K
Q

Question

OpenFaaS serve model using Tensorflow serving

2/18/2019

I'd like to serve Tensorfow Model by using OpenFaaS. Basically, I'd like to invoke the "serve" function in such a way that tensorflow serving is going to expose my model.

OpenFaaS is running correctly on Kubernetes and I am able to invoke functions via curl or from the UI.

I used the incubator-flask as example, but I keep receiving 502 Bad Gateway all the time.

The OpenFaaS project looks like the following

serve/
  - Dockerfile
stack.yaml

The inner Dockerfile is the following

FROM tensorflow/serving

RUN mkdir -p /home/app

RUN apt-get update \
    && apt-get install curl -yy

RUN echo "Pulling watchdog binary from Github." \
    && curl -sSLf https://github.com/openfaas-incubator/of-watchdog/releases/download/0.4.6/of-watchdog > /usr/bin/fwatchdog \
    && chmod +x /usr/bin/fwatchdog

WORKDIR /root/

# remove unecessery logs from S3
ENV TF_CPP_MIN_LOG_LEVEL=3

ENV AWS_ACCESS_KEY_ID=${AWS_ACCESS_KEY_ID}
ENV AWS_SECRET_ACCESS_KEY=${AWS_SECRET_ACCESS_KEY}
ENV AWS_REGION=${AWS_REGION}
ENV S3_ENDPOINT=${S3_ENDPOINT} 

ENV fprocess="tensorflow_model_server --rest_api_port=8501 \
    --model_name=${MODEL_NAME} \
    --model_base_path=${MODEL_BASE_PATH}"

# Set to true to see request in function logs
ENV write_debug="true"
ENV cgi_headers="true"
ENV mode="http"
ENV upstream_url="http://127.0.0.1:8501"

# gRPC tensorflow serving
# EXPOSE 8500

# REST tensorflow serving
# EXPOSE 8501

RUN touch /tmp/.lock
HEALTHCHECK --interval=5s CMD [ -e /tmp/.lock ] || exit 1

CMD [ "fwatchdog" ]

the stack.yaml file looks like the following

provider:
  name: faas
  gateway: https://gateway-url:8080

functions:
  serve:
    lang: dockerfile
    handler: ./serve
    image: repo/serve-model:latest
    imagePullPolicy: always

I build the image with faas-cli build -f stack.yaml and then I push it to my docker registry with faas-cli push -f stack.yaml.

When I execute faas-cli deploy -f stack.yaml -e AWS_ACCESS_KEY_ID=... I get Accepted 202 and it appears correctly among my functions. Now, I want to invoke the tensorflow serving on the model I specified in my ENV.

The way I try to make it work is to use curl in this way

curl -d '{"inputs": ["1.0", "2.0", "5.0"]}' -X POST https://gateway-url:8080/function/deploy-model/v1/models/mnist:predict

but I always obtain 502 Bad Gateway.

Does anybody have experience with OpenFaaS and Tensorflow Serving? Thanks in advance

P.S.

If I run tensorflow serving without of-watchdog (basically without the openfaas stuff), the model is served correctly.

-- spaghettifunk

docker

kubernetes

openfaas

tensorflow

tensorflow-serving

1 Answer

5/17/2019

Elaborating the link mentioned by @viveksyngh.

tensorflow-serving-openfaas:

Example of packaging TensorFlow Serving with OpenFaaS to be deployed and managed through OpenFaaS with auto-scaling, scale-from-zero and a sane configuration for Kubernetes.

This example was adapted from: https://www.tensorflow.org/serving

Pre-reqs:

OpenFaaS

OpenFaaS CLI

Docker

Instructions:

Clone the repo

$ mkdir -p ~/dev/

$ cd ~/dev/

$ git clone https://github.com/alexellis/tensorflow-serving-openfaas

Clone the sample model and copy it to the function's build context

$ cd ~/dev/tensorflow-serving-openfaas

$ git clone https://github.com/tensorflow/serving

$ cp -r serving/tensorflow_serving/servables/tensorflow/testdata/saved_model_half_plus_two_cpu ./ts-serve/saved_model_half_plus_two_cpu

Edit the Docker Hub username

You need to edit the stack.yml file and replace alexellis2 with your Docker Hub account.

Build the function image

$  faas-cli build

You should now have a Docker image in your local library which you can deploy to a cluster with faas-cli up

Test the function locally

All OpenFaaS images can be run stand-alone without OpenFaaS installed, let's do a quick test, but replace alexellis2 with your own name.

$ docker run -p 8081:8080 -ti alexellis2/ts-serve:latest

Now in another terminal:

$ curl -d '{"instances": [1.0, 2.0, 5.0]}' \
   -X POST http://127.0.0.1:8081/v1/models/half_plus_two:predict

{
    "predictions": [2.5, 3.0, 4.5
    ]
}

From here you can run faas-cli up and then invoke your function from the OpenFaaS UI, CLI or REST API.

$ export OPENFAAS_URL=http://127.0.0.1:8080

$ curl -d '{"instances": [1.0, 2.0, 5.0]}' $OPENFAAS_URL/function/ts-serve/v1/models/half_plus_two:predict

{
    "predictions": [2.5, 3.0, 4.5
    ]
}

-- RakTheGeek

Source: StackOverflow

KQ

OpenFaaS serve model using Tensorflow serving

Similar Questions

1 Answer

K
Q