vision-agent 0.2.195

Toolset for Vision Agent

VisionAgent is a library that helps you utilize agent frameworks to generate code to solve your vision task. Check out our discord for updates and roadmaps!

Table of Contents

• 🚀Quick Start
• 📚Documentation
• 🔍🤖VisionAgent
• 🛠️Tools
• 🤖LMMs
• 💻🤖VisionAgent Coder
• 🏗️Additional Backends

Quick Start

Web Application

The fastest way to test out VisionAgent is to use our web application. You can find it here.

Installation

To get started with the python library, you can install it using pip:

pip install vision-agent

Ensure you have both an Anthropic key and an OpenAI API key and set in your environment variables (if you are using Azure OpenAI please see the Azure setup section):

export ANTHROPIC_API_KEY="your-api-key" # needed for VisionAgent and VisionAgentCoder
export OPENAI_API_KEY="your-api-key" # needed for ToolRecommender

Basic Usage

To get started you can just import the VisionAgent and start chatting with it:

>>> from vision_agent.agent import VisionAgent
>>> agent = VisionAgent(verbosity=2)
>>> resp = agent("Hello")
>>> print(resp)
[{"role": "user", "content": "Hello"}, {"role": "assistant", "content": "{'thoughts': 'The user has greeted me. I will respond with a greeting and ask how I can assist them.', 'response': 'Hello! How can I assist you today?', 'let_user_respond': True}"}]
>>> resp.append({"role": "user", "content": "Can you count the number of people in this image?", "media": ["people.jpg"]})
>>> resp = agent(resp)

The chat messages are similar to OpenAI's format with role and content keys but in addition to those you can add media which is a list of media files that can either be images or video files.

Documentation

VisionAgent Library Docs

VisionAgent Basic Usage

Chatting and Message Formats

VisionAgent is an agent that can chat with you and call other tools or agents to write vision code for you. You can interact with it like you would ChatGPT or any other chatbot. The agent uses Clause-3.5 for it's LMM and OpenAI for embeddings for searching for tools.

The message format is:

{
    "role": "user",
    "content": "Hello",
    "media": ["image.jpg"]
}

Where role can be user, assistant or observation if the agent has executed a function and needs to observe the output. content is always the text message and media is a list of media files that can be images or videos that you want the agent to examine.

When the agent responds, inside it's context you will find the following data structure:

{
    "thoughts": "The user has greeted me. I will respond with a greeting and ask how I can assist them.",
    "response": "Hello! How can I assist you today?",
    "let_user_respond": true
}

thoughts are the thoughts the agent had when processing the message, response is the response it generated which could contain a python execution, and let_user_respond is a boolean that tells the agent if it should wait for the user to respond before continuing, for example it may want to execute code and look at the output before letting the user respond.

Chatting and Artifacts

If you run chat_with_artifacts you will also notice an Artifact object. Artifact's are a way to sync files between local and remote environments. The agent will read and write to the artifact object, which is just a pickle object, when it wants to save or load files.

import vision_agent as va
from vision_agent.tools.meta_tools import Artifact

artifact = Artifact("artifact.pkl")
# you can store text files such as code or images in the artifact
with open("code.py", "r") as f:
    artifacts["code.py"] = f.read()
with open("image.png", "rb") as f:
    artifacts["image.png"] = f.read()

agent = va.agent.VisionAgent()
response, artifacts = agent.chat_with_artifacts(
    [
        {
            "role": "user",
            "content": "Can you write code to count the number of people in image.png",
        }
    ],
    artifacts=artifacts,
)

Running the Streamlit App

To test out things quickly, sometimes it's easier to run the streamlit app locally to chat with VisionAgent, you can run the following command:

pip install -r examples/chat/requirements.txt
export WORKSPACE=/path/to/your/workspace
export ZMQ_PORT=5555
streamlit run examples/chat/app.py

You can find more details about the streamlit app here, there are still some concurrency issues with the streamlit app so if you find it doing weird things clear your workspace and restart the app.

