Train ML models

Your machines capture images in the field. ML model training turns those images into models that can classify what a camera sees or detect specific objects in real time. Viam handles the training infrastructure so you can focus on your data and your use case.

The training workflow

Training an ML model follows a repeating cycle:

1. Create a dataset. Collect images from your machines into a named, organization-level collection. You can pull images from multiple machines across different locations into a single dataset.

2. Label your images. Tag images for classification (the whole image gets a label) or draw bounding boxes for object detection (each object in the image gets a labeled rectangle). You can label manually or use an existing model to auto-annotate and review predictions.

3. Train a model. Submit a training job and Viam runs it on cloud infrastructure. No GPU provisioning, no framework installation. When training completes, the model is stored in your organization’s registry.

4. Deploy to your machine. Configure the appropriate ML model service on your machine (for example, tflite_cpu for TFLite models). Add a vision service to apply the model to live camera frames. The machine pulls the model from the registry automatically.

5. Iterate. Deploy the model, collect data on its predictions, auto-annotate new images with the current model, review the predictions, retrain, and redeploy. Each cycle tightens the feedback loop and improves accuracy. In ML this is called active learning.

Scale labeling with auto-predictions

Once you have a working model, you do not have to label every new image by hand. Use auto-predictions to have an existing model draft tags or bounding boxes for a dataset, then review the suggestions.

A typical active-learning loop:

1. Hand-label a starter dataset (20-50 images per class).
2. Train an initial model.
3. Capture more images from your machines.
4. Run auto-predictions against the new images using your initial model.
5. Accept or reject each prediction to produce verified labels.
6. Retrain with the expanded dataset.

This turns a small labeling effort into a growing, self-improving dataset.

Supported frameworks and hardware

Viam’s managed training handles TFLite and TensorFlow directly. For PyTorch, ONNX, or other frameworks, write a custom training script.

TFLite produces compact models optimized for edge devices without a GPU, and is the right choice for most Viam use cases. TensorFlow produces larger models that require more compute. Use PyTorch or ONNX when you are importing an existing model or need framework-specific features not available in TensorFlow.

Custom training scripts currently run in TensorFlow-based containers. You can install additional Python dependencies through setup.py. See custom training scripts for details.

Task types

The task type determines what the model learns to do. It must match how you labeled your dataset.

• Single label classification assigns exactly one label to each image. Your dataset should have tags with exactly one tag per image.
• Multi label classification assigns one or more labels to each image. Your dataset should have tags, and images may have multiple tags.
• Object detection finds objects in an image and draws bounding boxes around them with labels. Your dataset should have bounding box annotations.

Automatic deployment

When training completes, the model is stored in your organization’s registry. If a machine is already configured to use that model, Viam automatically deploys the new version. The machine pulls the latest version the next time it checks for updates. No manual download or restart is needed.

Dataset versioning

Each time you train a model from a dataset, Viam snapshots the dataset’s contents at that point. You can continue adding images and labels to a dataset after training without affecting previously trained models.