Apollo: An Exploration of Video Understanding in Large Multimodal Models

Apollo is a family of Large Multimodal Models (LMMs) that push the state-of-the-art in video understanding. It supports tasks including:

Long-form video comprehension
Temporal reasoning
Complex video question-answering
Multi-turn conversations grounded in video content

Apollo models excel at handling hour-long videos, balancing speed and accuracy through strategic design decisions. Our models outperform most 7B competitors at just 3B parameters and even rival 30B-scale models.

Key Highlights:

Scaling Consistency: Design decisions validated on smaller models and datasets effectively transfer to larger scales, reducing computation and experimentation costs.
Efficient Video Sampling: fps sampling and advanced token resampling strategies (e.g., Perceiver) yield stronger temporal perception.
Encoder Synergies: Combining SigLIP-SO400M (image) with InternVideo2 (video) delivers a robust representation, outperforming single encoders on temporal tasks.
ApolloBench: A streamlined evaluation benchmark (41x faster) that focuses on true video understanding capabilities.

Quick Start

Installation:

pip install -e .
pip install flash-attn --no-build-isolation

Inference Example:

import torch
from transformers import AutoModelForCausalLM
from apollo.mm_utils import (
    KeywordsStoppingCriteria,
    tokenizer_mm_token,
    ApolloMMLoader
)
from apollo.conversation import conv_templates, SeparatorStyle
from huggingface_hub import snapshot_download

model_url = "Apollo-LMMs/Apollo-3B-t32"
model_path = snapshot_download(model_url, repo_type="model")

device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    low_cpu_mem_usage=True
).to(device=device, dtype=torch.bfloat16)

tokenizer = model.tokenizer
vision_processors = model.vision_tower.vision_processor
config = model.config
num_repeat_token = config.mm_connector_cfg['num_output_tokens']
mm_processor = ApolloMMLoader(
    vision_processors,
    config.clip_duration,
    frames_per_clip=4,
    clip_sampling_ratio=0.65,
    model_max_length=config.model_max_length,
    device=device,
    num_repeat_token=num_repeat_token
)

video_path = "path/to/video.mp4"
question = "Describe this video in detail"
mm_data, replace_string = mm_processor.load_video(video_path)

conv = conv_templates["qwen_2"].copy()
conv.append_message(conv.roles[0], replace_string + "\n\n" + question)
conv.append_message(conv.roles[1], None)

prompt = conv.get_prompt()
input_ids = tokenizer_mm_token(prompt, tokenizer, return_tensors="pt").unsqueeze(0).to(device)

stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
stopping_criteria = KeywordsStoppingCriteria([stop_str], tokenizer, input_ids)

with torch.inference_mode():
    output_ids = model.generate(
        input_ids,
        vision_input=[mm_data],
        data_types=['video'],
        do_sample=True,
        temperature=0.4,
        max_new_tokens=256,
        top_p=0.7,
        use_cache=True,
        num_beams=1,
        stopping_criteria=[stopping_criteria]
    )

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

Citation

If you find this project useful, please consider citing:

@article{zohar2024apollo,
    title={Apollo: An Exploration of Video Understanding in Large Multimodal Models},
    author={Zohar, Orr and Wang, Xiaohan and Dubois, Yann and Mehta, Nikhil and Xiao, Tong and Hansen-Estruch, Philippe and Yu, Licheng and Wang, Xiaofang and Juefei-Xu, Felix and Zhang, Ning and Yeung-Levy, Serena and Xia, Xide},
    journal={arXiv preprint arXiv:2412.10360},
    year={2024}
}

For more details, visit the project website or check out the paper.