During the past weeks, I’ve had the opportunity to play a bit with Wasm and Go.

All those experiments led me to a write a proof of concepts that can illustrate everything I have said recently about:

  • Thinking the deep-learning stack like an Ops (see my post about NNRE/NNDK).
  • Capturing the real value of the training process (the knowledge) into a sequence of bits (the lightning talk I gave about it at the dotAI should be online soon).

Greetings professor Falken… Shall we play a game?

For a demo, I have developed a simple LSTM that can play the tic-tac-toe game. I am a fan of this example for AI-related kinds of stuff, it is indeed a “Madeleine” for me.

I will not go into every single detail here, and you can find the code here. But the principle is as follow:

An autonomous code based on an LSTM is learning the possible combinations of a winning tic-tac-toe board for token X.

Then, the weights of the LSTM (the knowledge) are exported (as a Gob file).

Another code (the NNRE) car read the knowledge file, applies it to the LSTM model and implements the mechanism to play the game. The interactive part is concurrent and synchronized via channels (did I told you how much I like this model of concurrency :)) That’s almost it.

Wasm ?

This was my first POC. Then I realized that I could run all of this in the browser.

I first created a table to represent the board. I gave every cell an ID: ttt0 ... ttt8.

Within my GO/Wasm code, I added an EventListener to trigger a callback when a click on a cell is made:

1for i, v := range []string{"ttt0", "ttt1", "ttt2", "ttt3", "ttt4", "ttt5", "ttt6", "ttt7", "ttt8"} {
2      m := mycb{ v, i, p }
3      js.Global.Get("document").Call("getElementById", v).Call("addEventListener", "click", js.NewCallback(m.cb))

Note mycb is a just a wrapper to pass some parameters to the “cb” method which is actually the callback.

When the AI player is playing, an event is triggered, and the corresponding letter is placed in the correct cell of the table.

The knowledge

I really wanted to show that the knowledge was strictly separated from the code. At first, I used the “fetch” transport method from from Johan Brandhorst.

 1import "github.com/johanbrandhorst/fetch"
 3c := http.Client{
 4        Transport: &fetch.Transport{},
 6resp, err := c.Get("/tictactoe.bin")
 7// ...
 8defer resp.Body.Close()
 9model := new(lstm.Model)
10dec := gob.NewDecoder(resp.Body)
11err = dec.Decode(model)

But this was not visual enough, so I decided to implent it in another way. What I wanted was the ability to upload the file directly to my pseudo server. (yes, actually, in regard to the front, the wasm file is a server).

The trick is to use an HTML “input” element and use file as type. The usage is clearly documented in the article Using files from web applications

1<input type="file" id="knowledgeFile" multiple="" size="1" style="width:250px" accept=".bin">

Now, in the Go code, I can make an API call to access FileList that contains File Objects as described in the documentation.

1files := js.Global.Get("document").Call("getElementById", "knowledgeFile").Get("files")
2if files.Length() == 1 {

The first element of the list is now the first and unique object that contains all the pieces of information of the uploaded file. Now I have to enter a loop to wait for it to be fully loaded (there must be a better and more idiomatic way, but let’s keep that for another time)…

1reader := js.Global.Get("FileReader").New()
2reader.Call("readAsDataURL", files.Index(0))
3for reader.Get("readyState").Int() != 2 {
4        fmt.Println("Waiting for the file to be uploaded")
5        time.Sleep(1 * time.Second)

Once we have the file, we can read its content as a data-url encoded value. First pass of decoding to get the payload, then second pass of Gob-decoding to extract the weight and instantiate a new model, et voilĂ !

1content := reader.Get("result").String()
2dataURL, err := dataurl.DecodeString(content)
3if err != nil {
4        return nil, err
6model := new(lstm.Model)
7dec := gob.NewDecoder(bytes.NewReader(dataURL.Data))
8err = dec.Decode(model)

Go and test it live!

All of this to get this result. You can try it and have fun.

Warning Just a couple of warning. It can hang a tab of your browser and even all the browser. I don’t think it works on a mobile.

TODO: The algo do not check for a winning move. A new game is triggered by clicking the Run button. The process is not bulletproof as it kills the current process to reload a new one. Therefore it hangs from times to time. A better option would be to catch an event within the Go file and to trigger a new game. But, well, it is a POC :D

Download a “knowledge” (Thos have been pre-trained with different hyper parameters).

Upload it here:

Load the WASM file (the file is 25Mo): Load

Wait for the file to be compiled (the run button will become available): Run