Cephalo is a series of multimodal materials science focused vision large language models (V-LLMs) designed to integrate visual and linguistic data for advanced understanding and interaction in human-AI or multi-agent AI frameworks.
A novel aspect of Cephalo's development is the innovative dataset generation method. The extraction process employs advanced algorithms to accurately detect and separate images and their corresponding textual descriptions from complex PDF documents. It involves extracting images and captions from PDFs to create well-reasoned image-text pairs, utilizing large language models (LLMs) for natural language processing. These image-text pairs are then refined and validated through LLM-based NLP processing, ensuring high-quality and contextually relevant data for training.
Cephalo can interpret complex visual scenes and generating contextually accurate language descriptions and answer queries.
The model is developed to process diverse inputs, including images and text, facilitating a broad range of applications such as image captioning, visual question answering, and multimodal content generation. The architecture combines a vision encoder model and an autoregressive transformer to process complex natural language understanding.
Cephalo provides a robust framework for multimodal interaction and understanding, including the development of complex generative pipelines to create 2D and 3D renderings of material microstructures as input for additive manufacturing methods.
This version of Cephalo, lamm-mit/Cephalo-Idefics-2-vision-10b-beta, is based on a merged expansion of the https://huggingface.co/lamm-mit/Cephalo-Idefics-2-vision-8b-beta and the HuggingFaceM4/idefics2-8b-chatty model. This method allows us to increase the depth of the model and focus on learning more complex representations and associations in deeper layers of the network.
The lamm-mit/Cephalo-Idefics-2-vision-10b-beta model is trained for two epochs, while the lamm-mit/Cephalo-Idefics-2-vision-10b-alpha version was trained for one epoch.
The model was trained in several stages:
Step 1: Train https://huggingface.co/lamm-mit/Cephalo-Idefics-2-vision-8b-beta by fine-tuning the HuggingFaceM4/idefics2-8b-chatty model.
Step 2: Combine the https://huggingface.co/lamm-mit/Cephalo-Idefics-2-vision-8b-beta decoder with the last 8 layers of the HuggingFaceM4/idefics2-8b-chatty decoder.
Step 3: Fine-tune the merged model, which now has 40 decoder layers and a total of 10b parameters.
The model was trained on a combination of scientific text-image data extracted from Wikipedia and scientific papers. For further details on the base model, see: https://huggingface.co/HuggingFaceM4/idefics2-8b-chatty. More details about technical aspects of the model, training and example applications to materials science problems are provided in the paper (reference at the bottom).
The lamm-mit/Cephalo-Idefics-2-vision-10b-beta model is suitable for one or more image inputs, wih prompts using the chat format as follows:
User: You carefully study the image, and respond accurately, but succinctly. Think step-by-step.
<image>What is shown in this image, and what is the relevance for materials design? Include a discussion of multi-agent AI.<end_of_utterance>
Assistant:
where the model generates the text after Assistant: . For multi-turn conversations, the prompt should be formatted as follows:
User: You carefully study the image, and respond accurately, but succinctly. Think step-by-step.
<image>What is shown in this image, and what is the relevance for materials design? Include a discussion of multi-agent AI.<end_of_utterance>
Assistant: The image depicts ants climbing a vertical surface using their legs and claws. This behavior is observed in nature and can inspire the design of multi-agent AI systems that mimic the coordinated movement of these insects. The relevance lies in the potential application of such systems in robotics and materials science, where efficient and adaptive movement is crucial.<end_of_utterance>
User: How could this be used to design a fracture resistant material?<end_of_utterance>
Assistant:
If you need to manually set the chat template:
IDEFICS2_CHAT_TEMPLATE = "{% for message in messages %}{{message['role'].capitalize()}}{% if message['content'][0]['type'] == 'image' %}{{':'}}{% else %}{{': '}}{% endif %}{% for line in message['content'] %}{% if line['type'] == 'text' %}{{line['text']}}{% elif line['type'] == 'image' %}{{ '<image>' }}{% endif %}{% endfor %}<end_of_utterance>\n{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}"
This code snippets show how to get quickly started on a GPU:
from PIL import Image
import requests
DEVICE='cuda:0'
from transformers import AutoProcessor, Idefics2ForConditionalGeneration
from tqdm.notebook import tqdm
model_id='lamm-mit/Cephalo-Idefics-2-vision-10b-beta'
model = Idefics2ForConditionalGeneration.from_pretrained( model_id,
torch_dtype=torch.bfloat16, #if your GPU allows
_attn_implementation="flash_attention_2", #make sure Flash Attention 2 is installed
trust_remote_code=True,
).to (DEVICE)
processor = AutoProcessor.from_pretrained(
f"{model_id}",
do_image_splitting=True
)
See section towards the end for more comments on model optimization, including quantization.
