docker-actdr

A dockerfile for a container that provides a simple HTTP grading interface

LICENSE: https://creativecommons.org/licenses/by-sa/4.0/legalcode CC-4.0-SA

Author(s): Tristan Swedish

Pre-Installation Steps

Make sure you've installed docker. See here for help Install Docker Community Edition. This should install everything so the docker command line tool is available on your system.

Installation

To get running right away, use the pre-built image (command line tested on macos 10.13, but should work on any 'nix):

docker pull tswedish/actdr

Installation done!

Getting Started

docker run -p 8080:8080 -d tswedish/actdr
export ACTDR_IP=localhost

Note: we've bound the port 8080 to our host machine (localhost) with the -p flag. This can be bound to a different local port if you don't want to use 8080!

Check it's status with stats:

docker stats

Note: The hash next to the running container is used to stop it, below we call this CONTAINER_HASH.

Once the docker container is up and running we can send it an image using curl:

export ACTDR_INPUT_IMG=input_image.png
curl -F file=@${ACTDR_INPUT_IMG} ${ACTDR_IP}:8080/grade

We can send png and jpg image, with any filename, specified as ${INPUT_IMG} above to the container at ip ${ACTDR_IP}. You should receive a JSON response with the prediction. If "success: false" in the returned JSON, something bad happened to the model, check that there is nothing wrong with the input file.

We could also send the request using python:

Here is an example of using the python requests library to make a post request to the service. you may need to pip install requests. We access the environment variable set above, except this time we relay an example "healthy" image from wikipedia. The following can be pasted into the python interpretor or run as a separate file.

import requests
import json
from io import BytesIO
import os
url = "http://"+os.environ["ACTDR_IP"]+":8080/grade"
norm_eye_url = "https://upload.wikimedia.org/wikipedia/commons/3/37/Fundus_photograph_of_normal_right_eye.jpg?download"
norm_jpg = BytesIO(requests.get(norm_eye_url).content)
pred = json.loads(requests.post(url,files={"file": ("norm.jpg", norm_jpg)}).content)
prediction = float(pred["grade_result"]["prediction"]) if pred["message"]=="OK" else "Undetermined"
std_prediction = float(pred["grade_result"]["prediction_std"]) if pred["message"]=="OK" else "Undetermined"
print("\n  Retina health score [0-4, 0 is healthy] (std deviation): %0.02f (%0.02f)\n" % (prediction, std_prediction))

This should print something like this:

Retina health score [0-4, 0 is healthy] (std deviation): 0.01 (0.07)

Which indicates the "healthy retina" image example on Wikipedia is indeed healthy!

Kill the docker container:

docker stop CONTAINER_HASH

Building and Installing directly

To build, run the following in the repo:

docker build -t actdr .

Notes

Docker seemed to change the way ports were exposed sometime around when I was developing this. I decided to leave the use of the -p flag above, which should work fine under most circumstances (since binding to the host computer simplifies the interface).

Acknowledgements

Thanks to the wonderful tutorial by aimbrain. This repo borrows heavily from actDR with further integration/overlap planned for the future. actDR was used to generate a production model made available on zenmodo. Please cite this repository if you use this work in academic work.