EgoGPT-0.5b-Demo

EgoGPT-0.5b-Demo

Model Summary

EgoGPT-0.5b-Demo is an omni-modal model trained on egocentric datasets, achieving state-of-the-art performance on egocentric video understanding. Built on the foundation of llava-onevision-qwen2-0.5b-ov, it has been finetuned on EgoIT-EgoLife-138k egocentric datasets, which contains EgoIT-99k and depersonalized version of EgoLife-QA (39k).

EgoGPT excels in two primary scenarios:

• Advanced Model Integration: EgoGPT combines LLaVA-OneVision and Whisper, improving its ability to process visual and auditory information.
• Outstanding Benchmark Performance: EgoGPT excels in egocentric benchmarks like EgoSchema, EgoPlan, and EgoThink, surpassing leading commercial and open-source models.

For further details, please refer to the following resources:

• 📰 Paper: https://arxiv.org/abs/2503.03803
• 🪐 Project Page: https://github.com/EvolvingLMMs-Lab/EgoLife
• 📦 Datasets: https://huggingface.co/datasets/lmms-lab/EgoIT-99K & https://huggingface.co/datasets/lmms-lab/EgoLife
• 🤗 Model Collections: https://huggingface.co/collections/lmms-lab/egolife-67c04574c2a9b64ab312c342

Usage

Installation

1. Clone this repository.

git clone https://github.com/egolife-ntu/EgoLife
cd EgoLife/EgoGPT

1. Install the dependencies.

conda create -n egogpt python=3.10
conda activate egogpt
pip install --upgrade pip
pip install -e .

3. Install the dependencies for training and inference.

```shell
pip install -e ".[train]"
pip install flash-attn --no-build-isolation

Quick Start

import argparse
import copy
import os
import re
import sys
import warnings

import numpy as np
import requests
import soundfile as sf
import torch
import torch.distributed as dist
import whisper
from decord import VideoReader, cpu
from egogpt.constants import (
    DEFAULT_IMAGE_TOKEN,
    DEFAULT_SPEECH_TOKEN,
    IGNORE_INDEX,
    IMAGE_TOKEN_INDEX,
    SPEECH_TOKEN_INDEX,
)
from egogpt.conversation import SeparatorStyle, conv_templates
from egogpt.mm_utils import get_model_name_from_path, process_images
from egogpt.model.builder import load_pretrained_model
from PIL import Image
from scipy.signal import resample


def setup(rank, world_size):
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "12355"
    dist.init_process_group("gloo", rank=rank, world_size=world_size)


def load_video(video_path=None, audio_path=None, max_frames_num=16, fps=1):
    if audio_path is not None:
        speech, sample_rate = sf.read(audio_path)
        if sample_rate != 16000:
            target_length = int(len(speech) * 16000 / sample_rate)
            speech = resample(speech, target_length)
        if speech.ndim > 1:
            speech = np.mean(speech, axis=1)
        speech = whisper.pad_or_trim(speech.astype(np.float32))
        speech = whisper.log_mel_spectrogram(speech, n_mels=128).permute(1, 0)
        speech_lengths = torch.LongTensor([speech.shape[0]])
    else:
        speech = torch.zeros(3000, 128)
        speech_lengths = torch.LongTensor([3000])

    vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
    total_frame_num = len(vr)
    avg_fps = round(vr.get_avg_fps() / fps)
    frame_idx = [i for i in range(0, total_frame_num, avg_fps)]
    if max_frames_num > 0 and len(frame_idx) > max_frames_num:
        uniform_sampled_frames = np.linspace(
            0, total_frame_num - 1, max_frames_num, dtype=int
        )
        frame_idx = uniform_sampled_frames.tolist()
    video = vr.get_batch(frame_idx).asnumpy()
    return video, speech, speech_lengths


def split_text(text, keywords):
    pattern = "(" + "|".join(map(re.escape, keywords)) + ")"
    parts = re.split(pattern, text)
    parts = [part for part in parts if part]
    return parts


def main(
    pretrained_path="checkpoints/EgoGPT-0.5b-Demo",
    video_path=None,
    audio_path=None,
    query="Please describe the video in detail.",
):
    warnings.filterwarnings("ignore")
    setup(0, 1)
    device = "cuda"
    device_map = "cuda"

    tokenizer, model, max_length = load_pretrained_model(
        pretrained_path, device_map=device_map
    )
    model.eval()

    conv_template = "qwen_1_5"
    question = f"<image>\n<speech>\n\n{query}"
    conv = copy.deepcopy(conv_templates[conv_template])
    conv.append_message(conv.roles[0], question)
    conv.append_message(conv.roles[1], None)
    prompt_question = conv.get_prompt()

    video, speech, speech_lengths = load_video(
        video_path=video_path, audio_path=audio_path
    )
    speech = torch.stack([speech]).to(device).half()
    processor = model.get_vision_tower().image_processor
    processed_video = processor.preprocess(video, return_tensors="pt")["pixel_values"]
    image = [(processed_video, video[0].size, "video")]

    parts = split_text(prompt_question, ["<image>", "<speech>"])
    input_ids = []
    for part in parts:
        if part == "<image>":
            input_ids.append(IMAGE_TOKEN_INDEX)
        elif part == "<speech>":
            input_ids.append(SPEECH_TOKEN_INDEX)
        else:
            input_ids.extend(tokenizer(part).input_ids)

    input_ids = torch.tensor(input_ids, dtype=torch.long).unsqueeze(0).to(device)
    image_tensor = [image[0][0].half()]
    image_sizes = [image[0][1]]
    generate_kwargs = {"eos_token_id": tokenizer.eos_token_id}

    cont = model.generate(
        input_ids,
        images=image_tensor,
        image_sizes=image_sizes,
        speech=speech,
        speech_lengths=speech_lengths,
        do_sample=False,
        temperature=0.5,
        max_new_tokens=4096,
        modalities=["video"],
        **generate_kwargs,
    )
    text_outputs = tokenizer.batch_decode(cont, skip_special_tokens=True)
    print(text_outputs)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--pretrained_path", type=str, default="lmms-lab/EgoGPT-0.5b-Demo"
    )
    parser.add_argument("--video_path", type=str, default=None)
    parser.add_argument("--audio_path", type=str, default=None)
    parser.add_argument(
        "--query", type=str, default="Please describe the video in detail."
    )
    args = parser.parse_args()
    main(args.pretrained_path, args.video_path, args.audio_path, args.query)

Citation

@inproceedings{yang2025egolife,
  title={EgoLife: Towards Egocentric Life Assistant},
  author={Yang, Jingkang and Liu, Shuai and Guo, Hongming and Dong, Yuhao and Zhang, Xiamengwei and Zhang, Sicheng and Wang, Pengyun and Zhou, Zitang and Xie, Binzhu and Wang, Ziyue and Ouyang, Bei and Lin, Zhengyu and Cominelli, Marco and Cai, Zhongang and Zhang, Yuanhan and Zhang, Peiyuan and Hong, Fangzhou and Widmer, Joerg and Gringoli, Francesco and Yang, Lei and Li, Bo and Liu, Ziwei},
  booktitle={The IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2025},
}