Llama-3.1-Dragonfly

Note: Users are permitted to use this model in accordance with the Llama 3.1 Community License Agreement.

Model Details

Dragonfly is a multimodal visual-language model, trained by instruction tuning on Llama 3.1.

• Developed by: Together AI
• Model type: An autoregressive visual-language model based on the transformer architecture
• License: Llama 3.1 Community License Agreement
• Finetuned from model: Llama 3.1

Model Sources

• Repository: https://github.com/togethercomputer/Dragonfly
• Paper: https://arxiv.org/abs/2406.00977

Uses

The primary use of Dragonfly is research on large visual-language models. It is primarily intended for researchers and hobbyists in natural language processing, machine learning, and artificial intelligence.

How to Get Started with the Model

💿 Installation

Create a conda environment and install necessary packages

conda env create -f environment.yml
conda activate dragonfly_env

Install flash attention

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

As a final step, please run the following command.

pip install --upgrade -e .

🧠 Inference

If you have successfully completed the installation process, then you should be able to follow the steps below.

Question: What is so funny about this image?

Load necessary packages

import torch
from PIL import Image
from transformers import AutoProcessor, AutoTokenizer

from dragonfly.models.modeling_dragonfly import DragonflyForCausalLM
from dragonfly.models.processing_dragonfly import DragonflyProcessor
from pipeline.train.train_utils import random_seed

Instantiate the tokenizer, processor, and model.

device = torch.device("cuda:0")

tokenizer = AutoTokenizer.from_pretrained("togethercomputer/Llama-3.1-8B-Dragonfly-v2")
clip_processor = AutoProcessor.from_pretrained("openai/clip-vit-large-patch14-336")
image_processor = clip_processor.image_processor
processor = DragonflyProcessor(image_processor=image_processor, tokenizer=tokenizer, image_encoding_style="llava-hd")

model = DragonflyForCausalLM.from_pretrained("togethercomputer/Llama-3.1-8B-Dragonfly-v2")
model = model.to(torch.bfloat16)
model = model.to(device)

Now, lets load the image and process them.

image = Image.open("./test_images/skateboard.png")
image = image.convert("RGB")
images = [image]
# images = [None] # if you do not want to pass any images

text_prompt = "user\n\nWhat is so funny about this image?assistant\n\n"

inputs = processor(text=[text_prompt], images=images, max_length=4096, return_tensors="pt", is_generate=True)
inputs = inputs.to(device)

Finally, let us generate the responses from the model

temperature = 0

with torch.inference_mode():
    generation_output = model.generate(**inputs, max_new_tokens=1024, eos_token_id=tokenizer.encode(""), do_sample=temperature > 0, temperature=temperature, use_cache=True)

generation_text = processor.batch_decode(generation_output, skip_special_tokens=False)

An example response.

The humor in this image comes from the surreal juxtaposition of a dog's face with the body of the Mona Lisa, a famous painting by Leonardo da Vinci.
The Mona Lisa is known for her enigmatic smile and is often considered one of the most famous paintings in the world. By combining the dog's face with
the body of the Mona Lisa, the artist has created a whimsical and amusing image that plays on the viewer 's expectations and familiarity with the
original paintings. The contrast between the dog's natural, expressive features and the serene, mysterious expression of the Mona Lisa creates a
humerous effect that is likely to elicit laughter

Training Details

See more details in the "Implementation" section of our paper.

Evaluation

See more details in the "Results" section of our paper.

🏆 Credits

We would like to acknowledge the following resources that were instrumental in the development of Dragonfly:

• Meta Llama 3.1: We utilized the Llama 3 model as our foundational language model.
• CLIP: Our vision backbone is CLIP model from OpenAI.
• Our codebase is built upon the following two codebases:
  • Otter: A Multi-Modal Model with In-Context Instruction Tuning
  • LLaVA-UHD: an LMM Perceiving Any Aspect Ratio and High-Resolution Images

📚 BibTeX

@misc{thapa2024dragonfly,
      title={Dragonfly: Multi-Resolution Zoom-In Encoding Enhances Vision-Language Models},
      author={Rahul Thapa and Kezhen Chen and Ian Covert and Rahul Chalamala and Ben Athiwaratkun and Shuaiwen Leon Song and James Zou},
      year={2024},
      eprint={2406.00977},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Model Card Authors

Rahul Thapa, Kezhen Chen, Rahul Chalamala

Model Card Contact

Rahul Thapa (rahulthapa@together.ai), Kezhen Chen (kezhen@together.ai)