README.md

metadata

license: cc
datasets:
  - liuhaotian/LLaVA-Instruct-150K
  - liuhaotian/LLaVA-Pretrain
language:
  - en
pipeline_tag: video-text-to-text

Model Card for LLaVA-Video-LLaMA-3

Please follow my github repo LLaVA-Video-Llama-3 for more details on fine-tuning LLaVA model with Llama-3 as the foundatiaon LLM.

Updates

• [6/4/2024] The codebase supports the video data fine-tuning for video understanding tasks.
• [5/14/2024] The codebase has been upgraded to llava-next (llava-v1.6). Now it supports the latest llama-3, phi-3, mistral-v0.1-7b models.

Model Details

• Video Frame Sampling: Considering we adopt CLIP-ViT-L-336px as the image encoder (576 tokens for one image) and the context window of LLaMA-3 is 8k, the video frame sampling rate is set as max(30, num_frames//10).
• Template: We follow the LLaVA-v1 template for constructing the conversation.
• Architecture: LLaVA architecture, visual encoder + MLP adapter + LLM backbone

How to Use

Please firstly install llava via

pip install git+https://github.com/Victorwz/LLaVA-Video-Llama-3.git

You can load the model and perform inference as follows:

from llava.conversation import conv_templates, SeparatorStyle
from llava.model.builder import load_pretrained_model
from llava.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
from PIL import Image
import requests
import cv2
import torch
import base64
import io
from io import BytesIO
import numpy as np

# load model and processor
device = "cuda" if torch.cuda.is_available() else "cpu"
model_name = get_model_name_from_path("weizhiwang/LLaVA-Video-Llama-3")
tokenizer, model, image_processor, context_len = load_pretrained_model("weizhiwang/LLaVA-Video-Llama-3", None, model_name, False, False, device=device)

# prepare image input
url = "https://github.com/PKU-YuanGroup/Video-LLaVA/raw/main/videollava/serve/examples/sample_demo_1.mp4"

def read_video(video_url):
    response = requests.get(url)
    if response.status_code != 200:
        print("Failed to download video")
        exit()
    else:
        with open("tmp_video.mp4", 'wb') as f:
            for chunk in response.iter_content(chunk_size=1024):
                f.write(chunk)
    
    video = cv2.VideoCapture("tmp_video.mp4")

    base64Frames = []
    while video.isOpened():
        success, frame = video.read()
        if not success:
            break
        _, buffer = cv2.imencode(".jpg", frame)
        base64Frames.append(base64.b64encode(buffer).decode("utf-8"))

    video.release()
    print(len(base64Frames), "frames read.")
    return base64Frames

video_frames = read_video(video_url=url)
image_tensors = []
samplng_interval = int(len(video_frames) / 10)
for i in range(0, len(video_frames), samplng_interval):
    rawbytes = base64.b64decode(video_frames[i])
    image = Image.open(io.BytesIO(rawbytes)).convert("RGB")
    image_tensor = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0].half().cuda()
    image_tensors.append(image_tensor)

# prepare inputs for the model
text = "\n".join(['<image>' for i in range(len(image_tensors))]) + '\n' + "Why is this video funny"
conv = conv_templates["llama_3"].copy()
conv.append_message(conv.roles[0], text)
conv.append_message(conv.roles[1], None)
prompt = conv.get_prompt()
input_ids = tokenizer_image_token(prompt, tokenizer, -200, return_tensors='pt').unsqueeze(0).cuda()

# autoregressively generate text
with torch.inference_mode():
    output_ids = model.generate(
        input_ids,
        images=image_tensors,
        do_sample=False,
        max_new_tokens=512,
        use_cache=True)

outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
print(outputs[0])

The image caption results look like:

The video is funny because it shows a baby girl wearing glasses and reading a book, which is an unusual and amusing sight. It is not common to see a baby wearing glasses and engaging in a reading activity, as they are still developing their motor skills and cognitive abilities. The image captures a cute and endearing moment, as the baby appears to be enjoying her time and learning to read. This scene can evoke a sense of warmth and delight in the viewer, as it showcases the innocence and curiosity of childhood.

Fine-Tune LLaVA-Llama-3 on Your Video Instruction Data

Please refer to a forked LLaVA-Video-Llama-3 git repo for fine-tuning data preparation and scripts. The data loading function and fastchat conversation template are changed due to a different tokenizer.

Citation

@misc{wang2024llavavideollama3,
  title={LLaVA-Video-Llama-3: A Video Understanding Multimodal LLM based on Llama-3-8B LLM backbone},
  author={Wang, Weizhi},
  year={2024}
}