creating createResNetModel1D

[mg64ve]

Hello, interesting project, thanks for sharing.
I am trying to write createResNetModel1D.
I have just taken createResNetModel2D and tried to convert to 1D layers.
I have some doubts on the following lines in squeezeAndExciteBlock2D:

    if( K$image_data_format() == "channels_first" )
      {
      block <- block %>% layer_permute( c( 4, 2, 3 ) )
      }
    x <- list( initial, block ) %>% layer_multiply()

I am using the following instead of c(4, 2, 3):

layer_permute( c( 4, 3 ) )

Not sure if it might be correct. What is the meaning of this permutation and when is supposed to be applied?

[ntustison]

@mg64ve --- good question. Unfortunately, my experience with both the squeeze-and-excite block as well as 1-D convolutional architectures are very limited. In fact, I haven't had a chance to test this yet. Since the 2-D case is

layer_permute( c( 4, 2, 3 ) )

and the 3-D case is

layer_permute( c( 5, 2, 3, 4 ) )

I suspect the 1-D case would be

layer_permute( c( 3, 2 ) )

but I'd have to do some more testing/thinking to be sure.

[mg64ve]

Thanks @ntustison
It seems to work.
If I have time to document and add an example I will create a pull request, I guess it might be useful for someone else.
However now I would like to transform createResNetModel1D 1D ResNet CNN neural network model in an autoencoder.
Any good idea on how to make it?

[ntustison]

Sorry, I don't. We have a couple of autoencoder code snippets but I haven't translated any of them to 1-D.

[mg64ve]

Ok @ntustison , but do you have 2D or 3D resnet autoencoder?

[ntustison]

No, we don't have that architecture implemented.