Release Notes 5.10

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

New Features and Enhancements

Validate Models on Desktop or Target Using the Validation Tab

Model Converter now supports model validation through the Validation tab. You can validate a model against the original model outputs or against reference outputs provided in the Code Generation tab.

Model Converter supports two validation modes: Target Validation and Desktop Validation. Use Target Validation to stream validation data to a connected target device over UART. This mode supports one quantization configuration at a time. Use Desktop Validation to run validation on the desktop using QEMU for the selected quantization configuration, or compare results across all supported quantization configurations. To know more, refer to Model Validation.

Explore the model using the Model Viewer feature of model converter

The Model Converter integrates the open-source Netron network visualizer, allowing you to inspect model layers and weights to better visualize, analyze, and understand model architecture. Using Model Viewer, you can inspect your model at different stages of optimization, including which operators are accelerated by the U55 NPU and how the model is converted from one framework to another, such as .h5, .keras, or .pt2 to .tflite. To know more, refer to Model Viewer.

Generated File Naming

The generated code file names no longer include the quantization type. For example, model_float.c/h and model_int8x8.c/h are now generated as model.c/h.

Multiple Input and Output Model support for .h5, .keras and .tflite

• Generate code for models with multiple inputs and outputs
• Use the NPZ data format for multiple input and output models
• Deploy multiple input and output models without modifying the original model architecture

Int16x8 Quantization is supported for Pytorch Models

You can quantize PyTorch models exported as .pt2 to INT16×8 by providing representative input samples for calibration. Achieve higher accuracy with a small increase in memory usage and inference time for models that benefit from larger activation data types.

Tensorflow .keras model support

DEEPCRAFT™ Model Converter now supports models built with TensorFlow 2 and Keras 3 through the .keras model file format.

Support for random calibration and validation data generation

The support for random calibration and validation data generation lets you deploy models and measure on-device inference time without preparing a dataset.

Generate code for PSOC™ Edge M55

You can generate code to deploy models on PSOC™ Edge M55 standalone. Models can be deployed to PSOC Edge E81 and E82 without the U55 NPU.

Fixes

• Resolved an issue that prevented the CLI tool from running in environments without a display adapter, such as Docker containers.
• Resolved an issue in the Use Provided Reference validation mode that caused incorrect validation results due to improper clipping of quantized data.
• Other small bug fixes and improvements