In keras 1 i used to write
def merge_mode(branches):
return #merge function
Merge([...], output_shape=(num_classes,), mode=merge_mode)
But now in keras 2.0 Merge is deprecated, and you can only add predefined functions like add, average, concatenate, ....
My question is how can i add a custom merge function in keras 2.0
If you're looking for a layer that has no trainable weights, which from your looking at your example, I guess is the case. Then you can use a Lambda layer.
I actually wrote one myself to create a siamese CNN using a squared distance measure. See example below:
from keras.layers import Lambda
# I create the layer structure first.
# This can then be added to a sequential model or used in a functional model.
merge_layer = Lambda(lambda x: K.square(x[0]-[x]), output_shape=lambda x: x[0])
# Inside a functional model you would use it like this:
block_output = merge_layer([left,right])
