September 2019 Updates

Small updates to our compiler and Python API

The Coral Team
September 25, 2019

These are incremental updates to our July release.

New Debian packages

We've improved our distribution process, making our software for all devices available with Debian packages (using apt-get).

If you're using a Dev Board, this doesn't affect you and you can get all updates like before, as follows:

sudo apt-get dist-upgrade

If you set up the USB Accelerator prior to this release, then you used a script called install.sh that manually installed our Python library and runtime library, which can cause conflicts with the new Debian-package installation. To transition to the new Debian packages, you must follow these steps:

# First uninstall the previous Python library and Edge TPU runtime
sudo pip uninstall edgetpu
# For the following, use the path corresponding to your system arch, either:
# x86_64-linux-gnu, arm-linux-gnueabihf, or aarch64-linux-gnu
ls -d /usr/lib/x86_64-linux-gnu/libedgetpu* | xargs sudo rm

# Now install the new Python library and Edge TPU runtime
sudo apt-get install python3-edgetpu libedgetpu1-std
# Or if you prefer the max operating frequency, install 'libedgetpu1-max'

From now on, you can upgrade these packages with apt-get.

Updated Edge TPU Compiler

We've released a minor update to the Edge TPU Compiler (version 2.0.2xx) with improved support for post-training quantization—especially those built with Keras—and support for the DeepLab v3 semantic segmentation model (try it with this example code).

You can get the new Edge TPU Compiler as follows:

sudo apt-get update

sudo apt-get upgrade edgetpu

Because this is a minor compiler update (from v2.0 to v2.0.2xx), the Edge TPU runtime remains the same (v12).

Updated Edge TPU Python library

We've updated the Edge TPU Python library to version 2.12.1.

This is mostly a cleanup of our API naming conventions, but in a backward-compatible way so existing code still works. And we added a new utility function, read_label_file(), to load your model labels.

To update, see the first section about improved Debian packages.

New GitHub repositories

We've released two new repositories on GitHub:

https://github.com/google-coral/edgetpu

This is the source code for the Edge TPU API (Python and C++), including related example code, test data, benchmarks, and more. We previously made this repository available coral.googlesource.com, but we are now moving this portion to GitHub. The coral.googlesource.com repos will remain active for Mendel source code only.
https://github.com/google-coral/tflite

This is a new collection of code examples that use the TensorFlow Lite APIs with Coral devices (instead of the Edge TPU APIs).

That's all for now. As always, please send us feedback at coral-support@google.com.