MLAdaptiveIntelligence/LLaVAction-7B

1---2license: cc-by-nc-sa-4.03datasets:4- lmms-lab/LLaVA-Video-178K5language:6- en7metrics:8- accuracy9base_model:10- lmms-lab/LLaVA-Video-7B-Qwen211pipeline_tag: video-text-to-text12library_name: transformers13tags:14- Action15- Video16- MQA17- multimodal18- VLM19- LLaVAction20- MLLMs21model-index:22- name: LLaVAction-7B23  results:24  - task:25      type: multimodal26    dataset:27      name: EgoSchema28      type: egoschema29    metrics:30    - type: accuracy31      value: 5932      name: accuracy33      verified: true34  - task:35      type: multimodal36    dataset:37      name: MVBench38      type: mvbench39    metrics:40    - type: accuracy41      value: 61.142      name: accuracy43      verified: true44  - task:45      type: multimodal46    dataset:47      name: NextQA48      type: nextqa49    metrics:50    - type: accuracy51      value: 82.852      name: accuracy53      verified: true54  - task:55      type: multimodal56    dataset:57      name: PercepTest58      type: percepTest59    metrics:60    - type: accuracy61      value: 70.262      name: accuracy63      verified: true64  - task:65      type: multimodal66    dataset:67      name: LongVideoBench68      type: longvideobench69    metrics:70    - type: accuracy71      value: 58.672      name: accuracy73      verified: true74  - task:75      type: multimodal76    dataset:77      name: VideoMME78      type: videomme79    metrics:80    - type: accuracy81      value: 63.982      name: accuracy83      verified: true84  - task:85      type: multimodal86    dataset:87      name: VideoMME (w-subs)88      type: videomme89    metrics:90    - type: accuracy91      value: 71.492      name: accuracy93      verified: true94---95 96# LLaVAction-7B97 98<div align="center">99<h2>LLaVAction: evaluating and training multi-modal large language models for action recognition100</h2>101 102[Shaokai Ye](https://yeshaokai.github.io/)<sup>1**</sup>&nbsp; 103[Haozhe Qi](https://people.epfl.ch/haozhe.qi)<sup>1**</sup>&nbsp;104 105[Alexander Mathis](https://mathislab.org/)<sup>1</sup><sup>†</sup>&nbsp;106[Mackenzie Weygandt Mathis](https://www.mackenziemathislab.org/mackenziemathis)<sup>1</sup><sup>†</sup><sup>‡</sup>&nbsp;107 108<sup>1</sup> EPFL109 110<sup>**</sup> First authors  <sup>†</sup> Senior Authors  <sup>‡</sup> Corresponding Author111 112\[[arXiv Paper](arxiv.org/abs/2503.18712)\] &nbsp; \[[Project Page](https://mmathislab.github.io/llavaction/)\] &nbsp; \[[Github Repo](https://github.com/AdaptiveMotorControlLab/LLaVAction)\] &nbsp; 113 114</div>115 116## Model Summary117The LLaVAction-7B model is trained on EPIC-KITCHENS-100-MQA, based on Qwen2 language model with a context window of 32K tokens.118This model supports at most 64 frames.119 120- **Project Page**:  [https://mmathislab.github.io/llavaction/](https://mmathislab.github.io/llavaction/)121- **Paper**: For more details, please check our [paper](https://arxiv.org/abs/tbd)122- **Repository**:  [https://github.com/AdaptiveMotorControlLab/LLaVAction](https://github.com/AdaptiveMotorControlLab/LLaVAction)123- **Point of Contact**: [Mackenzie Mathis](https://people.epfl.ch/mackenzie.mathis)124- **Languages**: English125- 126## Useage127 128### Intended use129The model was trained on EPIC-KITCHENS-100-MQA [dataset release pending] and [LLaVA-Video-178K](https://huggingface.co/datasets/lmms-lab/LLaVA-Video-178K). It has improved capability on understanding human egocentric actions from videos.130 131 132### Generation133We provide the simple generation process for using our model. For more details, you could refer to our [Github](https://github.com/AdaptiveMotorControlLab/LLaVAction).134 135```python136!pip install llavaction137 138from llavaction.model.builder import load_pretrained_model139from llavaction.mm_utils import get_model_name_from_path, process_images, tokenizer_image_token140from llavaction.