Edge TPU Python API

This repository contains an easy-to-use Python API to work with Coral devices:

• Dev Board
• USB Accelerator

You can run inference and do transfer learning.

Edge TPU Runtime

Run scripts/runtime/install.sh to install Edge TPU runtime or scripts/runtime/uninstall.sh to uninstall it.

Edge TPU Python API

1. Run scripts/build_swig.sh to build SWIG-based native layer for different Linux architectures. Build is Docker-based, so you need to have it installed.

2. Run make wheel to generate Python library wheel and then pip3 install $(ls dist/*.whl) to install it

Native C++ code

All native code is inside src folder. You can build everything using make command which invokes Bazel internally.

For example, run make tests to build all C++ unit tests or make benchmarks to build all C++ benchmarks. To get the list of all available make targets run make help. All output goes to out directory.

Linux

On Linux you can compile natively or cross-compile for 32-bit and 64-bit ARM CPUs.

To compile natively you need to install at least the following packages:

sudo apt-get install -y build-essential \
                        libpython3-dev \
                        libusb-1.0-0-dev \

and to cross-compile:

sudo dpkg --add-architecture armhf
sudo apt-get install -y crossbuild-essential-armhf \
                        libpython3-dev:armhf \
                        libusb-1.0-0-dev:armhf

sudo dpkg --add-architecture arm64
sudo apt-get install -y crossbuild-essential-arm64 \
                        libpython3-dev:arm64 \
                        libusb-1.0-0-dev:arm64

Compilation or cross-compilation is done by setting CPU variable for make command:

make CPU=k8      tests  # Builds for x86_64 (default CPU value)
make CPU=armv7a  tests  # Builds for ARMv7-A, e.g. Pi 3 or Pi 4
make CPU=aarch64 tests  # Builds for ARMv8, e.g. Coral Dev Board

macOS

You need to install the following software:

1. Xcode from https://developer.apple.com/xcode/
2. Xcode Command Line Tools: xcode-select --install
3. Bazel for macOS from https://github.com/bazelbuild/bazel/releases
4. MacPorts from https://www.macports.org/install.php
5. Ports of python interpreter and numpy library: sudo port install python35 python36 python37 py35-numpy py36-numpy py37-numpy
6. Port of libusb library: sudo port install libusb

Right after that all normal make commands should work as usual. You can run make tests to compile all C++ unit tests natively on macOS.

Docker

Docker allows to avoid complicated environment setup and build binaries for Linux on other operating systems without complicated setup:

make DOCKER_IMAGE=debian:buster DOCKER_CPUS="k8 armv7a aarch64" DOCKER_TARGETS=tests docker-build
make DOCKER_IMAGE=ubuntu:18.04  DOCKER_CPUS="k8 armv7a aarch64" DOCKER_TARGETS=tests docker-build