Tensorflow Serving gRPC Service

The complete example can be found at Tensorflow Serving gRPC - Truefoundry Examples

We will deploy a pre-trained MobileNet model from Tensorflow Hub

Project structure

Our final project structure will look like this

.
├── Dockerfile
├── client.py
├── german-shepherd.jpeg
├── deploy.py
└── models
    ├── imagenet_mobilenet_v3_small_075_224_classification_5
    │   └── 1
    │       ├── assets
    │       ├── saved_model.pb
    │       └── variables
    │           ├── variables.data-00000-of-00001
    │           └── variables.index
    └── models.config

Step 1: Download the model

mkdir -p models/imagenet_mobilenet_v3_small_075_224_classification_5/1
wget -O mobilenet_v3_small_075_224.tar.gz https://tfhub.dev/google/imagenet/mobilenet_v3_small_075_224/classification/5?tf-hub-format=compressed
tar -C models/imagenet_mobilenet_v3_small_075_224_classification_5/1/ -zxvf mobilenet_v3_small_075_224.tar.gz
rm mobilenet_v3_small_075_224.tar.gz

Create a models.config in models/

`models/models.config`

model_config_list {
  config {
    name: 'mobilenet-v3-small'
    base_path: '/mnt/models/imagenet_mobilenet_v3_small_075_224_classification_5'
    model_platform: 'tensorflow'
  }
}

Step 2: Dockerfile for Tensorflow serving

Create a Dockerfile with the following contents

`Dockerfile`

FROM tensorflow/serving:2.11.0
COPY models/ /mnt/models
COPY start_server.sh /app/start_server.sh
ENTRYPOINT ["/usr/bin/tensorflow_model_server", "--model_config_file=/mnt/models/models.config", "--port=9000", "--grpc_max_threads=8", "--enable_batching=true", "--enable_model_warmup=true"]

Step 3: Deploy it using Truefoundry

We can now deploy our service using Python SDK or YAML spec. Make sure you have servicefoundry installed and setup

Via Python SDK

Create a deploy.py

`deploy.py`

import argparse
import json
import logging
from servicefoundry import Service, Build, DockerFileBuild, Resources, Port, BasicAuthCreds, AppProtocol

logging.basicConfig(level=logging.INFO, format=logging.BASIC_FORMAT)

parser = argparse.ArgumentParser()
parser.add_argument(
    "--workspace_fqn",
    type=str,
    required=True,
    help="FQN of the workspace to deploy to",
)
args = parser.parse_args()

service = Service(
    name="mobilenet-v3-small-tf",
    image=Build(build_spec=DockerFileBuild()),
    resources=Resources(
        cpu_request=1,
        cpu_limit=1,
        memory_request=500,
        memory_limit=500,
    ),
    ports=[
        Port(
            port=9000,
            app_protocol=AppProtocol.grpc,
            # Note: Your cluster should allow subdomain based routing (*.yoursite.com) for gRPC to work correctly via public internet.
            # A host matching the wildcard base domain for the cluster can be explicitly configured by passing in `host`
        ),
    ],
)

service.deploy(workspace_fqn=args.workspace_fqn)

Now we can deploy by simply running this file with Workspace FQN

python deploy.py --workspace_fqn <YOUR WORKSPACE FQN HERE>

Via YAML Spec

Create a deploy.yaml with the following spec:

`deploy.yaml`

name: mobilenet-v3-small-tf
type: service
image:
  type: build
  build_spec:
    type: dockerfile
    dockerfile_path: './Dockerfile'
    build_context_path: './'
  build_source:
    type: local
ports:
  - port: 9000
    expose: true
    protocol: TCP
    app_protocol: grpc
replicas: 1
resources:
  cpu_limit: 0.5
  cpu_request: 0.2
  memory_limit: 500
  memory_request: 500

Now we can deploy by simply calling sfy deploy with Workspace FQN

sfy deploy --file deploy.yaml --workspace-fqn <YOUR WORKSPACE FQN HERE>

Step 4: Get the Endpoint URL

Once deployed, we can find the Service in the Deployments section of the platform and get the Endpoint

Step 5: Interacting with our Service

We can now interact with our service by passing the Endpoint from Step 6.

Install Tensorflow Serving Client

pip install numpy==1.22.4 tensorflow==2.11.0 tensorflow_serving_api==2.11.0

Download this image and save it as german-shepherd.jpeg

`german-shepherd.jpeg` — german-shepherd.jpeg

Finally, create a client.py that will send this image for prediction

`client.py`

import argparse
from google.protobuf.json_format import MessageToDict
import grpc
from tensorflow_serving.apis import predict_pb2, get_model_metadata_pb2, prediction_service_pb2_grpc
import tensorflow as tf
import numpy as np

parser = argparse.ArgumentParser()
parser.add_argument(
    "--host",
    type=str,
    required=True,
    help="Host of the deployed tf serving service",
)
args = parser.parse_args()

GRPC_MAX_RECEIVE_MESSAGE_LENGTH = 4096 * 4096 * 3  # Max LENGTH the GRPC should handle

channel = grpc.secure_channel(
        args.host,
        credentials=grpc.ssl_channel_credentials(),
)
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)

# Predict example
image = tf.io.read_file("german-shepherd.jpeg")
image = tf.io.decode_jpeg(image)
image = tf.image.convert_image_dtype(image, dtype=tf.int8)
image = tf.expand_dims(image, 0)

grpc_request = predict_pb2.PredictRequest()
grpc_request.model_spec.name = 'mobilenet-v3-small'
grpc_request.model_spec.signature_name = 'serving_default'
grpc_request.inputs['inputs'].CopyFrom(tf.make_tensor_proto(image, dtype=tf.float32))
predictions = stub.Predict(grpc_request, 10.0)

outputs_tensor_proto = predictions.outputs['logits']
shape = tf.TensorShape(outputs_tensor_proto.tensor_shape)
outputs = np.array(outputs_tensor_proto.float_val).reshape(shape.as_list())
outputs = tf.nn.softmax(outputs, axis=1)
top = tf.argmax(outputs, axis=1).numpy()[0]
print("Predicted Class:", top, outputs[0][top])

We can invoke our rpc by running the file

python client.py --host <YOUR ENDPOINT HERE>

E.g. If Endpoint were https://mobilenet-v3-small-tf-9000.abc.xyz.example.com

python client.py --host mobilenet-v3-small-tf-9000.abc.xyz.example.com

Note: Do not pass https:// in the --host option

This should get us output like

Predicted Class: 238