Backward propagation in Keras?
You simply don't. (Late edit: except when you are creating custom training loops, only for advanced uses)
Keras does backpropagation automatically. There's absolutely nothing you need to do for that except for training the model with one of the fit methods.
You just need to take care of a few things:
- The vars you want to be updated with backpropagation (that means: the weights), must be defined in the custom layer with the
self.add_weight()method inside thebuildmethod. See writing your own keras layers. - All calculations you're doing must use basic operators such as
+,-,*,/or backend functions. By backend, tensorflow/theano/CNTK functions are also supported.
This is all you need to have the automatic backpropagation working properly.
If your layers don't have trainable weights, you don't need custom layers, create Lambda layers instead (only calculations, no trainable weights).