traincv

No-code computer vision training — from dataset to deployable model in one command.

traincv lets you train production-grade computer vision models (detection, classification, segmentation) with a single command or a 3-line Python API. It wraps Ultralytics YOLO and torchvision behind a unified interface, auto-converts between common dataset formats (COCO, YOLO, Pascal VOC), offers zero-shot auto-labeling via SAM2 / YOLO-World / CLIP, and exports trained models to ONNX, TorchScript, or TFLite ready for deployment.

Built by Viet-Anh Nguyen at NRL.ai.

Why traincv?

• One-liner API — traincv.train("dataset/", task="detect") is a complete pipeline
• Plugin architecture — Register custom backbones, losses, and exporters
• Local-first — Trains on your GPU, no cloud credits or upload required
• Minimal core deps — Training stacks (YOLO / torchvision) are optional extras
• Production-ready — Auto-splitting, logging, checkpointing, resumable runs

Installation

pip install traincv

For training backends:

pip install traincv[yolo]          # Ultralytics YOLOv8/v11 for detect + segment
pip install traincv[torch]         # torchvision for classification + DeepLabV3
pip install traincv[autolabel]     # SAM2 + YOLO-World + CLIP for auto-annotation
pip install traincv[export]        # ONNX + TorchScript + TFLite exporters
pip install traincv[all]           # everything

Python 3.8+ supported (tested on 3.8, 3.9, 3.10, 3.11, 3.12, 3.13)

Quick Start

import traincv

# 1. Train an object detector (auto-detects YOLO-format dataset and uses YOLOv8n)
run = traincv.train(
    "datasets/pets/",      # expects images/ + labels/ or data.yaml
    task="detect",
    model="yolov8n",
    epochs=50,
    imgsz=640,
)
print(run.best_map50)       # best mAP50 on validation
print(run.weights_path)     # path to best checkpoint

# 2. Export to ONNX for deployment (compatible with anycv / anydeploy)
traincv.export(run.weights_path, format="onnx", out="pets.onnx")

# 3. Use the trained model immediately
preds = traincv.predict("pets.onnx", "new_image.jpg")

Models & Methods

Training backends

Dataset formats (auto-detected)

• YOLO — images/ + labels/ (one .txt per image) + data.yaml
• COCO — annotations.json with images / annotations / categories
• Pascal VOC — JPEGImages/ + Annotations/*.xml
• ImageFolder — One sub-folder per class (classification only)

traincv converts between formats on the fly:

traincv.convert("coco_dataset/", to="yolo", out="yolo_dataset/")

Auto-labeling (zero-shot annotation)

Install traincv[autolabel] to bootstrap labels without manual work:

# Generate YOLO-format bounding boxes using only text prompts
traincv.autolabel(
    "unlabeled_images/",
    method="yolo-world",
    classes=["person", "dog", "bicycle"],
    out="auto_labels/",
)

Export formats

API Reference

CLI Usage

# Train
traincv train datasets/pets/ --task detect --model yolov8n --epochs 50

# Evaluate
traincv evaluate runs/detect/best.pt --data datasets/pets/

# Export
traincv export runs/detect/best.pt --format onnx --out pets.onnx

# Auto-label with text prompts
traincv autolabel unlabeled/ --method yolo-world --classes "person,dog,cat"

# Convert datasets
traincv convert coco_ds/ --to yolo --out yolo_ds/

Examples

Full end-to-end: label, train, export

import traincv

# 1. Auto-label raw images with SAM2 + YOLO-World
traincv.autolabel("raw/", method="yolo-world",
                  classes=["product", "price_tag"], out="labeled/")

# 2. Split into train/val/test (80/10/10)
traincv.split("labeled/", ratios=(0.8, 0.1, 0.1))

# 3. Train YOLOv8s
run = traincv.train("labeled/", task="detect", model="yolov8s", epochs=100)

# 4. Export for mobile deployment
traincv.export(run.weights_path, format="tflite", out="products.tflite")

Fine-tune a classifier on ImageFolder data

import traincv

run = traincv.train(
    "flowers/",            # flowers/daisy/*.jpg, flowers/rose/*.jpg, ...
    task="classify",
    model="mobilenetv3",
    epochs=30,
    lr=1e-3,
)
print(run.best_accuracy)

Resume an interrupted run

traincv.train("datasets/pets/", task="detect", resume="runs/detect/exp5/last.pt")

License

MIT (c) Viet-Anh Nguyen