GitHub  •  LinkedIn  •  X  •  Discord

Telekinesis Iris

Telekinesis Iris is a computer vision platform for training, exporting, and deploying object detection and instance segmentation models on COCO datasets.

It includes:

• Local RF-DETR and SAM3-LoRA model integrations
• Unified PyTorch training, validation, checkpointing, export, and inference
• ONNX deployment for RF-DETR and native PyTorch deployment for SAM3-LoRA
• Rerun visualization for predictions

Release Model

Telekinesis Iris is currently in active development (pre-1.0). APIs may evolve between minor releases. Install the latest package version for the newest features and fixes.

Requirements

SAM3-LoRA requires access to Meta's gated facebook/sam3 model:

1. Sign in to Hugging Face and request access on the SAM3 model page.

2. Create a Hugging Face user access token with read access.

3. Export the token before training, exporting, or running SAM3-LoRA.

   macOS/Linux:

   export HF_TOKEN="hf_your_token"
   

   Windows PowerShell:

   $env:HF_TOKEN = "hf_your_token"
   

   These commands set the token only for the current terminal session. Export it again after opening a new terminal.

The token is not required for RF-DETR-only workflows. Keep it out of source control, configuration files, and logs.

Installation

1. Install Miniconda.

2. Create a Python 3.11 environment:

   conda create -n telekinesis-iris python=3.11
   

3. Activate the environment:

   conda activate telekinesis-iris
   

4. Install PyTorch and TorchVision for CUDA 12.8, then install the package:

   pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
   

    pip install -e ".[rfdetr]"
   

Example

Train an RF-DETR segmentation model on a COCO dataset:

from telekinesis.iris.dataset import COCODataset
from telekinesis.iris.models import RFDETR, RFDETRConfig, RFDETRVariant
from telekinesis.iris.trainer import Trainer

model = RFDETR(
    config=RFDETRConfig(
        variant=RFDETRVariant.SEG_NANO,
        num_classes=3,
        pretrained=True,
    )
)
training_dataset = COCODataset(
    "dataset/train",
    transforms=model.train_transforms,
    include_masks=model.requires_masks,
)

trainer = Trainer(
    model=model,
    training_dataset=training_dataset,
    output_dir="results/seg-nano",
    epochs=10,
    batch_size=4,
)

trainer.train()

Resources

• Telekinesis Examples
• Telekinesis Documentation
• Development guide
• Changelog

Support

For issues and questions:

• Create an issue.
• Contact support@telekinesis.ai.
• Join the Telekinesis Discord.