Release Notes 5.15

This section lists the new functionality, improvements and some of the major changes related to DEEPCRAFT™ Studio.

New Features and Enhancements

Dataset Node in Graph UX

We have introduced the Dataset node in Graph UX, a step toward a complete edge machine learning workflow in Graph UX. The Dataset node is a self-contained place to collect, import, label, and manage your data inside the graph, so your data is no longer scattered across folders and files.

Currently, collecting data means connecting a sensor node to a Tracks node, defining classes in a Predefined node, saving the sessions, and manually adding the data to a project. The Dataset node brings this whole process into a single node:

• Connect a sensor node to the Dataset node, and the data tracks are created automatically.
• Create label tracks, define classes, and label your data directly in the node.
• Import existing data into the same Dataset node.
• Send the data straight to a training project in DEEPCRAFT™ Studio when the data is ready.

The Dataset node also lays the groundwork for batch data processing. By connecting two Dataset nodes with processing nodes in between, you can clean or augment an entire dataset in one operation.

The Dataset node also comes with a rebuilt timeline that handles large volumes of data more efficiently. Because the timeline no longer renders every data point, zooming, scrolling, and navigating are smoother. You can also record data from high-bandwidth sensors, such as radar, without slowdowns or crashes. The improved timeline is available only in the Dataset node.

To know more, refer to Introduction to Dataset Node.

Train your object detection models using YOLO11n and YOLO26n architectures

You can now train object detection models on custom data in DEEPCRAFT™ Studio using the Ultralytics YOLO11n and YOLO26n architectures. YOLO11n balances accuracy, speed and efficiency. YOLO26n uses a lighter detection head and updated training methods that improve accuracy, especially on small objects. Its streamlined design also makes it well suited to efficient deployment. For step-by-step instructions on training an object detection model, refer to Training Object Detection Model.

Interpreterless mode available in in code generation

DEEPCRAFT™ Studio now offers Interpreterless mode that pre-interprets the model to generate the static C++ code that executes the neural network directly, removing the need for the TensorFlow Lite Micro interpreter at runtime. When this option is enabled, the runtime, memory planner and unused operator code, memory planner are not compiled into the final application, reducing the ROM footprint. This can provide memory savings for very small models (under 400 KB) that use a limited set of operator types. Refer to Optimization to know how to use the Interpreterless mode. This option is supported only on the PSOC 6 and PSOC Edge M33/NNLite targets.

Support for Int8x8 and Int16x8 quantization in code generation

You can now quantize models to Int8x8 or Int16x8 during code generation by providing representative input samples for calibration. Int8x8 minimizes inference time, memory, and bandwidth use, while Int16x8 offers higher accuracy for models that benefit from larger activation data types, with a small increase in memory usage and inference time. To know more, refer to Optimization.

Improvements

Local Model Training

Local Model Training now includes two new options in the training configuration: Use GPU for training and Imagifier Image.

You can train models on an NVIDIA GPU as well as on the CPU, which speeds up training for larger models. When you enable this option, training runs on the GPU using the CUDA-based container image. When you disable this option, training runs on the CPU using the CPU-based container image. You can now use a custom container image for training instead of the default image for your project type. By default, this option uses the standard image for your project type. To know more, refer to Local Model Training.

Labeling in Object Detection

You can now filter and sort the Image list to find the images you need faster. Click the Filter images button to filter by label, including Unlabeled, or by set (train, validation or test). Each active filter appears as a chip next to the search box, so you can see and remove filters at a glance. You can also sort the list by Name, Label Count or Set. Sorting by Label Count in ascending order lists unlabeled images first. When you filter or sort the list, the Next and Previous controls continue from the current image.

Reorganized Buttons in the Data and Training Tab

We have reorganized the Data and Training tabs to make common actions easier to find. The most-used actions now appear as buttons and related actions are grouped into dropdown menus.

• Data tab: The Prepare Data dropdown contains Edit Tracks, Merge Data, Augment Data and Generate Labels.
• Training tab: The Start Training dropdown let you select between Local or Cloud training. The Model Generation dropdown contains Generate Models, Create Custom Model and Import Architecture.

Fixes

Overall bug fixes and increased stability