First follow the Modal setup instructions found here.

1

Setting up your Image

Overeasy requires a few dependencies that are not available in the base Python image. In addition, to reduce cold start times its recommended to cache model weights in your image.

image.py
import modal
overeasy_image = (
    Image.debian_slim(python_version="3.11")
    .run_commands("apt-get update && apt-get install -y ffmpeg libsm6 libxext6")
    .pip_install("overeasy==0.1.6")
    .run_commands(
        "python -c 'import overeasy as ov; ov.models.warmup_models();'", gpu="any"
    )
    .env({
        "ANTHROPIC_API_KEY": "sk-...",
        "GOOGLE_API_KEY": "...",
        "OPENAI_API_KEY": "sk-..."
    })
)
2

Run Overeasy Code

Then you can run overeasy code remotely on a Modal container like this:

overeasy_script.py
import modal
from overeasy import *
from pydantic import BaseModel
from PIL import Image

app = modal.App() # Initialize modal app

@app.function(image=overeasy_image, gpu="a100") # Run the function remotely on a GPU
def run_overeasy(image: Image.Image):
    class PPE(BaseModel):
        hardhat: bool

    workflow = Workflow([
        BoundingBoxSelectAgent(classes=["person"]),
        NMSAgent(iou_threshold=0.5, score_threshold=0),
        SplitAgent(),
        DenseCaptioningAgent(),
        InstructorTextAgent(response_model=PPE),
        ToClassificationAgent(fn=lambda x: "has ppe" if x.hardhat else "no ppe"),
        JoinAgent()
    ])

    results, graph = workflow.execute(image)
    return results

@app.local_entrypoint() # Run the function locally
def main():
    image_path = "./examples/construction_workers.jpg"
    image = Image.open(image_path)
    results = run_overeasy.remote(image)
    results[0].visualize().save("modal_result.jpg")



modal run overeasy_script.py