Skip to content

CLI Commands

Command Overview

Command Purpose
yolomatic Main interactive TUI
yolomatic-train Train from a saved YAML config
yolomatic-predict Run YOLO/RF-DETR prediction workflows
yolomatic-sam Run SAM 3.1 segmentation inference
yolomatic-convert Convert Labelbox or Ultralytics NDJSON to YOLO/COCO, including pose
yolomatic-prepare Prepare and split datasets
yolomatic-benchmark Benchmark checkpoints with an HTML report
yolomatic-upload Upload or deploy checkpoints to Roboflow
yolomatic-tensorboard Launch TensorBoard for discovered runs
yolomatic-ultralytics Ultralytics Platform helper workflows
bump patch\|minor\|major\|VERSION Update the package version

All commands are available as uv run <command> from the repository root, or directly after uv tool install --python 3.12 yolomatic.

yolomatic

Launch the main interactive TUI. All primary workflows are accessible from this menu.

uv run yolomatic
# or
yolomatic

TUI menu options:

  • Configure Model
  • Configure Fine-Tune
  • Train Model
  • Predict
  • Benchmark Models
  • Augment Dataset
  • Convert Dataset Format
  • Prepare Dataset
  • Upload to Roboflow
  • Launch TensorBoard
  • SAM Segmentation
  • Ultralytics Platform

yolomatic-train

Train from a previously saved YAML config file. If multiple configs exist in configs/, a selector is shown.

uv run yolomatic-train

The smart training router reads the config's family field and dispatches to the correct trainer:

  • YOLO*, YOLOX → Ultralytics YOLO trainer
  • RF-DETR* → Native RF-DETR trainer
  • SAM* → HuggingFace SAM trainer
  • Detectron2 → Detectron2 trainer

Runtime prompts:

  • If ClearML is not configured: continue without ClearML or cancel
  • If CUDA is requested but unavailable: repair environment, fall back to CPU, or cancel

yolomatic-predict

Run prediction on a single image or a folder of images using a trained YOLO or RF-DETR checkpoint.

uv run yolomatic-predict [--mode MODE] [--weight PATH] [--source PATH] [--conf FLOAT] [--workers INT]
Flag Default Description
--mode interactive Prediction mode: single for one image, folder for batch directory
--weight interactive Path to a .pt weight file; omit to use the interactive selector
--source interactive Image file or folder path; omit to be prompted
--input-dir — Alias for --source in folder mode
--conf 0.25 Confidence threshold for predictions
--workers 1 Number of worker processes for folder prediction; set > 1 to enable multiprocessing

Examples:

# Interactive wizard
uv run yolomatic-predict

# Single image
uv run yolomatic-predict --mode single --weight runs/detect/train/weights/best.pt --source image.jpg

# Batch folder with multiprocessing
uv run yolomatic-predict --mode folder --weight runs/detect/train/weights/best.pt --source datasets/test/images --workers 4

# Lower confidence threshold
uv run yolomatic-predict --mode single --weight best.pt --source image.jpg --conf 0.5

yolomatic-sam

Run SAM 3.1 segmentation inference. Supports auto, text-prompted, and box-prompted modes.

uv run yolomatic-sam

This command is wizard-only — all options are presented interactively. See SAM 3.1 guide for authentication and mode details.

yolomatic-convert

Convert Labelbox or Ultralytics-platform NDJSON exports to YOLO or COCO format, including explicit YOLO Pose and COCO Pose targets. Includes concurrent image downloading.

uv run yolomatic-convert

Wizard-only. The wizard auto-detects whether the source is a Labelbox or Ultralytics-platform NDJSON. See NDJSON Conversion for details.

yolomatic-prepare

Prepare and split a dataset into train/val/test subsets using random, class-balanced, or smart-balanced strategies.

uv run yolomatic-prepare

Wizard-only. See Smart Split for the splitting algorithm details.

yolomatic-benchmark

Benchmark one or more same-task Ultralytics YOLO checkpoints or exported model artifacts against compatible YOLO dataset groups and generate an interactive HTML report.

uv run yolomatic-benchmark

Wizard-only. Requires:

  1. A trained, downloaded, or exported Ultralytics model artifact
  2. A compatible labeled dataset with data.yaml; select train, valid, test, or all

See Benchmarking guide for the full workflow.

yolomatic-upload

Upload YOLO checkpoints to Roboflow or deploy RF-DETR checkpoints via the RF-DETR deployment API.

uv run yolomatic-upload [--weight PATH] [--workspace SLUG] [--project-ids IDS] [--model-name NAME] [--model-type TYPE] [--version INT]
Flag Default Description
--weight interactive Path to checkpoint (best.pt, last.pt). Omit for interactive selector.
--workspace from .env Roboflow workspace slug. Falls back to ROBOFLOW_WORKSPACE in .env.
--project-ids from .env Comma-separated project IDs. Falls back to ROBOFLOW_PROJECT_IDS.
--model-name auto Versionless model name to register in Roboflow.
--model-type auto-detected Override the Roboflow model type (e.g., yolo26n, yolo26l, rf-detr-l).
--version 1 Roboflow project version for RF-DETR deployment.

Examples:

# Interactive wizard
uv run yolomatic-upload

# Direct upload with explicit args
uv run yolomatic-upload \
  --weight runs/segment/train/weights/best.pt \
  --workspace my-workspace \
  --project-ids my-project \
  --model-type yolo26l \
  --model-name train2-best

YOLO26 model type

YOLO26 uploads require a size-specific model type: yolo26n, yolo26s, yolo26m, yolo26l, or yolo26x. The generic yolo26 type is not valid.

See Cloud Upload guide for the full Roboflow workflow.

yolomatic-tensorboard

Launch TensorBoard for discovered training run directories.

uv run yolomatic-tensorboard

Scans the project tree for TensorBoard event files and starts TensorBoard on port 6006. See TensorBoard guide.

yolomatic-ultralytics

Helpers for Ultralytics Platform workflows (dataset download, NDJSON export conversion).

uv run yolomatic-ultralytics

Wizard-only.

bump

Update the project version in src/__version__.py and pyproject.toml.

uv run bump patch        # 5.0.0 → 5.0.1
uv run bump minor        # 5.0.0 → 5.1.0
uv run bump major        # 5.0.0 → 6.0.0
uv run bump 5.2.0        # explicit version

Related pages: Quickstart, Configuration, Cloud upload.