CHANGES.md

edgetpu release notes

This is a non-exhaustive summary of changes to the Edge TPU library, compiler, and runtime.

For pre-built downloads, see coral.ai/software.

Source code update, Edge TPU library 2.14.0 and compiler 2.1.302470888 (March 2020)

• C++ support for model pipelining
• Update Edge TPU python library to 2.14.0. New size parameter for run_inference
• Compiler update (new compiler options num_segments and intermediate_tensors)
• EfficientNet embedding extractor available for on-device learning with backprop

Edge TPU runtime v13 and compiler 2.0.291256449 (January 2020)

• Bug fixes
• Added support for ops: ExpandDims, Pack, Quantize, Sum, and TransposeConv
• First runtime release made available for Mac and Windows (compiler is still Linux only)

Edge TPU library 2.13.0 (January 2020)

• Updated the build based on the latest TensorFlow source (no API changes)
• First release made available for Mac and Windows

Edge TPU library 2.12.2 (November 2019)

• Performance optimizations

Edge TPU library 2.12.1 (September 2019)

• Python API signature style changes.
• New read_label_file() utility function.

Edge TPU compiler 2.0.267685300 (September 2019)

• Improved support for models built with full integer post-training quantization—especially those built with the Keras API.
• Added support for the DeepLab v3 semantic segmentation model.
• Still compatible with Edge TPU runtime v12.

Edge TPU library 2.11.1 (July 2019)

• New SoftmaxRegression API that allows you to perform transfer learning with on-device backpropagation.
• Re-implementation of the ImprintingEngine API so you can keep the pre-trained classes from the provided model, retrain existing classes without starting from scratch, and immediately perform inference without exporting to a new .tflite file. For details, see Retrain a classification model on-device with weight imprinting.

Edge TPU compiler 2.0 and runtime v12 (July 2019)

• Added support in libedgetpu.so for the TensorFlow Lite delegate API, allowing you to perform inferences directly from the TensorFlow Lite Python API (instead of using the Edge TPU Python API). For instructions, see Run inference with TensorFlow Lite in Python.
• Added support for models built with full integer post-training quantization.
• Added support for new EfficientNet-EdgeTPU models.

Edge TPU compiler 1.0 (May 2019)

• New offline Edge TPU Compiler.

Edge TPU libray 1.9.2 (April 2019)

• New edgetpu.h C++ API to perform inferencing on the Edge TPU, using the TensorFlow Lite C++ API.
• You can now run multiple models with multiple Edge TPUs.
• New edgetpu.utils.image_processing APIs to process images before running an inference.