[PYTHON] Create an API that returns data from a model using turicreate

Overview

There is an Apple machine learning library called turicreate. (It's machine learning models, so that should be fine)

A memo when I made a configuration to handle a model using this on GCP Reference: https://cloud.google.com/blog/products/ai-machine-learning/how-to-serve-deep-learning-models-using-tensorflow-2-0-with-cloud-functions

Constitution

Since there was no icon in the part where Cloud Run is written, the GAE icon is used instead.

Untitled (1).png

The user accesses GAE, and after that Cloud Run downloads the model from storage and works. Of course, it doesn't make sense to bring it from storage every time, so the loaded model is cached in memory.

Trouble + solution

In the reference configuration, Tensorflow model is loaded by Cloud Functions and returned as API, but in the case of Turicreate, probably it will not be possible to build and an error will occur when trying to resolve the dependency. (It's the same with GAE, it may be good to change the Cloud Build settings, but I didn't know if it would be touched by GAE or Cloud Functions)

Therefore, I decided to use Cloud Run, which allows you to create, deploy, and run your own execution image.

Implementation

The implementation itself is nothing, almost following the configuration of python + CloudRun

`main.py`


import turicreate
import os
from flask import Flask, request, jsonify
from google.cloud import storage
import zipfile

model = None

app = Flask(__name__)


def download_model(bucket_name, source_blob_name, dest_blob_name):
    storage_client = storage.Client()
    bucket = storage_client.get_bucket(bucket_name)
    blob = bucket.blob(source_blob_name)

    blob.download_to_filename(dest_blob_name)
    with zipfile.ZipFile(dest_blob_name) as modelZip:
        modelZip.extractall('.')


@app.route('/')
def root():
    global model
    request_json = request.get_json()

    if request.args and 'userId' in request.args:
        userId = request.args.get('userId')
    else:
        return jsonify({'message': 'userId is not found'}), 400
    if 'limit' in request.args:
        limit = int(request.args.get('limit'))
    else:
        limit = 10

    if model is None:
        load_model()

    result = model.recommend(users=[userId], k=limit)

    random.shuffle(result)
    return jsonify({'result': result})

def load_model():
    global model

    os.environ['GOOGLE_APPLICATION_CREDENTIALS'] = './credential.json'
    download_model("learning-data", "model.zip", "./model.zip")
    model = turicreate.load_model('./model')
    

if __name__ == '__main__':
    app.run(host='127.0.0.1', port=int(
        os.environ.get('PORT', 8080)), debug=False)

FROM python:3.7

ENV APP_HOME /app
WORKDIR $APP_HOME
COPY . .

RUN pip install -r requirements.txt

CMD [ "python", "main.py" ]

`requirements.txt`


turicreate==6.1
flask==1.1.1
gunicorn==20.0.4
google-cloud-storage==1.26.0

GAE is OK as long as you access the completed service

Task

In the above case, what happens if the instance continues to remain, I want to make it work by checking the timing of model update
How to update when the number of instances increases

→ I'm thinking about how to set the data in memory to key: time, value: model so that the model data is read asynchronously first at a separate timing.

This area may be helpful https://cloud.google.com/run/docs/tips?hl=ja