Skip to content

Latest commit

 

History

History
108 lines (82 loc) · 4.8 KB

install-using-docker.md

File metadata and controls

108 lines (82 loc) · 4.8 KB
Raw

Install TensorFlow Deploy using Docker

Depending on the use of TensorFlow Deploy, two installation scenarios will be covered.

Installation of TensorFlow Deploy communication tools

In order to be able to communicate with the service easily, you can install tensorflow-deploy-utils python package. It's recommended, but still optional.

Package can be installed directly in python

$ pip install --user tensorflow-deploy-utils

or by using virtualenv

$ python3 -m venv ~/tfd-env
$ source ~/tfd-env/bin/activate
$ pip install tensorflow-deploy-utils

TensorFlow Deploy as a model repository

First, let's assume we would like to use tfd only as a repository with no possibility of communication with TensorFlow Serving.

# Prepare an environment to run tfd
$ mkdir ~/tfdeploy
# Download docker's image
$ docker pull grupawp/tensorflow-deploy
# Run the image
$ docker run --name tfd -v ~/tfdeploy:/tfdeploy -p 9500:9500 grupawp/tensorflow-deploy --storage_filesystem_base_path=/tfdeploy

Entering those three easy commands gives us a ready-to-go repository that provides us with a possibility of storing and versioning models as well as modules created by using TensorFlow.

Test tfd:

# Ping tfd
$ curl 0:9500/ping

You can find more information about configuration in section Configure TensorFlow Deploy and REST API in Communicate using REST API.

TensorFlow Deploy as a deployment tool

However, we may sometimes want to use tfd full potential and pair it together with tfs.

Prepare enviroment and download needed files from the repository.

# Prepare an environment to run tfd
$ mkdir ~/tfdeploy

# Download empty configuration for TensorFlow Serving so that it can run without models (by default it requires a model to run and if it is missing one it results in a loop of errors impossible to stop - endpoint reload_config is not working)
$ wget -O ~/tfdeploy/empty.config https://github.com/grupawp/tensorflow-deploy/raw/master/demo/data/empty.config

# Download example model
$ wget https://github.com/grupawp/tensorflow-deploy/raw/master/demo/data/example_model.tar

# Create a dns file storing details of our tfs instance with its IP and GRPC communication port
# Make sure the tfs instance IP address is correct. It may differ depending on your Docker environment and currently running cointainers.
$ echo 'tfs-team1-project1 172.17.0.3:8500' > ~/tfdeploy/hosts

Install tfd and tfs.

# Download TensorFlow Deploy image
$ docker pull grupawp/tensorflow-deploy
# Run tfd. Optionaly, add -d to run container as a daemon.
$ docker run --name tfd -v ~/tfdeploy:/tfdeploy -p 9500:9500 grupawp/tensorflow-deploy --discovery=plaintext --discovery_plaintext_hosts_path=/tfdeploy/hosts --storage_filesystem_base_path=/tfdeploy

# Download TensorFlow Serving image
$ docker pull tensorflow/serving
# Run tfs. Optionaly, add -d to run container as a daemon.
$ docker run --name tfs-1 -v ~/tfdeploy/models:/models -p 8500:8500 -p 8501:8501 tensorflow/serving --model_config_file=/models/empty.config --allow_version_labels_for_unavailable_models

Test it by using utils

# Upload model
$ tfd_deploy_model --host localhost --team team1 --project project1 --name name1 --path example_model.tar
# List models
$ tfd_list_models --host localhost

or by using REST API

# Upload model
$ curl -X POST -H "Content-Type: multipart/form-data" -F "archive_data=@example_model.tar" 0:9500/v1/models/team1/project1/names/name1
# Reload models within a team and project
$ curl -X POST 0:9500/v1/models/team1/project1/reload
# Check model on tfs
$ curl 0:8501/v1/models/name1

TensorFlow Deploy for a mass production

Service discovery configuration

While running multiple projects with at least one TensorFlow Serving instance assigned to each of them, it is difficult to keep DNS text file in order. Especially, when we are using tools like Kubernetes with running tfs together with autoscaller on it. In such case, it is worth considering using assistance of an administrator who can configure DNS service on the hardware running TensorFlow Deploy so that each tfs instance will have a common suffix, e.g. tfd. Then we can use DNS maintenance options embedded in tfd

$ docker run --name tfd -v ~/tfdeploy:/tfdeploy -p 9500:9500 grupawp/tensorflow-deploy --discovery=dns --discovery_dns_service_suffix=tfd --storage_filesystem_base_path=/tfdeploy

Shared drive storage configuration

Shared drive storage of tfd and all instances of tfs may be an issue in a mass production environment. The simplest solution is to create a (NFS) network resource and mount it to all the instances.