Fast Style Transfer Service with S4TF and Vapor

Swift for TensorFlow model wrapped in a HTTP/REST service, with basic front-end included.

Tested with Ubuntu and Swift for TensorFlow v0.5-rc0.

Live demo: https://s4tf-style-transfer.herokuapp.com

REST API

API tries to be compatible with TF Serving Predict API.

Expected input image format is JPEG, output format is PNG (base64-encoded).

Request:

POST /v1/models/style-transfer:predict
{
    "instances": [{"b64": "ABC=="}]
}

Response:

200 OK
{
    "predictions": [{"b64": "ABC=="}]
}

Docker

Checkout repo, build image and run:

$ docker build --tag=s4tf-style-transfer .
$ docker run --rm -p 8080:80 s4tf-style-transfer

Open http://localhost:8080/