corecv 0.1.4

CoreCV: Production-ready Computer Vision Library

CoreCV - Unified Vision Engine

Production-Ready Computer Vision Engine with Edge-Aware Optimization & Unified Deployment API

Overview • Key Features • Installation • Quickstart • Model Zoo • Edge Export • Documentation

---

Overview

CoreCV is a high-performance, production-grade computer vision library built on PyTorch. It bridges the gap between research agility and edge deployment constraints by consolidating model construction, training loops, multi-source inference, and hardware-targeted export into a unified facade API (CoreModel).

With CoreCV, you can instantiate any vision architecture with a single string, train with automatic mixed precision and edge graph rewrites, run GPU-accelerated inference across heterogeneous sources, and export directly to ONNX and ExecuTorch formats.

---

Key Features

• 🎯 Unified Facade API (CoreModel): Manage complete model lifecycles through three clean methods: .train(), .predict(), and .export().
• 🦁 Built-in Model Zoo: 13 Backbones (ResNet, MobileNetV3, ConvNeXt, ViT), 2 Feature Necks (FPN, PANet), and 5 Task Heads (Linear, ASPP Decoder, ResUNet Decoder, Decoupled Anchor-Free, Query Transformer Head).
• ⚡ Polymorphic & Dynamic Component Assembly: Instantiate models via plain backbone string names (e.g. "resnet50"), raw Python dictionaries, or .yaml files. Swap necks and heads on the fly (neck="panet", head="query_detection").
• 🚀 Edge-First Optimization:
  • TargetRewriter: Automatic edge-hardware graph rewrites (GELU → ReLU, SiLU → Hardswish, LayerNorm channel collapses).
  • MetaProber: Zero-VRAM shape propagation and static-graph audit using PyTorch meta device tensors.
• 📦 Multi-Format Export: One-line export pipeline producing optimized ONNX (.onnx) and ExecuTorch (.pte) artifacts.
• 🛡️ Type-Safe Static Registry: CoreRegistry with pre-instantiation signature checking and signature kwarg filtering.

---

Installation

Install CoreCV via uv (recommended for fast dependency resolution) or standard pip:

Using uv

uv pip install corecv

Or within a uv project workspace:

uv add corecv

Using pip

pip install corecv

---

Quickstart

1-Line Config-Driven Training (YAML / Dict)

from corecv.api import CoreModel

# 1. Load entire model architecture, dataset, & training config from a single YAML file
model = CoreModel("configs/detection_config.yaml")

# 2. Train - Auto-builds model, dataloaders, loss functions, and executes pipeline!
model.train()

3-Line Python Quickstart

from corecv.api import CoreModel

# 1. Instantiate model directly using a backbone string name
model = CoreModel("resnet18", task="classification", num_classes=10)

# 2. Train with edge-aware optimizations
model.train(data="./dataset", epochs=10, lr=1e-3, target_hardware="edge")

# 3. Predict & Export to edge deployment formats
predictions = model.predict("test_image.jpg", topk=5)
paths = model.export(format="onnx", target_hardware="edge")

---

Model Zoo Catalogue

CoreCV features a modular, decoupled architecture where any backbone can be paired with any neck and head:

Category	Component Family	Registry Keys	Description
Backbones	ResNet	resnet18, resnet34, resnet50, resnet101	Residual convolutional feature extractors
	MobileNetV3	mobilenet_v3_small, mobilenet_v3_large	Ultra-lightweight edge backbones
	ConvNeXt	convnext_tiny, convnext_small, convnext_base, convnext_large	Modernized conv-nets with transformer design choices
	ViT	vit_tiny, vit_small, vit_base	Vision Transformers with SimplePyramidAdapter
Necks	FPN	fpn	Top-down Feature Pyramid Network
	PANet	panet	Path Aggregation Network (FPN + bottom-up path)
Heads	Classification	linear_classification	Global average pooling + linear classifier
	Segmentation	aspp_decoder	DeepLabV3+ style ASPP decoder
		resunet_decoder	U-Net style residual skip-connection decoder
	Detection	decoupled_anchor_free	YOLOX/FCOS style decoupled conv head
		query_detection	RT-DETR/D-FINE style query transformer decoder

Flexible Component Swapping

from corecv.api import CoreModel

# Detection: MobileNetV3-Large + PANet Neck + Query Transformer Head
detector = CoreModel(
    "mobilenet_v3_large",
    task="detection",
    neck="panet",
    head="query_detection",
    neck_channels=128,
    num_classes=80,
)

# Segmentation: ConvNeXt-Tiny + ResUNet Decoder
segmentor = CoreModel(
    "convnext_tiny",
    task="segmentation",
    head="resunet_decoder",
    decoder_channels=128,
    num_classes=19,
)

---

Multi-Source Inference

CoreModel.predict() processes single image files, image folders, NumPy arrays, or PyTorch tensors out of the box with GPU-native preprocessing and FP16 support:

# Infer on single image file, folder, or tensor
results = model.predict(
    source="data/test_images/",
    conf_threshold=0.3,
    half_precision=True,  # FP16
    batch_size=16,
    weights="checkpoints/best.pt",  # On-the-fly weights loading
)

for res in results:
    if res.detection:
        print(f"Image {res.image_path}: {len(res.detection.boxes)} boxes detected")

---

Edge Hardware Export

CoreCV compiles and verifies models for edge deployment without allocating GPU VRAM:

# Export model to ONNX & ExecuTorch simultaneously with edge rewrites
paths = model.export(
    format="both",  # Produces .onnx and .pte
    target_hardware="edge",  # Applies GELU->ReLU & SiLU->Hardswish rewrites
    opset=18,
    output_path="exports/detector_edge",
)

print(f"ONNX Model: {paths['onnx']}")
print(f"ExecuTorch Model: {paths['executorch']}")

---

Documentation

The official CoreCV documentation is hosted on GitHub Pages:

👉 https://liturriago.github.io/corecv/

Build and Serve Locally

# Build the documentation site
uv run mkdocs build --strict

# Start the live preview server
uv run mkdocs serve

Once running, open http://127.0.0.1:8000 in your browser.

---

License

This project is licensed under the MIT License.