require model name for ImageClassifier transfer learning

[shiffman]

This is the present code for the transfer learning example:

classifier = new ml5.ImageClassifier(video, modelLoaded);

Since we are planning to eventually have an empty ImageClassifier (see ml5js/ml5-library#121 and ml5js/ml5-library#128) I think we should require that you explicitly reference the pre-trained model. There is no default one. This is also more clear to an end-user as to what is going on.

This might change a bit if we go with the FeatureExtractor idea, but for now the example should probably be:

classifier = new ml5.ImageClassifier('MobileNet', video, modelLoaded);

@cvalenzuela Is this already supported or do we have to add support for the above to the library?

[cvalenzuela]

Yes, agree! We do need to add support for this. It shouldn't be much work.

Although maybe this is better to make together with the FeatureExtractor class. I started working on a branch for this yesterday. I hope to pick it up this weekend and submit a PR including it.

[shiffman]

Great. I am starting to wrap my head around the idea of a FeatureExtractor and I think it could work really well. Since transfer learning with an image classification model a la "Teachable Machine" will likely end up one of the most common applications of this, maybe we should still allow it to happen via the ImageClassifier natively. But to do regression or other more custom stuff, the FeatureExtractor would be required?

[cvalenzuela]

Yep, I think keeping the transfer learning inside ImageClassifier is a good idea.
I will continue to work on creating the feature extractor class then.

[cvalenzuela]

Closed, see ml5js/ml5-library#142