Upsampling layer for convolutional autoencoder

Question

I've been trying to build a convolutional autoencoder with Mathematica 12.3. For this, I need to reverse the action of PoolingLayer[]. Is there any layer similar to Keras's UpSampling1D? Could, perhaps, ReplicateLayer[] be handy here?

I just found this thread that talks about implementing an upsampling operation with DeconvolutionLayer[].

Answer 1

From what I see, UpSampling1D is equivalent to:

upsampling1DLayer[size_] := 
 NetChain[{ReplicateLayer[size, -2], 
   TransposeLayer[{-2 <-> -1, -3 <-> -2}], FlattenLayer[-1], 
   TransposeLayer[{-2 <-> -1}]}]

Applying to the example shown in Keras' documentation:

di = {{{0, 1, 2}, {3, 4, 5}}, {{6, 7, 8}, {9, 10, 11}}}
upsampling1DLayer[2]@di

{{{0., 1., 2.}, {0., 1., 2.}, {3., 4., 5.}, {3., 4., 5.}}, {{6., 7., 8.}, {6., 7., 8.}, {9., 10., 11.}, {9., 10., 11.}}}