Skip to Content
DEEPCRAFT™ Studio 5.15 has arrived. Read more →

Release Notes 6.1

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

New Features and Enhancements

Deploy PyTorch models using the ExecuTorch runtime

DEEPCRAFT™ Model Converter now supports the ExecuTorch runtime. You can now deploy PyTorch models on PSOC™ Edge using ExecuTorch as an alternative to TensorFlow Lite Micro. ExecuTorch is available in both the GUI and CLI versions of DEEPCRAFT™ Model Converter for code generation and validation, and the setting is saved in the project file (.byom).

ExecuTorch is available for the PSOC™ Edge M55+U55 only. For other targets, the Use ExecuTorch option is disabled.

You can use ExecuTorch in one of the following ways:

  • Convert a PyTorch (.pt2) model: Select a PyTorch (.pt2) model and enable the Use ExecuTorch checkbox in General Settings. Model Converter exports the model to ExecuTorch and generates the code. When the checkbox is disabled, Model Converter converts .pt2 models to TensorFlow Lite as in earlier versions.

  • Use a pre-exported ExecuTorch (.pte) model: Select an ExecuTorch (.pte) model. Model Converter automatically enables and locks the Use ExecuTorch checkbox, because the model is already exported.

Note the following limitations when using ExecuTorch:

  • Only Float and Int8x8 quantization types are supported. To use Int16x8 quantization, disable the Use ExecuTorch checkbox.
  • Pre-exported ExecuTorch (.pte) models cannot be re-quantized. Validation compares only the quantization already applied to the model.
  • Use ExecuTorch requires a PyTorch (.pt2) or ExecuTorch (.pte) model. Model Converter displays an error if you enable it with a TensorFlow Keras (.h5, .keras) or TensorFlow Lite (.tflite) model.

To know how ExecuTorch compares to TensorFlow Lite Micro, refer to Deployment runtimes: ExecuTorch and TensorFlow Lite Micro. Refer to Use ExecuTorch in Graphical User Interface and Command Line Interface for more details.

Select the Tensor Format for calibration and validation data

The new Tensor Format option specifies the layout of the calibration and validation data. By default, Model Converter uses channels first (NCHW) for ExecuTorch and channels last (NHWC) for TensorFlow Lite. Set the Tensor Format explicitly to reuse the same calibration and validation data for both runtimes. Model Converter transposes the data internally when the selected format differs from the format the runtime expects. To know more, refer to Advanced Data Options.

Anonymous usage and crash reporting

Model Converter now collects anonymous usage and crash data to help in identifying bugs and prioritizing improvements. When you launch Model Converter for the first time, the license agreement dialog displays a new Share anonymous usage and crash data checkbox. Enable the checkbox to share the data, or clear it if you do not want to share.

Model Converter collects the following data:

  • Application start and stop events.
  • Code generation and on-target validation runs, including start and finish, duration, and whether the run succeeded, failed, or cancelled.
  • The options used for a run, such as target device, target API, quantization type, sparsity, and ExecuTorch.
  • The model format only, for example tflite. File names, model architecture, and weights are never collected.
  • Crash details, including the exception type and call stack. Exception messages are not collected.
  • One active installation ping every 28 days per installation.

You can change this setting at any time. In the GUI, navigate to File> Preferences> Privacy and select or clear the Share Anonymous Usage Data option. In the CLI, run telemetry --on to enable data sharing, telemetry --off to disable it, or telemetry --status to view the current setting. The GUI and CLI share the same setting, and changes apply immediately without restarting the application.

The collected data cannot be linked to your identity, Imagimob Cloud account, or project data. Model Converter includes only a random session ID and does not write any data to disk.

Fixes

  • The Target API in Advanced Options now updates correctly when you load a project file.
  • Empty or random calibration and validation datasets no longer fail with an error stating that the path is not a .csv or .npz file. Model Converter now generates random samples as intended in both the GUI and the CLI.
  • For macOS:
    • Code generation and validation no longer fail with the error no such file or directory: 'brew'. Model Converter now finds Homebrew tools correctly.
    • When required Homebrew packages are missing, Model Converter now displays a message with the install command: brew install cairo snappy libjpeg gnutls libslirp colima.
    • Library import crashes on macOS 27 are now fixed.