Theano support? about yad2k HOT 1 CLOSED

Yes and no.

If you're using a YOLO_v2 model (yolo.cfg and yolo-voc.cfg) with a passthrough layer then this depends on tf.space_to_depth. This is not in the Keras backend and I don't know if Theano has an equivalent function. I have seen an implementation of this using Theano here and would like to switch to something like that eventually. Ideally, space_to_depth will get incorporated into the Keras backend.

The other supported models like darknet19 and tiny-yolo-voc should work fine.

Also, note that yolo_head() which post-processes the CNN outputs is implemented as a Keras backend function and not a layer. yolo_eval() which is used for filtering predicted boxes based on score and non-maximum-suppression currently uses tf.non_max_suppression. A Theano or pure Python equivalent can be used instead.

from yad2k.