README.md

gRPC Python Docker example

In this guide, we are going to build a simple gRPC client and server that take an image as input and return a negative and resized version of the image.

1. Dockerfile and dependencies

Nothing specific, just pip install the grpcio module for gRPC communication and the numpy + Pillow libraries to manipulate our image.

We are also going to use the pickle library (included in Python) to transform our numpy image, read by Pillow, into bytes.

FROM ubuntu:bionic

RUN apt-get update
RUN apt-get install python3 python3-pip -y

COPY ./requirements.txt /requirements.txt
RUN pip3 install -r /requirements.txt

Our requirements.txt file :

numpy==1.19.0
Pillow==8.1.1
grpcio==1.38.0
protobuf==3.17.1

2. Docker-compose file

To bind our local files to our container files and execute the program easily, let's create a docker-compose.yml file :

version: '3.3'

services:

    client:
        build: .
        command: python3 /usr/app/client.py
        volumes:
            - ./input:/usr/app/input    # Our input image directory
            - ./output:/usr/app/output  # Our output image directory
            - ./client.py:/usr/app/client.py:ro
            - ./grpc_compiled:/usr/app/grpc_compiled
        depends_on: 
            - server

    server:
        build: .
        command: python3 /usr/app/server.py
        volumes:
            - ./server.py:/usr/app/server.py:ro
            - ./grpc_compiled:/usr/app/grpc_compiled

3. Defining your proto file

gRPC works with .proto files to know which data to handle. Let's create a image_transform.proto file :

syntax = "proto3";

package flavienbwk;

service EncodeService {
    rpc GetEncode(sourceImage) returns (transformedImage) {}
}

// input
message sourceImage {
    bytes image = 1; // Our numpy image in bytes (serialized by pickle)
    int32 width = 2; // Width to which we want to resize our image
    int32 height = 3; // Height to which we want to resize our image
}

// output
message transformedImage {
    bytes image = 1; // Our negative resized image in bytes (serialized by pickle)
}

4. Compile your .proto file

.proto files must be compiled with the grpcio-tools library to generate two classes that will be used to perform the communication between our client and server

First, install the grpcio-tools library :

pip3 install grpcio-tools

And compile our image_transform.proto file with :

python3 -m grpc_tools.protoc -I. --python_out=./grpc_compiled --grpc_python_out=./grpc_compiled image_transform.proto

Files image_transform_pb2.py and image_transform_pb2_grpc.py files will appear in grpc_compiled/

5. Client

Our client.py file reads the image which becomes a numpy array and sends the query to the server along with the resize information. Then we save the image returned by the server in ./output/eiffel-tower-transformed.jpg.

from PIL import Image
import numpy as np
import pickle
import grpc
import sys

sys.path.append("/usr/app/grpc_compiled")
import image_transform_pb2
import image_transform_pb2_grpc

def run():
    channel = grpc.insecure_channel('server:13000')
    stub = image_transform_pb2_grpc.EncodeServiceStub(channel)
    image_np = np.array(Image.open('/usr/app/input/eiffel-tower.jpg'))
    image = Image.fromarray(image_np.astype('uint8')) # Transforming np array image into Pillow's Image class
    query = image_transform_pb2.sourceImage(
        image=pickle.dumps(image),
        width=320,
        height=180
    )
    response = stub.GetEncode(query)
    image_transformed = pickle.loads(response.image)
    image_transformed.save('/usr/app/output/eiffel-tower-transformed.jpg')

if __name__ == "__main__":
    run()

6. Server

Our server.py file receives the image, resize width and height. Then it returns a resized, negative-transformed image to the client.

from concurrent import futures
import numpy as np
import pickle
import grpc
import time
import sys

sys.path.append("/usr/app/grpc_compiled")
import image_transform_pb2
import image_transform_pb2_grpc

def image_to_negative(image: np.ndarray) -> np.ndarray:
    """Transforms a classic image into its negative"""
    negative = image.copy()
    for i in range(0, image.size[0]-1):
        for j in range(0, image.size[1]-1):
            pixelColorVals = image.getpixel((i,j))
            redPixel    = 255 - pixelColorVals[0] # Negate red pixel
            greenPixel  = 255 - pixelColorVals[1] # Negate green pixel
            bluePixel   = 255 - pixelColorVals[2] # Negate blue pixel
            negative.putpixel((i,j),(redPixel, greenPixel, bluePixel))
    return negative

class EService(image_transform_pb2_grpc.EncodeServiceServicer):

    def GetEncode(self, request, context):
        print("Received job !")
        image = pickle.loads(request.image)
        image = image.resize((request.width, request.height))
        image_transformed = image_to_negative(image)
        return image_transform_pb2.transformedImage(image=pickle.dumps(image_transformed))

def serve():
    server = grpc.server(futures.ThreadPoolExecutor(max_workers=4))
    image_transform_pb2_grpc.add_EncodeServiceServicer_to_server(EService(),server)
    server.add_insecure_port('[::]:13000')
    server.start()
    print("Server started. Awaiting jobs...")
    try:
        while True: # since server.start() will not block, a sleep-loop is added to keep alive
            time.sleep(60*60*24)
    except KeyboardInterrupt:
        server.stop(0)

if __name__ == '__main__':
    serve()

7. Run !

So let's run docker-compose up !

You will see our original eiffel-tower.jpg image will transform into its negative and resized version eiffel-tower-transformed.jpg

eiffel-tower.jpg (640px / 360px)	eiffel-tower-transformed.jpg (320px / 180px)