How to "Add" multiple layers in a CNN architecture?

1

Hi, I have a question to implement “Add” in DL4J:

In Python, in my CNN model, I use “Add” to add several convolution layers together (same output form), such as:

layer_all_added = add([layer1, layer2, layer3])

then layer_all_added will be used as input for the next convolutional layer.

How can I implement it in DL4J?

Thanks.

2

@FanDev could you flesh this out a bit? Which API are you using? The ComputationGraph or Samediff? What is the add supposed to do? is that supposed to be an automatic concatneation? An element wise add?

3

Sure, let me explain it further.

In my neural network, I applied multiple dilation convolutions with different dilation rates, such as:

layer0 = Conv3D(neuron_size, (3,3,3), activation= 'relu', padding = 'same')(inputs)  //regular conv layer

layer1 = Conv3D(neuron_size, (3,3,3), activation='relu', padding='same', dilation_rate=1)(layer0) //dilation conv layer with dilation rate = 1
layer2 = Conv3D(neuron_size, (3,3,3), activation='relu', padding='same', dilation_rate=2)(layer1) //dilation conv layer
layer3 = Conv3D(neuron_size, (3,3,3), activation='relu', padding='same', dilation_rate=4)(layer2) //dilation conv layer
Add_All_Dilation = add([layer1, layer2, layer3])

merge1 = concatenate([UpSampling3D(size=(2,2,2))(Add_All_Dilation),layer0], axis = 4)
...

Since layer1, layer2, layer3 have the same output format, I am able to add their weights together to create a more strong feature map.

SO my questions are:

  1. How to perform “Add” in DL4J? Do I need to consider this as an additional layer?
  2. (extra question) How can I assign the concatenate (in DL4J called “MergeVetax”?) with axis=4 in DL4J?

Thanks.

4

@FanDev thanks a lot… You’ll want to use the element wise vertex with Op.add:

You would take the output names of those 3 layers as inputs.

1 Like