Telekinesis Agentic Skill Library

Appearance

Iris: Computer Vision Model Training ​

SUMMARY

Iris (telekinesis-iris) is the computer vision model training package within the Telekinesis Agentic Skill Library. It provides a unified framework for preparing datasets, training and fine-tuning models, evaluating performance, and exporting trained models for downstream inference and deployment.

0:00 / 0:00
Install Telekinesis Skill Library
Set up your Python environment, configure your Telekinesis API key, and install the Skill Library before getting started with Iris.
Open installation →
Explore the Telekinesis Quickstart
Get familiar with the Skill Library through guided examples covering computer vision, perception, and robotics workflows.
Open quickstart →

Iris Training Workflow ​

Iris standardizes the model training lifecycle across supported architectures. Prepare a COCO dataset, configure a trainable model, execute training and evaluation, and export the selected checkpoint for downstream inference and deployment.

1 · Prepare data
COCODataset
loads images, boxes, class labels, and optional masks from COCO annotations
↓
2 · Configure the model
TrainableModel
provides the model, preprocessing, loss, evaluation, and export behavior
↓
3 · Train and evaluate
Trainer
runs training and validation, reports metrics, and saves resumable checkpoints
↓
4 · Export
Export a Model
converts the selected checkpoint into a deployment-ready artifact
↓
5 · Deploy
Run Inference
loads the exported artifact behind a model-family-neutral inference API
↓
6 · Predict
Run Inference
returns boxes, scores, class IDs, optional class names, and optional masks

When to Use Iris? ​

Use Iris when you need to train or fine-tune computer vision models on custom datasets for application-specific tasks.

Typical use cases include:

  • Object detection: Train models to recognize and localize custom object classes.

  • Instance segmentation: Train models to identify individual objects and generate pixel-level masks.

  • Domain adaptation: Fine-tune pretrained models for new objects, environments, and imaging conditions.

  • Model evaluation: Track training performance and evaluate models using standardized metrics.

  • Model export: Prepare trained models for inference and integration into downstream applications.

What Does Iris Provide? ​

Iris provides a shared dataset, training, evaluation, checkpointing, export, and inference workflow.

Supported Models ​

Iris currently supports the following model families:

ModelTraining TasksExport Format
YOLOXObject DetectionONNX (.onnx)
RF-DETRObject Detection · Instance SegmentationONNX (.onnx)
SAM3-LoRAPrompt-driven Instance Segmentation · LoRA Fine-tuningIris Bundle (.pt)

Guides ​

Prepare a Dataset

Load COCO detection and segmentation datasets, including Roboflow-style splits.

Prepare data →

Train RFDETR

Train an RF-DETR detection or instance-segmentation model and resume from checkpoints.

Start training →

Train YOLOX

Train a fast object detector with Mosaic, MixUp, EMA, and multiscale training.

Start training →

Train SAM3-LoRA

Fine-tune SAM3 for prompted instance segmentation with compact LoRA weights.

Start training →

Export a Model

Export RFDETR, YOLOX, or SAM3-LoRA checkpoints as inference-ready artifacts.

Export a model →

Configuration

Reference the shared trainer, model, export, and post-processing settings.

View options →