Skip to content

Migrating to SLEAP 1.5

SLEAP 1.5 represents a major milestone with significant architectural improvements, performance enhancements, and new installation methods. Here are the key changes:

Major Changes

UV-Based Installation

SLEAP 1.5+ now uses uv for installation, making it much faster than previous methods. Get up and running in seconds with our streamlined installation process.

PyTorch Backend

Neural network backend switched from TensorFlow to PyTorch, providing:

  • Much faster training and inference speeds
  • Modern deep learning capabilities
  • Improved developer experience
  • Multi-GPU training

Standalone Libraries

SLEAP GUI is now supported by two new packages for modular workflows:

SLEAP-IO

I/O backend for handling labels, processing .slp files, and data manipulation. Essential for any SLEAP workflow and can be used independently for data processing tasks.

SLEAP-NN

PyTorch-based neural network backend for training and inference. Perfect for custom training pipelines, remote processing, and headless server deployments.

Torch Backend Changes

New Backbones

SLEAP 1.5 introduces three powerful new backbone architectures (check here for more details):

  • UNet - Classic encoder-decoder architecture for precise pose estimation
  • SwinT - Swin Transformer for state-of-the-art performance
  • ConvNeXt - Modern convolutional architecture with improved efficiency

Legacy Support

We've maintained full backward compatibility:

  • GUI Support: SLEAP now uses a new YAML-based config file structure, but you can still upload and work with old SLEAP JSON files in the GUI. For details on converting legacy SLEAP 1.4 config/JSON files to the new YAML format, see our conversion guide.
  • Using TensorFlow Model Weights: Continue to support running inference on SLEAP <1.4 TensorFlow model weights (UNet backbone only). Check using legacy models for more details.

What's New in v1.6

SLEAP 1.6 builds on the v1.5 foundation with additional features and improvements:

Unified CLI

The new sleap command provides a single entry point for all SLEAP functionality:

sleap                    # Launch the GUI
sleap doctor             # System diagnostics
sleap train ...          # Training (via sleap-nn)
sleap predict ...        # Inference (via sleap-nn)
sleap show labels.slp    # View labels summary (via sleap-io)

See Command Line Interfaces for the full list of commands.

Label Quality Control

New Analyze → Label QC... menu for automated detection of labeling errors including: - Temporal jitter (unstable predictions across frames) - Visibility inconsistencies - Scale anomalies - Potential identity swaps

See Label Quality Control for details.

Instance Size Distribution

New Analyze → Instance Size Distribution... widget helps determine optimal crop sizes for top-down models by analyzing the distribution of instance bounding box sizes in your labeled data.

See Instance Size Distribution for details.

ONNX/TensorRT Export

Export trained models to ONNX or TensorRT format for 3-6x faster inference:

sleap export-model models/my_model --format onnx

Additional Improvements

  • Real-time inference progress: Live FPS and ETA display during inference
  • Filter overlapping instances: New controls in training/inference dialogs
  • Crop size visualization: Overlay showing crop region for top-down models

For a complete list of changes, see our Changelog.