fuzzy-octo-disco

Small project of face recognition and image manipulation in Python using face_recognition and Pillow library. It is now used in the Dank Face Bot project, as a separate micro-service. The main use of this project is to identify faces on picture, by rotating it, and then return croped pictures of this faces.

Try it with

python example.py

Don't forget to install dependecies in a Python virtualenv first !

pip install -r requirements.txt

See my gist for more details on how to setup virtualenvwrapper with zsh.

Choice of architecture

These program need to communicate with other services, as dank-face-bot, which sends pictures, and expects some cropped faces back. But how can we transfer the pictures, an heavy load, between the services ?

Originally, the Picture class needed to use a file path to open the file. First, I thougth about sending the pictures through gRPC as bytes. But gRPC isn't made for large transfer, and I would have to change the class.

Therefore, I just send the path of the file and this find-faces service get the image how he can. On Kubernetes, this means I'll need a shared volume. On docker-compose, I use a shared volume.

We use a NFS shared volume : https://github.com/mappedinn/kubernetes-nfs-volume-on-gke

It's deployed with the main DFB repo.

Test

Use the pytest library. Just run

pytest

Add the -s arg to display print.

Deploy

See the main DFB repo for reminders about k8s, Travis and etc...

GRPC

Generate the proto

Install grpcio and grpcio-tools with pip. Then

python -m grpc_tools.protoc -I ./find_faces/grpc_service --python_out=./find_faces/grpc_service --grpc_python_out=./find_faces/grpc_service grpc/find-faces.proto

Fix path import issue for GRPC

There is an issue with the way protoc generate the pb files and __init__.py, which create a module.

@see protocolbuffers/protobuf#1491

Anyway, a fix for now : In find_faces/grpc_service/find_faces_pb2_grpc.py change

import find_faces_pb2 as find_faces__pb2

to

from . import find_faces_pb2 as find_faces__pb2

This isn't a great solution however...

Test GRPC

A small test file in find_face/grpc_service/client.py:

python -m find_faces.grpc_service.client

Use npm grpcc.

grpcc -p find_faces/grpc_service/find-faces.proto -a localhost:50051 -i

Then, in the REPL

client.findFaces({picture: {path: "./pictures/lena.png"}}, pr)