"id":41,"name":"linter"},{"id":28,"name":"lms"},{"id":16,"name":"logging"},{"id":76,"name":"low-code"},{"id":90,"name":"message-queue"},{"id":42,"name":"mobile-app"},{"id":18,"name":"monitoring"},{"id":36,"name":"networking"},{"id":7,"name":"node-version"},{"id":55,"name":"nosql"},{"id":57,"name":"observability"},{"id":46,"name":"orm"},{"id":52,"name":"os"},{"id":14,"name":"parser"},{"id":74,"name":"react"},{"id":82,"name":"real-time"},{"id":56,"name":"robot"},{"id":65,"name":"runtime"},{"id":32,"name":"sdk"},{"id":71,"name":"search"},{"id":63,"name":"secrets"},{"id":25,"name":"security"},{"id":85,"name":"server"},{"id":86,"name":"serverless"},{"id":70,"name":"storage"},{"id":75,"name":"system-design"},{"id":79,"name":"terminal"},{"id":29,"name":"testing"},{"id":12,"name":"ui"},{"id":50,"name":"ux"},{"id":88,"name":"video"},{"id":20,"name":"web-app"},{"id":35,"name":"web-server"},{"id":43,"name":"webassembly"},{"id":69,"name":"workflow"},{"id":87,"name":"yaml"}]" returns me the "expected json"