This repository contains work in progress towards Swift TensorFlow as a Service. The idea is similar to TFaaS with implementation of web server and TF components in Swift. The web framework is based on Vapor.
For TF model we use model produced by SwiftTFExample.
In order to start with Vapor you need to build its toolbox.
# NOTE: to build vapor please use Apple/Linux vanila swift toolchain
# do not use TF toolchain
# download vapor toolbox
git clone https://github.com/vapor/toolbox.git
cd toolbox
# optional you may check out particular branch/tag
git checkout 18.2.2
# build vapor tool
swift build -c release --disable-sandbox --enable-test-discovery --verbose
# copy vapor executable to your favorite OS location
sudo cp .build/release/vapor /usr/local/bin
Now, we can setup a new project (this one is already done):
vapor new <YourProjectName>
and, start coding. To build the code just use
swift build
from your project area.
To run the new code please use
swift run
The server will start on port 8080.
You may add new area Model to your project where you can store
your TF models, e.g. when I added model.tf.npy TF model file it
appears at a load time
# next the following will appear on your screen
loading weights from: model/model.tf.npy
[ NOTICE ] Server starting on http://127.0.0.1:8080
To place a new request please use
curl http://localhost:8080/inference -H "Content-Type: application/json" -d '{"row":[5.9, 3.0, 4.2, 1.5]}'
The response will look like:
{"request":{"row":[5.0999999046325684,3.2999999523162842,1.7000000476837158,0.5]},"classes":["Iris setosa","Iris versicolor","Iris virginica"],"predictions":[0.84249889850616455,0.091388963162899017,0.066112160682678223]}