Flow

This is work-in-progress. The program may be buggy. You are warned!

Flow is a flow-based automation tool that is capable of running simple Machine Learning programs. In Flow, users mainly work with a drag-and-drop flow-based "graph" to create Machine Learning components (trainers and minimizers) and external ones (sending tweets, fetch HTTP endpoints, etc.).

Under the hood, Flow uses Tensorflow to do all the Machine Learning tasks.

Flow-based Graph

Due to different iterations of this program, you may come across terms like "receipts", "elements" or "graphs", which all mean the same thing -- a flow-based graph. A flow-based graph consists of "nodes" that have different "inports" and "outports". Inports are used for receive values and outports are for sending out values. Those outports can be connected with other inports of different nodes in the same graph. Those connections transfer values.

A node is associated with a type that does computations to produce values for outports. The node, when running, is initialized from its type to create "instances".

A graph itself can also be a type. For this case, a graph can have graph-level inports and outports.

Unlike other flow-based programming models like NoFlo, Inports and outports of a node are always associated with some values -- in another word, it is not pulses, but a states. Nodes which send data to external endpoints (for example, sending tweets) usually requires an inport to be presented for a certain period of time (for example, five seconds). Respectively, nodes which receive data from external endpoints usually have their own refresh rate.

Machine Learning

Currently, the Machine Learning part of Flow is work-in-progress. It can only handle really simple ones like MNIST handwritten digits. An example is shown below.

Quickstart

The program consists of several components. They are all required to be running for the whole program to be functional.

Mongo Database

Have a Mongo database running in localhost of the default port. Refer to MongoDB's documentation for how to do this.

flow-tensorflow Server

Start the flow-tensorflow server. Refer to this repository for how to do this.

Flow Server in Development Mode

Install Boot. Go to this project's root and run:

boot dev

You should see something like this:

Starting reload server on ws://localhost:41555
Writing boot_reload.cljs...
Writing boot_cljs_repl.cljs...

Starting file watcher (CTRL-C to quit)...

Compiling {less}... 87 changed files.
Adding :require adzerk.boot-reload to editor.cljs.edn...
nREPL server started on port 45597 on host 127.0.0.1 - nrepl://127.0.0.1:45597
Adding :require adzerk.boot-cljs-repl to editor.cljs.edn...
Compiling ClojureScript...
• js/editor.js
Elapsed time: 29.413 sec

Connect to the nREPL using your favorite Clojure tools (Emacs' cider, for example). If you don't have one installed, simply run this command:

boot -C repl -cp 45597

In which 45597 is the nREPL port opened as shown above.

In the REPL, run:

(in-ns 'dev)
(start)

And you should see the server starts running.

Register a User and Create Your First Graph

Go to http://127.0.0.1:10555 and you should see Flow's welcome page. Register a user here (Sign Up button on the top-right), and go to http://127.0.0.1:10555/graphs/home. You will see your home graph.

A MNIST Trainer

Currently to create a trainer the procedure is a little bit tedious, but nonetheless here are the steps.

MNIST Minimizer

Create a new graph by pressing the start ("*") icon on top-left. You should see a new graph node appears. Press the edit button on that node to enter the graph.

In here, we use the "New Instance" search box on top-right to search for and add appropriate nodes. The result would be something like below:

Create two graph inports y and y_ (using the inports panel on the right), connect y to tensorflow/log's in inport, and connect y_ to tensorflow/time's b inport.

Create one graph outport out and connect it to tensorflow/mean's out outport.

Look at the location of your current browser, remember the minimizer's id (the long string after 127.0.0.1:10555/graphs/.

You can also press the edit button just right after the graph's name (on top-left) to give the graph a better name (for example, "Example Minimize Graph").

MNIST Trainee

Go back to your home graph (by pressing browser's go-back button). Create another new graph (by the star "*" button on top-left). Another new graph node should appear. Click the edit button on that node.

In here, we only need a softmax node to make it functional. The result would look something like below:

Create a graph inport x, and connect it to tensorflow/softmax's in. Create a graph outport y, and connect it to tensorflow/softmax's out.

Look at the location of your current browser, remember the trainee's id (the long string after 127.0.0.1:10555/graphs/.

Modify the Server, Just a Little Bit...

Go to project folder's src/clj/flow/walker/coordinators/primitives.clj. Around line 106, change that to trainee's id as you got above. Around line 112, change that to minimizer's id as you got above.

Now, you need to restart the server. Go to the REPL and run:

(reset)

Restart the Server and Create the Final Graph

Go to 127.0.0.1:10555/graph/home again. Create a graph like this:

Once you connected everything, the server will start to train MNIST models and you are done!