Multi-label model support for image classification

[Antal]

It would be great to have the option for multi-label problems in VisualML image classification.

In this use case the target is usually an array of one hot encoded classes and an image can belong to one or more classes. An image can be classified not as a single class (the class with the highest probability), but as multiple classes.

This includes the model top end to not be of type

Dense(n_classes, activation='softmax')

but of type

Dense(n_classes, activation='sigmoid')

Loss function should be 'binary_crossentropy'

[Jason]

Adding my support for this feature:

I want to label data that is not actually an image, but it is convenient enough to convert my data to a graph, which can then be saved as an image, and annotated (the task I am trying to solve already classifies this data using a visual representation.  The labels can then be applied back to the raw data.  However, one sample of my data can be positive for many classes.  Regardless of the model type used this feature would be very useful (in my case, I'm partitioning my models such that each independent model can call positive/negative for the feature it is trained on.)

I would add that this doesn't apply only to images.  Any type of data/problem space could benefit from multi-class annotation.  I think it should be an option for any labeling task to be single class OR multi-class.

[Jason]

Adding my support for this feature:

I want to label data that is not actually an image, but it is convenient enough to convert my data to a graph, which can then be saved as an image, and annotated (the task I am trying to solve already classifies this data using a visual representation.  The labels can then be applied back to the raw data.  However, one sample of my data can be positive for many classes.  Regardless of the model type used this feature would be very useful (in my case, I'm partitioning my models such that each independent model can call positive/negative for the feature it is trained on.)

I would add that this doesn't apply only to images.  Any type of data/problem space could benefit from multi-class annotation.  I think it should be an option for any labeling task to be single class OR multi-class.