fpaupier/tensorflow-serving_sidecar

Serve machine learning models using tensorflow serving

Note: This repo is now archived and won't see further development

How to serve your machine learning model with tensorflow-serving

This repo offers detailed How-to about serving your tensorflow models with tensorflow-serving. It can be followed as a step-by-step tutorial to serve your tensorflow model or other models with some adjustments.

Scope of the project

I do not cover the parts of training or exporting models here. I focus solely on serving a model for inference.

What's in the box?

Two tutorials:

1. A basic tutorial to get you quickly serve an object detection model with tensorflow-serving. Expect to get such an awesome object-detection up and running in less than 10 mins.

2. An advanced tutorial to deploy your tensorflow server docker image on Google Cloud Platform. Unleash the power of GCP to build a scalable machine learning server running in a kubernetes cluster.

The proposed object detection model is here to get you started quickly, feel free to use yours for more fun!

The reader is expected to be familiar with tensorflow, knowing how to export a model for inference will be helpful (tensorflow-server works with savedModel). Knowing docker or kubernetes are not pre-requisites since the commands used are simple and explained when needed.

Those tutorials are highly inspired by tensorflow-serving official documentation with some tips and more detail on the installation process.

0. Install the project

The installation process is thoroughly described in the docs/setup.md. It covers everything you need to do prior being able to serve a model with tensorflow-serving.

1. Serve your first model and perform inference with tensorflow-serving

Play Time!

To make sure the installation went smoothly, get your first inference result from the object detection model tensorflow serving. Follow the first tutorial to serve an object detection model on your machine docs/tf_server_local.md.

2. Deploy your model on the cloud for high availability

Now that we made sure our inference server works great on local, let's deploy it on the cloud at production scale. Let's dive in the docs/tf_server_k8s.md.

---

Resources

Additional information you may find useful

Google proposes a managed solution - Google Cloud ML Engine - to serve your saved_models.pb models. I do not focus on it since google documentation is a more comprehensive source of information.

Pros

• Deploy new models easily

Cons

• Less flexibility
• As of today, you are limited in size for your savedModel.pb file to 250 MB. (That may change)

---

Credits

The object_detection directory comes from the tensorflow-model repository. It offers useful utils functions to tag the image returned from the model.

Feel free to investigate the models on the tensorflow-model repo since they are well documented and often comes with useful tutorials.

tensorflow-serving_sidecar's Issues

variables empty!

I have a question:

1.download faster_rcnn_resnet101_coco_2018_01_28 from model_zoo as you mentioned
2. rename dir name to 001
3.run TF Serving successfully!

But,why they need to rewrite exporter.py ?

Missing LICENSE file

Hi.

It would be helpful if this repo had a LICENSE file, so terms of reuse were clear. Instructions are here: https://docs.github.com/en/github/building-a-strong-community/adding-a-license-to-a-repository .

I think you have a cool example, and I'd like to share it with some folks here at Google, but we can't do anything with it without a license.

Thanks!

client file for videos

how to do tf-serving for video file ?

This is giving syntax error

tensorflow-serving_sidecar/client.py

Line 163 in 2680f63

print(f'\n\nPre-processing input file {image_path}...\n')

All the print statements starting with "f" is giving error with both python and python3. Do we know why?

cannot import name 'plot_util'

while running client.py script , i am getting error as - cannot import name 'plot_util'.
i do not see 'plot_util.py' file under object_detection/utils.

Let me know your views.
Thanks