Tools

There are a variety of tools for the model or the user to use. Some are executed locally while others are hosted for you. You can easily access them yourself, for example if you want to run owl_v2_image and visualize the output you can run:

import vision_agent.tools as T
import matplotlib.pyplot as plt

image = T.load_image("dogs.jpg")
dets = T.owl_v2_image("dogs", image)
# visualize the owl_v2_ bounding boxes on the image
viz = T.overlay_bounding_boxes(image, dets)

# plot the image in matplotlib or save it
plt.imshow(viz)
plt.show()
T.save_image(viz, "viz.png")

Or if you want to run on video data, for example track sharks and people at 10 FPS:

frames_and_ts = T.extract_frames_and_timestamps("sharks.mp4", fps=10)
# extract only the frames from frames and timestamps
frames = [f["frame"] for f in frames_and_ts]
# track the sharks and people in the frames, returns segmentation masks
track = T.florence2_sam2_video_tracking("shark, person", frames)
# plot the segmentation masks on the frames
viz = T.overlay_segmentation_masks(frames, track)
T.save_video(viz, "viz.mp4")

You can find all available tools in vision_agent/tools/tools.py, however the VisionAgent will only utilizes a subset of tools that have been tested and provide the best performance. Those can be found in the same file under the FUNCION_TOOLS variable inside tools.py.

Custom Tools

If you can't find the tool you are looking for you can also add custom tools to the agent:

import vision_agent as va
import numpy as np

@va.tools.register_tool(imports=["import numpy as np"])
def custom_tool(image_path: str) -> str:
    """My custom tool documentation.

    Parameters:
        image_path (str): The path to the image.

    Returns:
        str: The result of the tool.

    Example
    -------
    >>> custom_tool("image.jpg")
    """

    return np.zeros((10, 10))

You need to ensure you call @va.tools.register_tool with any imports it uses. Global variables will not be captured by register_tool so you need to include them in the function. Make sure the documentation is in the same format above with description, Parameters:, Returns:, and Example\n-------. The VisionAgent will use your documentation when trying to determine when to use your tool. You can find an example use case here for adding a custom tool. Note you may need to play around with the prompt to ensure the model picks the tool when you want it to.

Can't find the tool you need and want us to host it? Check out our vision-agent-tools repository where we add the source code for all the tools used in VisionAgent.

LMMs

All of our agents are based off of LMMs or Large Multimodal Models. We provide a thin abstraction layer on top of the underlying provider APIs to be able to more easily handle media.

from vision_agent.lmm import AnthropicLMM

lmm = AnthropicLMM()
response = lmm("Describe this image", media=["apple.jpg"])
>>> "This is an image of an apple."

Or you can use the OpenAI chat interaface and pass it other media like videos:

response = lmm(
    [
        {
            "role": "user",
            "content": "What's going on in this video?",
            "media": ["video.mp4"]
        }
    ]
)

VisionAgent Coder

Underneath the hood, VisionAgent uses VisionAgentCoder to generate code to solve vision tasks. You can use VisionAgentCoder directly to generate code if you want:

>>> from vision_agent.agent import VisionAgentCoder
>>> agent = VisionAgentCoder()
>>> code = agent("What percentage of the area of the jar is filled with coffee beans?", media="jar.jpg")

Which produces the following code:

from vision_agent.tools import load_image, florence2_sam2_image

def calculate_filled_percentage(image_path: str) -> float:
    # Step 1: Load the image
    image = load_image(image_path)

    # Step 2: Segment the jar
    jar_segments = florence2_sam2_image("jar", image)

    # Step 3: Segment the coffee beans
    coffee_beans_segments = florence2_sam2_image("coffee beans", image)

    # Step 4: Calculate the area of the segmented jar
    jar_area = 0
    for segment in jar_segments:
        jar_area += segment['mask'].sum()

    # Step 5: Calculate the area of the segmented coffee beans
    coffee_beans_area = 0
    for segment in coffee_beans_segments:
        coffee_beans_area += segment['mask'].sum()

    # Step 6: Compute the percentage of the jar area that is filled with coffee beans
    if jar_area == 0:
        return 0.0  # To avoid division by zero
    filled_percentage = (coffee_beans_area / jar_area) * 100

    # Step 7: Return the computed percentage
    return filled_percentage

To better understand how the model came up with it's answer, you can run it in debug mode by passing in the verbose argument:

>>> agent = VisionAgentCoder(verbosity=2)

Detailed Usage

You can also have it return more information by calling generate_code. The format of the input is a list of dictionaries with the keys role, content, and media:

>>> results = agent.generate_code([{"role": "user", "content": "What percentage of the area of the jar is filled with coffee beans?", "media": ["jar.jpg"]}])
>>> print(results)
{
    "code": "from vision_agent.tools import ..."
    "test": "calculate_filled_percentage('jar.jpg')",
    "test_result": "...",
    "plans": {"plan1": {"thoughts": "..."}, ...},
    "plan_thoughts": "...",
    "working_memory": ...,
}

With this you can examine more detailed information such as the testing code, testing results, plan or working memory it used to complete the task.

