server.md

The backend server lives in ImageSearchServer.py and can be run with:

python3 ImageSearchServer.py

If you go to the server in a browser (i.e., send a GET request), you will just find a simple test page that lets you manually enter API parameters in a form and see the response. The search API is simple: send a POST to /imagesearch with the (urlencoded) body containing img_id and optionally max_results, both of which should be integers. You will get a JSON response that looks something like this:

{
    "images": [
        {
            "img_id": INTEGER,
            "diff": FLOATING_POINT,
        },
        ...
    ],
    "errstr": STRING
}

The errstr is used to give a human-readable error message if something goes wrong (mainly for debugging if the status code returned is not 200). The diff field indicates how different the image is from the image that was passed in the request, and the images array is sorted by this field (so most similar images are first). The img_id field will be an integer corresponding to an image ID in the database. If max_results was passed in the request, the server will not return more than that number of results in the images array.

More notes:

• The server runs on port 8000 by default. Pass the --port <port_number> option at startup to change it.
• The server will vectorize all the images it finds in Mysql when it starts up and search through those (but images added to Mysql after the server is already started will not be searched; that feature is forthcoming).
• The vectorizer used will be an autoencoder by default, which expects that a saved model is available (i.e., assumes you have already run train_model.py). Since this may make testing the server slow or troublesome if you don't have a good GPU for training/eval, you can pass --vectorizer-type "flat" to the server startup and it will use an alternative "dumb" vectorizer implementation that does not require a neural network.
• The images will be sourced from Mysql by default, which is probably what you want. If you want to directly load images from the filesystem instead, you can pass --image-source-type "filesystem" to the server at startup, but be warned that the img_id field will be hard to interpret (IDs will be assigned to files on the fly during data loading).