If you need to manually set the chat template:
IDEFICS2_CHAT_TEMPLATE = "{% for message in messages %}{{message['role'].capitalize()}}{% if message['content'][0]['type'] == 'image' %}{{':'}}{% else %}{{': '}}{% endif %}{% for line in message['content'] %}{% if line['type'] == 'text' %}{{line['text']}}{% elif line['type'] == 'image' %}{{ '<image>' }}{% endif %}{% endfor %}<end_of_utterance>\n{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}"
tokenizer = AutoTokenizer.from_pretrained(base_model_id, use_fast=True)
tokenizer.chat_template = IDEFICS2_CHAT_TEMPLATE
processor.tokenizer = tokenizer
Simple inference example:
from transformers.image_utils import load_image
image = load_image("https://d2r55xnwy6nx47.cloudfront.net/uploads/2018/02/Ants_Lede1300.jpg")
# Create inputs
messages = [
{
"role": "user",
"content": [
{"type": "image"},
{"type": "text", "text": "What is shown in this image, and what is the relevance for materials design? Include a discussion of multi-agent AI."},
]
},
]
prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
# Get inputs using the processor
inputs = processor(text=prompt, images=[image], return_tensors="pt")
inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
# Generate
generated_ids = model.generate(**inputs, max_new_tokens=500)
generated_texts = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(generated_texts)
Next we provide a convenience function for inference. This function takes the model, processor, question, and images, along with messages and images objects for repeated chat-like interactions with the model.
def ask_about_image (model, processor, question,
images_input=[],
verbatim=False,
temperature=0.1,
show_image=False,
system="You are a biomaterials scientist who responds accurately. ",
init_instr = "",
show_conversation=True,
max_new_tokens=256,
messages=[],
images=[],
use_Markdown=False,
):
query = question
images_input=ensure_list(images_input)
if len (images)==0:
if len (images_input)>0:
for image in tqdm (images_input) :
if is_url(image):
image= load_image(image)
images.append (image)
if show_image:
display ( image )
if len (messages)==0:
base_message = {
"role": "user",
"content": [
{"type": "text", "text": system + init_instr},
# Image messages will be added dynamically here
{"type": "text", "text": query}
]
}
# Ensure the images_input is a list
images_input = ensure_list(images_input)
# Add image messages dynamically
image_messages = [{"type": "image"} for _ in images_input]
base_message["content"][1:1] = image_messages # Insert image messages before the last text message
# Append the constructed message to messages list
messages.append(base_message)
else:
messages.append (
{
"role": "user",
"content": [
{"type": "text", "text": query
}
]
}
)
if verbatim:
print (messages)
text = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(text=[text.strip()], images=images, return_tensors="pt", padding=True).to(DEVICE)
generated_ids = model.generate(**inputs, max_new_tokens=max_new_tokens, temperature=temperature, do_sample=True)
generated_texts = processor.batch_decode(generated_ids[:, inputs["input_ids"].size(1):], skip_special_tokens=True)
messages.append (
{
"role": "assistant",
"content": [ {"type": "text", "text": generated_texts[0]}, ]
}
)
formatted_conversation = format_conversation(messages, images)
# Display the formatted conversation, e.g. in Jupyter Notebook
if show_conversation:
if use_Markdown:
display(Markdown(formatted_conversation))
else:
display(HTML(formatted_conversation))
return generated_texts, messages, images
question = "What is shown in this image, and what is the relevance for materials design? Include a discussion of multi-agent AI."
