torch-gcp-fn

Deploy Image classification on GCP as cloud function.

About Cloud Function

Cloud function has good response time and ability to scale. These can be configured accordingly using requirements.txt. For more understanding, this post describes how TF is used for deploying classification model.

App

This app uses imagenet pre-trained model resnet-18 available in torchvision to give image classification according to 1000 predefined categories in imagenet_class_index.json.

Setup

1. local machine

• clone this repo
• cd $REPO;zip -r torch_gcp_fn.zip .

This will make a zip file of this repo.

2. Cloud Function

• On GCP console, go to Cloud Functions and click Create Function.
• Fill name as desired, add memory 2GB, set Triger HTTP.
• Under Source code, select Zip upload.
• Select Runtime as Python3
• set functions to execute to handler
• Expand Environment variables, networking, timeouts and more
• set Timeout to max as 540
• click create

Response

• once the function is launched correctly, there will be green tick
• to start inference and testing, run following with updated values

import requests

function_url = "" # add url path to cloud function previously created
file_path = "" # add local image path 
resp = requests.post(function_url, files={"file":open(file_path,'rb')})
print(resp.json())

Note

In the requirements.txt, to install pytorch full path to wheel is given, instead if only torch was written it would have given errors. Common error encountered was code3, INVALID_ARGUMENT.