0
votes

I’m working on a TJBot visual recognition application.

I have some particular questions about retraining.

My case is similar to this:

I have a model trained for cats, dogs and a negative category (not cats or dogs).

The visual recognizer sometimes gets the wrong answer.

What I’d like to do is ‘automatically’ retrain the model, programmatically using an API whenever it gets the wrong answer.

For example after making the wrong determination of a dog when it was actually a cat, I’d like to turn around and call the visual recognition retraining api and indicate: this picture is a cat. This picture is not a dog.

Questions: 1) Is there an API for ‘partial retraining’ using just one image? I don’t want to have to upload hundreds of photos again to completely retrain the model if I don’t have to.

2) In this case I’d like to say this is a positive example of a cat, and a negative example of a dog. As I understand it there’s no way to indicate ‘negative example of a dog’. I can only say ‘negative example of cat and dog’. Is there a way to indicate ‘negative example of just one of the trained classes?

3) If there’s no API, can I accomplish #1 or #2 using the ‘studio’ website? How?

Thanks, Andy Citron (IBM retired)

1

1 Answers

0
votes

Thanks for your question.

  1. Yes, you can send a single image as the data for retraining a classifier. However, 1 image is unlikely to affect your classifier much. We suggest sending at least 10 new images in each retraining request.

  2. If sent as a "negative_example" the image should be neither a cat or a dog. The classes in a classifier should be mutually exclusive. A photo of a cat and dog together does not work well for training a system which is trying to tell the difference between the two types. To the classifier, the types of answers it can give are defined by the classes, or if negative_examples are used "none of the above". So in your example, every image in the world gets classified by this model as either cat, dog, or neither.

  3. The API is documented here: https://www.ibm.com/watson/developercloud/visual-recognition/api/v3/curl.html?curl#update-classifier