Support for neural net regression models about deepdetect

[beniz]

Note: support for float 'labels' for regression requires modifying the Datum protobuf format from Caffe as it defines an int32 as the reference value.

[beniz]

FTR, turning Caffe label into a float is related to BVLC/caffe#881 and BVLC/caffe#2139

[revilokeb]

Changing label in Datum to float allows for regressing a single float. A more general approach is also described in https://github.com/BVLC/caffe/issues/881by Evan Shelhamer, i.e. storing any number of floats in a second lmdb by using data in Datum, later slice the data in the train_val.prototxt. Are there plans to incorporate multi-regression in dede, too?

[beniz]

Yes, I had definitely well noted the comment by Evan! My rational at the time was, like many things involving Caffe, that the amount of work for getting the full and clean functionality working was too much to be done properly immediately vs the quick hack that seemed not too awful in the end.

Multi-regression is a real need that I do believe requires its own ticket and focus. The plan is definitely there, though it will not be immediate. Ticket to come then, and some discussion in it as well.