Image Classifier update

[OrpheasK]

This PR fixes the issue #166 that @jackbdu brought into attention and also offers a chance to decide upon for the use/name of the topk parameter in conjunction with kNumber as already mentioned in issue #83 by @lindapaiste.

Also recently mentioned in #83 is the support of local loading of TM models among other subjects, which can be incorporated in this or another PR depending on the priority each aspect will be given.

[ziyuan-linn]

This looks good to me! I don't really have a strong opinion on the naming of topk, but perhaps another option could be classes. classes: 4 returns the top 4 classifications.

Personally I think we should only use the ml5.imageClassifier({topk: 4}) and not myImageClassifier.classify(image, 4), so the interface is more consistent with the other models. This might be a breaking change though, and I don't feel strongly about this either.

[lindapaiste]

resultCount might also make sense. Or something along those lines that’s more intuitive and less technical.

[shiffman]

Personally I think we should only use the ml5.imageClassifier({topk: 4}) and not myImageClassifier.classify(image, 4), so the interface is more consistent with the other models. This might be a breaking change though, and I don't feel strongly about this either.

@ziyuan-linn can you elaborate on this? What other models follow this pattern? Perhaps it's better to conform the other models to how we are doing it here? I think having the # of classes returned specified with classify() is more intuitive?

[ziyuan-linn]

@shiffman The other pretrained models like BodySegmentation and FaceMesh only accept a media element and an optional callback on the detection methods and do not accept any kind of options. For example faceMesh.detect(image, gotFaces);. All the options are set during the ml5.faceMesh() call.

I think setting the number of classes returned is unique to ImageClassifier (and maybe soundclassifier?).

I agree that the number of classes on the classify method is more intuitive.

[shiffman]

Apologies for letting this PR languish for a bit! Based on the discussion we had at our previous Thursday meeting, we decided to allow for passing in the number of desired classes + confidence scores to be returned as an argument to the classify() function itself. We can then extend this pattern to other models that work in a similar way (right now it would only be ml5.neuralNetwork() for the "classification" task. I'll open an issue regarding incorporating this in examples and website documentation.