TensorFlow Serving on ARM

TensorFlow Serving cross-compile project targeting linux on common arm cores from a linux amd64 / x86_64 build host.

Contents

• Overview
• Docker Images
• Build From Source
• Legacy Builds
• Disclosures
• Disclaimer

Overview

The basis of this project is to provide an alternative build strategy for tensorflow/serving with the intention of making it relatively easy to cross-build CPU optimized model server docker images targeting common linux arm platforms. Additonally, a set of docker image build targets is maintained and built for some of the popular linux arm platforms and hosted on Docker Hub.

Upstream Project: tensorflow/serving

Docker Images

Hosted on Docker Hub: emacski/tensorflow-serving

Usage Documentation: TensorFlow Serving with Docker

Note: The project images are desinged to be functionally equivalent to their upstream counter part.

Quick Start

On many consumer / developer 64-bit and 32-bit arm platforms you can simply:

docker pull emacski/tensorflow-serving:latest
# or
docker pull emacski/tensorflow-serving:2.6.0

Refer to TensorFlow Serving with Docker for configuration and setting up a model for serving.

Images

emacski/tensorflow-serving:[Tag]

Example

# on beaglebone black
docker pull emacski/tensorflow-serving:2.6.0-linux_arm_armv7-a_neon_vfpv3

Aliases

emacski/tensorflow-serving:[Alias]

Examples

# on Raspberry PI 3 B+
docker pull emacski/tensorflow-serving:2.6.0-linux_arm64
# or
docker pull emacski/tensorflow-serving:latest-linux_arm64

Manifest Lists

emacski/tensorflow-serving:latest

Examples

# on Raspberry PI 3 B+
docker pull emacski/tensorflow-serving
# or
docker pull emacski/tensorflow-serving:latest
# the actual image used is emacski/tensorflow-serving:latest-linux_arm64
# itself actually being emacski/tensorflow-serving:[Latest-Version]-linux_arm64_armv8-a

emacski/tensorflow-serving:[Version]

Example

# on Raspberry PI 3 B+
docker pull emacski/tensorflow-serving:2.6.0
# the actual image used is emacski/tensorflow-serving:2.6.0-linux_arm64
# itself actually being emacski/tensorflow-serving:2.6.0-linux_arm64_armv8-a

Debug Images

As of version 2.0.0, debug images are also built and published to docker hub. These images are identical to the non-debug images with the addition of busybox utils. The utils are located at /busybox/bin which is also included in the image's system PATH.

For any image above, add debug after the [Version] and before the platform suffix (if one is required) in the image tag.

# multi-arch
docker pull emacski/tensorflow-serving:2.6.0-debug
# specific image
docker pull emacski/tensorflow-serving:2.6.0-debug-linux_arm64_armv8-a
# specific alias
docker pull emacski/tensorflow-serving:latest-debug-linux_arm64

Example Usage

# start a new container with an interactive ash (busybox) shell
docker run -ti --entrypoint /busybox/bin/sh emacski/tensorflow-serving:latest-debug-linux_arm64
# with an interactive dash (system) shell
docker run -ti --entrypoint sh emacski/tensorflow-serving:latest-debug-linux_arm64
# start an interactive ash shell in a running debug container
docker exec -ti my_running_container /busybox/bin/sh

Back to Top

Build From Source

Build / Development Environment

Build Host Platform: linux_amd64 (x86_64)

Build Host Requirements:

• git
• docker

For each version / release, a self contained build environment devel image is created and published. This image contains all necessary tools and dependencies required for building project artifacts.

git clone [email protected]:emacski/tensorflow-serving-arm.git
cd tensorflow-serving-arm

# pull devel
docker pull emacski/tensorflow-serving:latest-devel
# or build devel
docker build -t emacski/tensorflow-serving:latest-devel -f tensorflow_model_server/tools/docker/Dockerfile .

All of the build examples assume that the commands are executed within the devel container:

# interactive shell
docker run --rm -ti \
    -w /code -v $PWD:/code \
    -v /var/run/docker.sock:/var/run/docker.sock \
    emacski/tensorflow-serving:latest-devel /bin/bash
# or
# non-interactive
docker run --rm \
    -w /code -v $PWD:/code \
    -v /var/run/docker.sock:/var/run/docker.sock \
    emacski/tensorflow-serving:latest-devel [example_command]

Config Groups

The following bazel config groups represent the build options used for each target platform (found in .bazelrc). These config groups should be treated as mutually exclusive with each other and only one should be specified in a build command as a --config option.

Build Project Image Target

//tensorflow_model_server:project_image.tar

Build a project maintained model server docker image targeting one of the platforms specified by a project config group as listed above. The resulting image can be found as a tar file in bazel's output directory.

bazel build //tensorflow_model_server:project_image.tar --config=linux_arm64_armv8-a
# or
bazel build //tensorflow_model_server:project_image.tar --config=linux_arm_armv7-a_neon_vfpv4
# each build creates a docker loadable image tar in bazel's output dir

Load Project Image Target

//tensorflow_model_server:project_image

Same as above, but additionally bazel attempts to load the resulting image onto the host, making it immediately available to the host's docker.

Note: host docker must be available to the build container for final images to be available on the host automatically.

bazel run //tensorflow_model_server:project_image --config=linux_arm64_armv8-a
# or
bazel run //tensorflow_model_server:project_image --config=linux_arm_armv7-a_neon_vfpv4

Build Project Binary Target

//tensorflow_model_server

Build the model server binary targeting one of the platforms specified by a project config group as listed above.

Note: It's not recommended to use these binaries as standalone executables as they are built specifically to run in their respective containers, but they may work on debian 10 like systems.

bazel build //tensorflow_model_server --config=linux_arm64_armv8-a
# or
bazel build //tensorflow_model_server --config=linux_arm_armv7-a_neon_vfpv4

Build Image for Custom ARM Target

//tensorflow_model_server:custom_image.tar

Can be used to fine-tune builds for specific platforms. Use a "Base" type config group and custom compile options. For linux_arm64 and linux_arm options see: https://releases.llvm.org/10.0.0/tools/clang/docs/CrossCompilation.html

# building an image tuned for Cortex-A72
bazel build //tensorflow_model_server:custom_image.tar --config=linux_arm64 --copt=-mcpu=cortex-a72
# look for custom_image.tar in bazel's output directory

Back to Top

Legacy Builds

Legacy GitHub Tags (prefixed with v)

• v1.11.1
• v1.12.0
• v1.13.0
• v1.14.0

Note: a tag exists for both v1.14.0 and 1.14.0 as this was the current upstream tensorflow/serving version when this project was refactored

Legacy Docker Images

The following tensorflow serving versions were built using the legacy project structure and are still available on DockerHub.

• emacski/tensorflow-serving:[Version]-arm64v8
• emacski/tensorflow-serving:[Version]-arm32v7
• emacksi/tensorflow-serving:[Version]-arm32v7_vfpv3

Versions: 1.11.1, 1.12.0, 1.13.0, 1.14.0

Back to Top

Disclosures

This project uses llvm / clang toolchains for c++ cross-compiling. By default, the model server is statically linked to llvm's libc++. To dynamically link against gnu libstdc++, include the build option --config=gnulibcpp.

The base docker images used in this project come from another project I maintain called Discolix (distroless for arm).

Back to Top

Disclaimer

• Not an ARM expert
• Not a Bazel expert (but I know a little bit more now)
• Not a TensorFlow expert
• Personal project, so testing is minimal

Should any of those scare you, I recommend NOT using anything here. Additionally, any help to improve things is always appreciated.

Back to Top