Support for a concatenate layer

[jvdp1]

State-of-the-art:

• https://keras.io/api/layers/merging_layers/concatenate/

[milancurcic]

Looks like the implementation would be straightforward. What are some uses of this?

[jvdp1]

In my context I have two types of input (DNA samples and environmental factors). I would like to investigate different types of layers (CNN for DNA and Dense for environmental factors).

[jvdp1]

As discussed here is a potential example:

  !Define the Convolutional network branch for the feature 1
  net_feat1 = network([ &
    input(ninput1), &
    reshape([1,28,28]), &
    conv1d(filters=8, kernel_size=3, activation=relu()), &
    maxpool1d(pool_size=2), &
    dense(64, activation=relu()) &
  ])

  !Define the Dense network branch for the feature 2
  net_feat2 = network([ &
    input(ninput2), &
    dense(32, activation=relu()) &
  ])

  !Combine the outputs of the two network branches and define the final dense layers for predi
ction
  net_combi = network([&
    input = concatenate(net_feat1%predict(), net_feat2%predict()), &
    dense(64, activation=relu(), &
    dense(1, activation=linear()) &
  ])

  !Train the model including its two branches
  epochs: do n = 1, num_epochs

    call net_combi % train( &
      input = [input%features1, input%features2], &  !could be a single array
      output_data, &
      batch_size=32, &
      epochs=1 &
    )

Prior art: Keras concatanate

[milancurcic]

Thanks, Jeremie, that's clear!