Multi-turn conversations

You can have multi-turn conversations with vision-agent as well, giving it feedback on the code and having it update. You just need to add the code as a response from the assistant:

agent = va.agent.VisionAgentCoder(verbosity=2)
conv = [
    {
        "role": "user",
        "content": "Are these workers wearing safety gear? Output only a True or False value.",
        "media": ["workers.png"],
    }
]
result = agent.generate_code(conv)
code = result["code"]
conv.append({"role": "assistant", "content": code})
conv.append(
    {
        "role": "user",
        "content": "Can you also return the number of workers wearing safety gear?",
    }
)
result = agent.generate_code(conv)

Additional Backends

E2B Code Execution

If you wish to run your code on the E2B backend, make sure you have your E2B_API_KEY set and then set CODE_SANDBOX_RUNTIME=e2b in your environment variables. This will run all the agent generated code on the E2B backend.

Anthropic

AnthropicVisionAgentCoder uses Anthropic. To get started you just need to get an Anthropic API key and set it in your environment variables:

export ANTHROPIC_API_KEY="your-api-key"

Because Anthropic does not support embedding models, the default embedding model used is the OpenAI model so you will also need to set your OpenAI API key:

export OPEN_AI_API_KEY="your-api-key"

Usage is the same as VisionAgentCoder:

>>> import vision_agent as va
>>> agent = va.agent.AnthropicVisionAgentCoder()
>>> agent("Count the apples in the image", media="apples.jpg")

OpenAI

OpenAIVisionAgentCoder uses OpenAI. To get started you just need to get an OpenAI API key and set it in your environment variables:

export OPEN_AI_API_KEY="your-api-key"

Usage is the same as VisionAgentCoder:

>>> import vision_agent as va
>>> agent = va.agent.OpenAIVisionAgentCoder()
>>> agent("Count the apples in the image", media="apples.jpg")

Ollama

OllamaVisionAgentCoder uses Ollama. To get started you must download a few models:

ollama pull llama3.1
ollama pull mxbai-embed-large

llama3.1 is used for the OllamaLMM for OllamaVisionAgentCoder. Normally we would use an actual LMM such as llava but llava cannot handle the long context lengths required by the agent. Since llama3.1 cannot handle images you may see some performance degredation. mxbai-embed-large is the embedding model used to look up tools. You can use it just like you would use VisionAgentCoder:

>>> import vision_agent as va
>>> agent = va.agent.OllamaVisionAgentCoder()
>>> agent("Count the apples in the image", media="apples.jpg")

WARNING: VisionAgent doesn't work well unless the underlying LMM is sufficiently powerful. Do not expect good results or even working code with smaller models like Llama 3.1 8B.

Azure OpenAI

AzureVisionAgentCoder uses Azure OpenAI models. To get started follow the Azure Setup section below. You can use it just like you would use VisionAgentCoder:

>>> import vision_agent as va
>>> agent = va.agent.AzureVisionAgentCoder()
>>> agent("Count the apples in the image", media="apples.jpg")

Azure Setup

If you want to use Azure OpenAI models, you need to have two OpenAI model deployments:

1. OpenAI GPT-4o model
2. OpenAI text embedding model

Then you can set the following environment variables:

export AZURE_OPENAI_API_KEY="your-api-key"
export AZURE_OPENAI_ENDPOINT="your-endpoint"
# The deployment name of your Azure OpenAI chat model
export AZURE_OPENAI_CHAT_MODEL_DEPLOYMENT_NAME="your_gpt4o_model_deployment_name"
# The deployment name of your Azure OpenAI text embedding model
export AZURE_OPENAI_EMBEDDING_MODEL_DEPLOYMENT_NAME="your_embedding_model_deployment_name"

NOTE: make sure your Azure model deployment have enough quota (token per minute) to support it. The default value 8000TPM is not enough.

You can then run VisionAgent using the Azure OpenAI models:

import vision_agent as va
agent = va.agent.AzureVisionAgentCoder()

---

Q&A

How to get started with OpenAI API credits

1. Visit the OpenAI API platform to sign up for an API key.
2. Follow the instructions to purchase and manage your API credits.
3. Ensure your API key is correctly configured in your project settings.

Failure to have sufficient API credits may result in limited or no functionality for the features that rely on the OpenAI API. For more details on managing your API usage and credits, please refer to the OpenAI API documentation.