Test replacing Watson Visual Recognition with TF

[dalelane]

The sound model support is currently implemented in JavaScript, using TensorflowJS to train and run models in the browser.

https://github.com/IBM/taxinomitis/blob/316faf5087b635241801b143e2a252882f97ab84/public/components/soundtraining/soundtraining.service.js#L113-L265

I'd like to explore doing something similar for images projects.

Advantages:

• Removes the need for managing API keys
• Potentially faster to train models (avoids the queuing that sometimes happens with the cloud service)

Disadvantages:

• Increases hardware requirement for student - would a browser running on a Raspberry Pi, for example, manage?
• Prevents projects being made in non-Scratch platforms (e.g. App Inventor) without implementing something native for those environments as well

At this point, I'd like to evaluate whether the models would perform well enough for this to be an option.

[dalelane]

This is now implemented, and live in the site. Access to it is currently restricted to a small number of beta testers while it can be tested more thoroughly.

[dalelane]

Testing has been successful so far, so I'm enabling the feature while flagging it as a beta and leaving Watson Visual Recognition in as an option.