photo-manager-classifier

This project provides a deep learning image auto-tagging classifier as a JSON REST API. It is intended to be used as part of this project.

It currently uses Caffe2 and the squeezenet pre-trained model. In time I intend to transfer and train from a dataset that has more relevant categories and better accuracy. This could even be user contributed eventually.

Setup

git clone [email protected]:damianmoore/photo-manager.git
cd photo-manager-classifier/

Edit the file docker-compose.yml to mount your directory of photos as the volume /photos then build and run.

docker-compose build
docker-compose up

It will take a while to build the Docker image as it is based off a large deep-learning image that is over 900MB to download. This can be reduced in future as we don't need all the included resources and trained models. It works for now as a means to get up and running.

Querying

Once the server is running you should be able to use curl or HTTPie to issue requests, passing a path to a file in the mounted volume directory.

curl:

curl -H "Content-Type: application/json" -X GET -d '{"path": "/photos/IMG_6085.jpg"}' http://localhost:8888/categories/

HTTPie:

http GET localhost:8888/categories/ path='/photos/IMG_6085.jpg'

Response

You should get a response (after a few seconds) like this:

{
    "categories": [
        [
            "acoustic guitar",
            0.4404420852661133
        ],
        [
            "rifle",
            0.07927235215902328
        ],
        [
            "bannister, banister, balustrade, balusters, handrail",
            0.05965603142976761
        ],
    ],
    "path": "/photos/IMG_6085.jpg",
    "version": 0
}