TRAIN · AIVQC Trainer

Before a model reaches the line, it goes through Trainer.

You mark the defects and set the goal and time — automatic training does the rest. Trainer turns images of your product into a tested inspection model without machine-vision expertise and publishes it to AIVQC Production stations.

Inspection quality starts with data, not with the model. In Trainer you import images and recordings, mark every defect, confirm defect-free images and freeze a version of the dataset. Training always uses that frozen copy, and the test set stays independent — the model never learns from it.

Training is automatic: you set a time budget and a quality goal, and Trainer runs successive attempts, adjusts parameters and keeps the best model by validation results. You check the candidate on images and video through the same inference path the line uses, and only when it meets the agreed criteria is a deployment package created.

You don't need a machine-learning specialist to work with Trainer — a process or quality engineer who knows the product and its defects is enough. Trainer works outside the rhythm of production, so experiments don't affect the ongoing inspection, and nothing reaches the line until you deliberately deploy it.

Automatic training

A time budget and a quality goal instead of manual tuning. Trainer chooses the parameters, improves them after every attempt and keeps the best model.

Leak-free data

Validated, deduplicated import, groups for shots of the same part and frozen dataset versions — test results are not artificially inflated.

Evaluation like the line

The same inference path as Production, testing on video from the target camera and a recommended confidence threshold.

Deployment with evidence

A package with the model, thresholds and measured results. Export stays blocked while any result is below target.

WORKFLOW

Five workspaces, one path from image to package.

Each Trainer workspace is one stage of the work. The Overview screen always suggests the next step for the current project and shows what is still missing before deployment.

01
Dataset
Import images, recordings or existing datasets, review warnings, split the data and freeze a dataset version.
02
Annotation
A box around every defect and, optionally, every product. Defect-free images are explicitly confirmed as negatives.
03
Training
Automatic attempts within a time budget — Trainer chooses and improves the parameters, not you.
04
Evaluation
Testing on images and video through the same inference path as the line, and choosing the confidence threshold.
05
Deployment
An .aivqcpkg package with the model, thresholds and quality evidence, published through Server to a selected station.

AUTOMATIC TRAINING

You don't need to be a machine-vision expert.

You know your product and what counts as a defect. Choosing the model, batch size, learning rate and number of epochs is machine-learning specialist knowledge — and that is exactly the part Trainer does for you.

You only set a time budget and a quality goal (mAP@50). Trainer checks the data is ready, fits the parameters to your dataset and hardware, then runs successive complete training attempts. After each one it analyses the results and improves the next — until it reaches the goal or runs out of time. What you get at the end is the best model, not the last one.

Results stalling? Trainer lowers the learning rate and trains longer. Getting worse? It returns to the best known configuration. Unstable? It makes optimization gentler.
Model letting defects through or rejecting good parts? The next attempt adjusts in the direction that is missing — no guesswork.
Out of GPU memory? Trainer retries with a smaller batch on its own, without using up an attempt from the budget.
The model is chosen on validation data only, and the test set only reports — so the report shows how the model handles images it has never seen.
When new line images arrive, you fine-tune the previous model instead of starting from scratch.
Training runs in the background — on this computer or on a GPU worker on the local network — with progress shown at the bottom of every workspace.
For advanced users: a manual mode with the full parameter set and four detection architectures to compare.

Step	Traditionally — an ML specialist	In AIVQC Trainer
Choosing the model	Comparing architectures and configurations	A recommendation fitted to your data and hardware
Training parameters	Hand-tuning batch size, learning rate and epochs	Chosen automatically and adjusted after every attempt
Data checks	Custom scripts and reviews	A readiness check before any GPU time is spent
Stalling, regression, out of memory	Reading logs and restarting by hand	Automatic response within safe limits
Picking the best model	Collating results by hand	Best validation result plus an independent test report
Preparing for the line	Converting and verifying the model	ONNX export (FP32 or INT8) with metrics recalculated

Your role: mark the defects, set the goal and time, review the result in Evaluation. The decision to deploy is always yours.

DATA

A dataset you can trust.

Most model problems start in the data: duplicated frames, missed defects, a test set too similar to the training set. Trainer catches them before the first GPU hour is spent. How many images you realistically need is covered in How many images do you need to train an AI inspection?

