Add anchor_generator, box_matcher and non_max_supression

[sineeli]

Add below layers for RetinaNet

1. Non Max Suppression
2. Anchor Generator
3. Box Matcher

[sineeli]

Closed previous PR and made changes as per new keras-hub rename

[mattdangerw]

Looks good! Minor comments. Also, a heads up @fchollet is quickly adding bounding box support to core Keras, soon we will be able to move our implementation, though some porting might be required to the final API Francois cooked up.

[sineeli]

@mattdangerw where ever necessary if there are general ops. that we are using in layers should be casted toself.compute_dtype is it correct and is it mandatory to convert?

[mattdangerw]

@mattdangerw where ever necessary if there are general ops. that we are using in layers should be casted toself.compute_dtype is it correct and is it mandatory to convert?

Not sure I totally get the question, but in general, layers should make sure they are doing computation with the compute dtype. This will often happen automatically. Variables and inputs will automatically be cast to compute dtype in call. But when, say making a new array of floats inside call for whatever reason, that should usually be done with dtype=self.compute_dtype.

[sineeli]

Not sure I totally get the question, but in general, layers should make sure they are doing computation with the compute dtype. This will often happen automatically. Variables and inputs will automatically be cast to compute dtype in call. But when, say making a new array of floats inside call for whatever reason, that should usually be done with dtype=self.compute_dtype.

Thanks @mattdangerw for clarification, I wanted to know when we declare new arrays in layer call, clear now.