Release 2.6.0

What's Changed

Breaking Change: TPC Settings (#1635)

• Changed the existing get_target_platform_capabilities API (Breaking Change). Read more.

tpc = mct.get_target_platform_capabilities(tpc_version='1.0', 
                                           device_type='imx500')

• Added a new get_target_platform_capabilities_sdsp API to get TPC based on SDSP Converter version. Read more.

tpc = mct.get_target_platform_capabilities_sdsp(sdsp_version='3.14')

• Migrated features from the EdgeMDT-TPC repository.

Added TPC v6.0

• Added new TPC version with support for torch.take and torch.index_select layers. (#1570, #1591)

New API : MCTWrapper (#1578)

• Developed a new API that allows you to set parameters, perform quantization, and export model all at once. Read more.

wrapper = mct.MCTWrapper()
_, quantized_model = wrapper.quantize_and_export(float_model=float_model,
                                                 representative_dataset=representative_dataset,
                                                 method='PTQ',
                                                 framework='pytorch',
                                                 use_mixed_precision=False)

• Added tutorials of MCTWrapper. Try it!

Log Output Correction

• Modification of log output functionality using the logging module. (#1610)

Bug Fixes

• Modified weight quantization for Sin, Cos, and Exp layers in TPC v5.0 and v6.0. (#1647)

Tutorials

We've added new tutorials for various task. Try it!

• PyTorch Tutorials

  • YOLOX for Object Detection (#1558)
  • DeepLabv3+ for Semantic Segmentation (#1563)

• Keras Tutorials

  • EfficientDet for Object Detection (#1596)