mvkit

Machine vision inspection for production lines — a framework you install, not a library you call.

mvkit runs an unattended inspection loop: wait for a trigger, grab a frame, run an inspection chain, report the verdict to the line, repeat. Cameras, triggers, result sinks and inspection modules are all plugins, so a new site is a configuration change rather than a code change.

No PyTorch in the default install. The runtime is onnxruntime + numpy + opencv.

Install

pip install mvkit              # CPU
pip install "mvkit[gpu]"       # NVIDIA CUDA via onnxruntime-gpu
pip install "mvkit[genicam]"   # GenICam/GenTL industrial cameras

On GPU, upgrade with --no-deps. onnxruntime and onnxruntime-gpu ship the same onnxruntime module and overwrite each other. onnxruntime (CPU) is a core dependency, so a plain pip install -U mvkit pulls it back in and you silently drop to CPU:

pip uninstall -y onnxruntime         # after installing mvkit[gpu], keep one build
pip install -U --no-deps mvkit       # upgrades from now on

mvkit doctor reports this state and prints the way back.

Then:

mvkit serve                    # http://127.0.0.1:8000 — dashboard and REST API
mvkit doctor                   # what is installed, what is missing, and why

The dashboard opens on a demo file camera reading the bundled sample images, so a fresh install has something to look at before any hardware arrives.

What it does

The runtime keeps capture, inference and dispatch on separate threads with bounded queues, so a slow PLC never stalls inference and a fast line never silently piles up frames. Camera drop-outs reconnect with backoff; a watchdog heartbeat tells the line when the inspector itself has stopped.

Inspection modules

Model weights are not bundled in the wheel. They are downloaded per module on first use, or imported from a disk:

mvkit models download --module patchcore
mvkit models import D:\mvkit_assets
mvkit models status

Teaching a part

Upload five to ten good samples in the dashboard; no defect images are required and no labelling is involved. The result is stored on your machine and never leaves it, and a product folder can be copied to another machine as it is.

Teaching from the command line is not available yet.

What needs to be installed separately

Two things cannot come from PyPI.

Industrial camera SDKs. harvesters talks to a GenTL producer (.cti), and that file ships with the camera vendor's runtime — Basler pylon, Hikrobot MVS, Teledyne FLIR Spinnaker, and so on. Install the vendor runtime and mvkit will find the producer on its own. Order does not matter; the search happens when a camera is opened, not at install time.

PyTorch, for mvkit[ocr] and mvkit[train] only. PyPI metadata cannot express which CUDA build you need, so install it yourself first:

pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install "mvkit[train]"

mvkit doctor reports both, and distinguishes "not installed" from "broken".

Where things are stored

A product folder is self-contained: copy it to another machine and inspection works there. Retraining caches live outside it and can be deleted.

Line PCs often have a small system drive. mvkit doctor prints the artifacts path with free space; move it with mvkit config set artifacts_dir D:\....

Evaluation

All modules are available for three months. After that, barcode, presence, count and color keep working; the rest need a licence.

An expired licence never stops a running line. Inspection in progress continues; only starting a new run and teaching new products are blocked, and the dashboard warns 30, 7 and 1 days ahead.

Your teaching results are yours. They are not deleted when a licence expires and can always be exported.

For a licence, email pashidl.lab@gmail.com with your name, organisation and intended use.

License

Proprietary. Evaluation and non-commercial research use only — see LICENSE. Contact the copyright holder for commercial licensing.