DeepLTK New Version Released v9.0

DeepLTK v9.0.2 brings a major upgrade in capability and performance. This release adds new layer types, smarter data handling, and substantial optimizations that boost both training and inference speeds - achieving up to 2x faster execution across a wide range of models.

Note: This version introduces structural changes that might break backward compatibility with previous releases in some cases.

New Features

1. New Layers and Activation Functions DeepLTK now includes Layer Normalization and Scale layers, along with the GELU activation function – a modern choice used in architectures such as ConvNeXt and Transformer-based networks. These allow building next-generation deep learning models in LabVIEW.

   

2. Dry Run Mode for Resource-Free Evaluation

   A new Dry Run feature allows users to compute model metrics and verify configurations without allocating CPU or GPU resources.

3. 

   Ref-Based Layer Architecture

   Layers are now managed as references (Refnums) instead of Data Value References (DVRs), which significantly reduces overhead and improves latency. This might break backward compatibility if NN.ctl and NN_Layer.ctl clusters are passed through subVIs.

4. Smarter Data Handling for Custom Datasets

   Dataset types now include a variant element, enabling flexible and efficient handling of custom datasets. This improvement enhances performance and efficiency in anomaly detection tasks.

   
   

Performance Optimizations

1. 2x Faster Inference on GPUs: Particularly for deep networks with many layers.

2. Optimized GPU Training: Improved efficiency for large-scale models.

3. Accelerated ShortCut Layers: Reduced latency in residual network structures.

4. Thread Control API: Configure and optimize multithreading in CPU mode.

Together, these enhancements make DeepLTK more scalable and responsive for real-time AI applications in LabVIEW.

Usability and Developer Experience

1. New NN_Get_T_dT.vim API: Old NN_Get_T_dT.vi vi was replaced with its vim version to simplify benchmarking process.

2. Enhanced Debugging: Error messages now include layer name, type, and index for faster issue localization.

   
   

Stability and Bug Fixes

This release includes a wide range of stability improvements and memory management fixes, including:

1. More accurate metric computation and FLOP calculations.

2. Fixed GPU-specific activation and layer training issues.

3. Improved error propagation and YOLO layer validation.

4. Corrected documentation and refined example UIs.

Summary

Version 9.0.2 is a major step forward for DeepLTK – faster, more efficient, and ready for modern deep learning architectures. With new layers, improved control, and smarter data handling, it makes AI development in LabVIEW simpler and more powerful.

Explore the updated examples and tutorials on our GitHub repository.

For detailed information about all changes in this version please refer to DeepLTK’s release notes.