5

There seems no built-in support for Mean Pooling layer for RNN in Keras. Anyone knows how to wrap one?

enter image description here http://deeplearning.net/tutorial/lstm.html

Isilmë O.
  • 1,668
  • 3
  • 23
  • 35
  • Look at this answer, use [TemporalMeanPooling](http://stackoverflow.com/a/43014228/1105231) may help you. – XY WANG Mar 25 '17 at 08:35

1 Answers1

7

Keras has a layer AveragePooling1D for that. If you use the graph API, you should be able to do something like:

model.add_node(AveragePooling1D(...), 
               inputs=['h0', 'h1', ..., 'hn'], 
               merge_mode='concat', ...)
Mikael Rousson
  • 2,247
  • 1
  • 15
  • 16
  • Do you have any example for Sequential model? Eg. how to connect AveragePooling1D with an RNN layer that has return_sequence enabled. – Isilmë O. Feb 14 '16 at 13:11
  • I don't have an example but it looks like should be able to just add an `AveragePooling1D` layer to your `Sequential` after the `RNN` layer and specify the correct dimension for the pooling. – Mikael Rousson Feb 14 '16 at 15:39
  • @IsilmëO - since you've accepted the answer, can you show a complete example? I'm unable to follow the `AveragePooling1D` layer with a `Dense(1)` layer – mossaab Apr 01 '16 at 03:27