url1 = "https://d2r55xnwy6nx47.cloudfront.net/uploads/2018/02/Ants_Lede1300.jpg"
response, messages,images= ask_about_image ( model, processor, question,
images_input=[url1,],
temperature=0.1,
system= '', init_instr='You carefully study the image and provide detailed answers. Think step-by-step.\n\n',
show_conversation=True,
max_new_tokens=512, messages=[], images=[])
Sample output:
Image by Vaishakh Manohar
The image shows a group of ants moving in coordinated patterns on a surface. This illustrates the concept of multi-agent AI, which involves the study and simulation of complex systems involving multiple agents (in this case, ants) interacting with each other and their environment. The relevance for materials design is in understanding how these natural systems exhibit emergent behaviors such as self-organization, which can inspire the development of new materials and systems that mimic these natural processes. By studying the movement patterns of ants, researchers can gain insights into how to design materials that exhibit similar emergent properties, leading to improved performance in various applications. Multi-agent AI involves creating models that describe the interactions between individual agents and their environment, allowing for the simulation of complex systems with multiple interacting components. This approach can be applied to various fields, including materials science, where understanding emergent behaviors at the microscopic level can lead to the design of new materials with enhanced properties.
The schematic below shows a visualization of the approach to generate datasets for training the vision model. The extraction process employs advanced algorithms to accurately detect and separate images and their corresponding textual descriptions from complex PDF documents. It involves extracting images and captions from PDFs to create well-reasoned image-text pairs, utilizing large language models (LLMs) for natural language processing. These image-text pairs are then refined and validated through LLM-based NLP processing, ensuring high-quality and contextually relevant data for training.
The image below shows reproductions of two representative pages of the scientific article (here, Spivak, Buehler, et al., 2011), and how they are used to extract visual scientific data for training the Cephalo model.
If your GPU allows, load and run inference in half precision (torch.float16 or torch.bfloat16).
model = AutoModelForVision2Seq.from_pretrained(
"lamm-mit/Cephalo-Idefics-2-vision-10b-beta",
+ torch_dtype=torch.float16,
).to(DEVICE)
Vision encoder efficiency
Given the high resolution supported, the vision part of the model can be memory hungry depending on your configuration. If you are GPU-memory-constrained, you can:
do_image_splitting=False when initializing the processor (AutoProcessor.from_pretrained). There are no changes required on the model side. Note that only the sft model has been trained with image splitting.size= {"longest_edge": 448, "shortest_edge": 378} when initializing the processor (AutoProcessor.from_pretrained). In particular, the longest_edge value can be adapted to fit the need (the default value is 980). We recommend using values that are multiples of 14. There are no changes required on the model side.do_image_splitting=True is especially needed to boost performance on complex tasks where a very large image is used as input. The model was fine-tuned with image splitting turned on. For simple tasks, this argument can be safely set to False.
Using Flash-attention 2 to speed up generation
Mke sure to install flash-attn. Refer to the original repository of Flash Attention for the package installation. Simply change the snippet above with:
model = AutoModelForVision2Seq.from_pretrained(
"lamm-mit/Cephalo-Idefics-2-vision-10b-beta",
+ torch_dtype=torch.bfloat16,
+ _attn_implementation="flash_attention_2",
).to(DEVICE)
4 bit quantization with bitsandbytes
+ from transformers import BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
bnb_4bit_compute_dtype=torch.bfloat16
)
model = AutoModelForVision2Seq.from_pretrained(
"lamm-mit/Cephalo-Idefics-2-vision-10b-beta",
+ torch_dtype=torch.bfloat16,
+ quantization_config=quantization_config,
).to(DEVICE)
Please cite as:
@article{Buehler_Cephalo_2024,
title={Cephalo: Multi-Modal Vision-Language Models for Bio-Inspired Materials Analysis and Design},
author={Markus J. Buehler},
journal={arXiv preprint arXiv:2405.19076},
year={2024}
}