traincv 0.2.0

No-code labeling and training toolkit for computer vision

TrainCV

A no-code labeling and training toolkit for computer vision. Train object detection, classification, and segmentation models with minimal code via CLI or Python API.

Supports local HuggingFace models. Pre-trained models can be downloaded from HuggingFace Hub and cached locally for offline training and fine-tuning.

Built by NRL.ai

---

Installation

# Core (no training backends)
pip install traincv

# With PyTorch backend
pip install traincv[torch]

# With YOLO backend
pip install traincv[yolo]

# Everything
pip install traincv[full]

Quick Start

CLI

# Train an object detection model
traincv train --task detection --data ./dataset --model yolov8n --epochs 50

# Train an image classifier
traincv train --task classification --data ./images --model mobilenetv2

# Export a trained model to ONNX
traincv export --model ./best.pt --format onnx

# Validate dataset format
traincv validate --data ./dataset --format coco

Python API

import traincv

# Train a detector
traincv.train(task="detection", data="./dataset", model="yolov8n", epochs=50)

# Export to ONNX
traincv.export(model="./best.pt", format="onnx")

# Load and inspect a dataset
dataset = traincv.Dataset.load("./dataset", format="coco")
print(dataset.summary())

Supported Tasks

Task	Models	Backend
Object Detection	YOLOv8n/s/m/l/x, YOLOv5	ultralytics
Image Classification	MobileNetV2, ResNet18/34/50, EfficientNet-B0	torchvision
Segmentation	YOLOv8n/s/m-seg, DeepLabV3	ultralytics / torchvision

Dataset Formats

TrainCV supports the following annotation formats:

Format	Description	File Type
COCO	COCO JSON format	.json
YOLO	YOLO darknet format	.txt per image
VOC	Pascal VOC XML format	.xml per image

COCO Format

dataset/
  images/
    img001.jpg
    img002.jpg
  annotations/
    instances.json

YOLO Format

dataset/
  images/
    img001.jpg
    img002.jpg
  labels/
    img001.txt
    img002.txt

Pascal VOC Format

dataset/
  JPEGImages/
    img001.jpg
  Annotations/
    img001.xml

Model Zoo

Model	Task	Size	Speed
yolov8n	Detection	6.2M	Fast
yolov8s	Detection	11.2M	Fast
yolov8m	Detection	25.9M	Medium
yolov8l	Detection	43.7M	Slow
yolov8x	Detection	68.2M	Slow
mobilenetv2	Classification	3.5M	Fast
resnet18	Classification	11.7M	Fast
resnet50	Classification	25.6M	Medium
efficientnet_b0	Classification	5.3M	Fast

Export Formats

Format	Extension	Use Case
ONNX	.onnx	Cross-platform inference
TorchScript	.torchscript	PyTorch production

Configuration

Training parameters can be set via CLI flags, Python kwargs, or a YAML config file:

# config.yaml
task: detection
model: yolov8n
data: ./dataset
epochs: 50
batch_size: 16
lr: 0.001
augmentation: medium
input_size: [640, 640]
device: auto

traincv train --config config.yaml

Data Augmentation

Built-in augmentation presets:

Preset	Transforms
light	Horizontal flip, slight rotation
medium	Flip, rotation, color jitter, scale
heavy	Flip, rotation, color jitter, scale, mosaic, mixup

Local-First / Edge AI

This package supports downloading pre-trained models from HuggingFace Hub for offline use. After initial download, no internet connection is required for training or fine-tuning.

# Pre-download a model for offline use
python -m traincv download yolov8n

# Download a classification model
python -m traincv download mobilenetv2 --task classification

import traincv

# Download and cache a model locally
model_path = traincv.download_model("yolov8n")

# Train using the cached model (works offline)
traincv.train(task="detection", data="./dataset", model=str(model_path))

Links

• Website: nrl.ai
• GitHub: github.com/vietanhdev/traincv
• Author: Viet-Anh Nguyen

License

MIT License - see LICENSE for details.