CVMD

中文文档

One unified inference interface for deployment-ready vision models.

Why CVMD

Use one consistent inference interface across multiple vision model families, without dragging training code into deployment.

• One API for multiple model families: run YOLO and DETR-style models with the same model(image) workflow.
• TorchScript-first deployment: load export-ready weights directly and keep inference environments lean.
• Easy to adopt, easy to scale: start from simple single-image inference, then extend to sliding-window or Ray-based distributed pipelines.

Installation

pip install cvmd

Quick Start

import imageio.v3 as iio
from cvmd import build

model = build("yolov11det", weights="yolo11l.torchscript", device="cuda")
model.load_model()

image = iio.imread("image.jpg")
results = model(image)
# results: [x1, y1, x2, y2, confidence, class]

Supported Models

Core API

• build(model_name_or_cls, **kwargs): build a model instance by name or class.
• list_models(): list registered model names.
• register_model(*names): register a custom model class.

Detection models return:

# np.ndarray, shape=(N, 6)
# [x1, y1, x2, y2, confidence, class]

Segmentation models return:

# (detections, masks)
# detections: np.ndarray, shape=(N, 6)
# masks: np.ndarray, shape=(N, H, W)

More Docs

• Chinese README
• Guide
• Examples and tests

Development

git clone <this repository>
cd cvmd
uv sync --dev