🧑‍💻 autocast - Stop writing boilerplate

If you work in Computer Vision or Machine Learning, you have likely written this code a thousand times:

# The "Dance of the Types" 🕺
points = prediction.detach().cpu().numpy()
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(points)

It is verbose. It is brittle. And worst of all, it breaks your flow. You stop thinking about geometry and start thinking about memory layouts, device placement, and wrappers.

I got tired of writing these same three lines in every project. I wanted a way to just say “give me a PointCloud” and have Python figure out the rest.

So, I built autocast.

A Graph of Types

autocast treats type conversion as a routing problem. Instead of manually writing the conversion chain, you just define the start and the end.

from autocast import acast
import open3d as o3d

# "Just give me a PointCloud, I don't care how."
pcd = acast(prediction, o3d.geometry.PointCloud)

Under the hood, autocast maintains a Directed Graph of all registered types.

When you ask for a conversion from A to C:

1. It checks if a direct A -> C converter exists.
2. If not, it runs a Breadth-First Search (BFS) to find the shortest path (e.g., A -> B -> C).
3. It executes the chain automatically.

Why autocast is Cool

1. It Handles the “Messy” Stuff

Deep learning pipelines often spit out complex nested structures: lists of dicts of tensors. Converting these manually requires writing recursive loops. autocast handles collections natively.

data = {
    "boxes": [torch.tensor(...), torch.tensor(...)],
    "scores": torch.tensor(...)
}

# Convert the WHOLE structure to NumPy arrays in one line
numpy_data = acast(data, np.ndarray)

2. Extensible (Monkey-Patch Friendly)

You can extend the graph with your own custom types using the @acaster decorator.

@acaster(MyCustomBBox, torch.Tensor)
def bbox_to_tensor(bbox):
    return torch.tensor([bbox.x, bbox.y, bbox.w, bbox.h])

Now, any time you call acast(bbox, np.ndarray), autocast knows it can go MyCustomBBox -> Tensor -> NumPy.

Try it out

It’s available on PyPI now.

Note: The package name on PyPI is acaster, but you import it as autocast.

pip install acaster

Check out the code on GitHub: github.com/marcbenedi/autocast

I’d love to hear if this saves you as much time as it saved me!