Import of images (JPG, PNG, BMP, WebP), recordings and camera captures — with validation, SHA-256 hashing and duplicate detection even when file names differ.
Automatic recognition of existing datasets: Pascal VOC, YOLO, COCO, LabelMe and VIA. On an error the import pauses and shows the image instead of silently accepting bad boxes.
Dataset checks: class distribution with a warning for underrepresented defects, very small boxes, images without a product.
A defect-free image must be explicitly confirmed as negative — the model does not learn from images nobody has looked at.
Product and session groups: shots of the same part and frames from one recording always stay in one set, so test results are not artificially inflated.
Frozen snapshot: later annotation fixes do not change results you have already seen — after changes you prepare a new version.

EVALUATION

What you see in Trainer is what the station will do.

Evaluation runs the exported ONNX model through the same inference path AIVQC Production uses. The same images and the same threshold give the same result on the station. How to choose between false rejects and escaped defects is covered in False accept or false reject?

Testing on images outside training and on video from the target camera — at least 10 minutes covering good products, every known defect, several products at once and partial occlusions.
A recommended confidence threshold calculated from the evaluation data (best F1), with a range of acceptable values.
Per-class metrics: precision, recall, F1, mAP@50 and mAP@50–95. Low recall means escaped defects; low precision means false rejects.
Acceptance criteria (quality, FPS, latency) stored in the project. No package is created while any result is below target or the target-computer measurement is missing.

DEPLOYMENT AND FEEDBACK

Packages to the line, difficult images back.

The approved model goes into an .aivqcpkg package with its classes, preprocessing, thresholds, acceptance criteria and measured quality evidence. Trainer checks the dataset fingerprint — if images or annotations changed after training, no package is created. Publishing goes through AIVQC Server to a selected station, and AIVQC Production verifies the package before activation and keeps the previous recipe for rollback.

Stations can send selected images — for example only NOK or low-confidence results — through Server to an inbox in Trainer. That is the most valuable material for the next model version: the images the model got wrong.

Lines, stations, camera roles and triggers (presence sensor, PLC product sequence, encoder) are described once in settings and reused by every deployment.
A station is connected with a one-time code. Trainer never receives the station's credential — Server issues and stores it.
A new recipe is never switched in during an inspection in progress.

PLACE IN THE PLATFORM

Trainer prepares what Server later publishes.

A model version approved in Trainer becomes a package that AIVQC Server distributes only to selected stations. Trainer never connects directly to the production line — that separation between experimental work and the production environment is deliberate.

FAQ

AIVQC Trainer

Do I need machine-vision or machine-learning expertise to use Trainer?+

No. Automatic training chooses the model and parameters, improves them between attempts and picks the best result, while the Overview screen suggests the next step. What matters is knowing the process and the defects — you decide what counts as a defect. A manual mode with the full parameter set is available for advanced users.

Does Trainer need our own data, or do you help collect it?+

During the pilot we jointly define what images are needed and help plan how to collect them. In ongoing operation, production data is collected by your team under the agreed imaging conditions.

Is the number of training runs limited during the pilot?+

No. During the pilot you train as often as you need — every annotation fix or new batch of line images can go straight into the next automatic training run.

Does Trainer work without a graphics card?+

Yes. Annotation, evaluation and deployment do not need a GPU. Training can run on the CPU (slower, for experiments) or on a GPU training worker on the local network.

Do images and models leave the site?+

No. Projects, datasets and training stay on your computers, and exchange with the lines goes through the local AIVQC Server on the plant network.

Can I use a model trained outside AIVQC?+

Yes — an ONNX model with its class file can be imported, evaluated and deployed if it meets the AIVQC Production detection contract. It goes through the same acceptance criteria as a model trained in Trainer.

AIVQC PILOT

Try it on your own line.

Start with a 30-day pilot on one station and see the result on images of your own product before you decide on deployment.

Have a quality engineer and IT staff? Your own team can deploy AIVQC — you pay for a licence, not for integrator hours. Sign up for the beta →

1 Trainer · 1 station
Up to 3 defect classes
30 days + free extension
No cloud

Join the pilot →Deploying with your own team