Using the libraries in non-browser NodeJS about teachablemachine-community

[traumverloren]

I found this issue when I was trying to figure out how to use my teachable machine model in Node.js. I thought I would comment in case others end up here wondering how to use their model in Node.js with Tensorflow.js. I created a starting template that might be helpful to others. (FYI it is not using the teachable machine library, only its exported trained model.)

https://github.com/traumverloren/teachable-machine-image-nodejs-starter

[irealva]

Hi @tr7zw and @steamp0rt thanks for filing the issue. We are working on the savedmodel export it should be fixed soon.

And you're right we have not implemented nodeJs support. It's good to know that you would like to use it in this environment and we'll keep it in mind for future planning.

[tr7zw]

After some playing around I was able to get it running in NodeJs with teachablemachine/image by doing some workarounds with jsdom and canvas, still a "native" version would be nice.

[irealva]

Just a note that the exporting to Tensorflow SavedModel is fixed in latest release.

[tr7zw]

I tried that too, but it looks like "@teachablemachine/image" isn't made for NodeJs at all. It tries to use Canvases and the dom to work with the images. Hopefully, a Node compatible version gets released soon.

how does someone make something as useful as this... without thinking "hey someone might want to use this out-of-browser"

[tr7zw]

Well, I'm stuck trying to get it running inside Java. The Savemodel export is broken, and the Keras model is incompatible with deeplearning4j's Keras import. Also a bit confusing that all examples are with a webcam, not showing how to get it running in NodeJS/Java/Browser/whatever to classify normal images via upload/post/filesystem.

[hapticdata]

Hi,
Just to add clarification. The model you create with Teachable Machine will work with node.js + Tensorflow.js; however our library does not support node.js.
We absolutely did consider use outside the browser which is why we implemented converters for our models to tensorflow and tflite and have even released a video showing a tflite model working offline on a small coral board.

Having the library work as attempted above would not be possible because it would require including node's 'fs' module in the browser and other aspects such as converting an image to a tensor would be different than in browser.

We will add language to the readme to be more explicit these modules are meant for web users.
thanks

Thanks for clarifying!

[luke-foxed]

After some playing around I was able to get it running in NodeJs with teachablemachine/image by doing some workarounds with jsdom and canvas, still a "native" version would be nice.

Could you possibly share how you achieved that? I've been trying to do the same with canvas but have had no luck.

[tr7zw]

After some playing around I was able to get it running in NodeJs with teachablemachine/image by doing some workarounds with jsdom and canvas, still a "native" version would be nice.

Could you possibly share how you achieved that? I've been trying to do the same with canvas but have had no luck.

Sure. I quickly threw together a git containing everything needed here.
Just keep in mind that I'm not a Javascript/node/web developer, this setup is enough to be able to expose multiple models as REST API for other projects(languages) to use, but probably has some horrible flaws if you want to have it facing the web^^.

[luke-foxed]

Thanks! I got something working too after a lot of trial and error, it's along the same idea as yours. Thanks for sharing!

[irealva]

Added your repo to this section, thanks for that!

https://github.com/googlecreativelab/teachablemachine-community/tree/master/libraries