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, IGNORE_INDEX141from llavaction.conversation import conv_templates, SeparatorStyle142from PIL import Image143import requests144import copy145import torch146import sys147import warnings148from decord import VideoReader, cpu149import numpy as np150warnings.filterwarnings("ignore")151 152#Your video (it assumes an egocentric view point)153video_path = "XXXX"154 155#These are the prompts we trained with, but you can test others:156perspective_prompt = "You are seeing this video from egocentric view and you are the person. Your hands are sometimes interacting with objects. What action are you doing?"157task_prompt = "Describe in details what you see from the video frames."158 159def load_video(video_path, max_frames_num,fps=1,force_sample=False):160    if max_frames_num == 0:161        return np.zeros((1, 336, 336, 3))162    vr = VideoReader(video_path, ctx=cpu(0),num_threads=1)163    total_frame_num = len(vr)164    video_time = total_frame_num / vr.get_avg_fps()165    fps = round(vr.get_avg_fps()/fps)166    frame_idx = [i for i in range(0, len(vr), fps)]167    if len(frame_idx) > max_frames_num or force_sample:168        sample_fps = max_frames_num169        uniform_sampled_frames = np.linspace(0, total_frame_num - 1, sample_fps, dtype=int)170        frame_idx = uniform_sampled_frames.tolist()171        frame_time = [i/vr.get_avg_fps() for i in frame_idx]172    spare_frames = vr.get_batch(frame_idx).asnumpy()173    # import pdb;pdb.set_trace()174    return spare_frames,frame_time,video_time175 176pretrained = "MLAdaptiveIntelligence/LLaVAction-7B"177model_name = "llava_qwen"178device = "cuda"179device_map = "auto"180tokenizer, model, image_processor, max_length = load_pretrained_model(pretrained, None, model_name, torch_dtype="bfloat16", device_map=device_map)  # Add any other thing you want to pass in llava_model_args181model.eval()182max_frames_num = 64183video,frame_time,video_time = load_video(video_path, max_frames_num, 1, force_sample=True)184video = image_processor.preprocess(video, return_tensors="pt")["pixel_values"].cuda().to(torch.bfloat16)185video = [video]186conv_template = "qwen_1_5"  # Make sure you use correct chat template for different models187time_instruction = f"The video lasts for {video_time:.2f} seconds, and {len(video[0])} frames are uniformly sampled from it. "188question = DEFAULT_IMAGE_TOKEN + f"\n{time_instruction}\n{perspective_prompt} {task_prompt}"189 190conv = copy.deepcopy(conv_templates[conv_template])191conv.append_message(conv.roles[0], question)192conv.append_message(conv.roles[1], None)193prompt_question = conv.get_prompt()194input_ids = tokenizer_image_token(prompt_question, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(device)195 196cont = model.generate(197    input_ids,198    images=video,199    modalities= ["video"],200    do_sample=False,201    temperature=0,202    max_new_tokens=4096,203)204text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)[0].strip()205print(text_outputs)206```207 208 209## Training210 211See details in Ye et al. 2025: arxiv.org/abs/2503.18712212 213### Model214- **Architecture**: SO400M + Qwen2215- **Initialized Model**: lmms-lab/LLaVA-Video-7B-Qwen2216- **Data**: A mixture of LLaVA-178K and EPIC-KITCHENS-100-MQA, 2 epochs, full model217- **Precision**: bfloat16218 219 220### Hardware & Software221GPUs: 32 * Nvidia GH-200 (for whole model series training)222Orchestration: HuggingFace Trainer223Neural networks:  PyTorch224 225## Citation226 227arXiv: arxiv.org/abs/2503.18712228 229```bibtex230@article{YeQi2025llavaction,231  title={LLaVAction: evaluating and training multi-modal large language models for action recognition},232  author={Ye, Shaokai and Qi, Haozhe and Mathis, Alexander and Mathis, Mackenzie W.},233  journal={arXiv preprint},234  year={2025}235}236```