variational autoencoder

[dfalbel]

setting up variational autoencoder
not ready to merge

[dfalbel]

@jjallaire this is now ready to merge

[dfalbel]

@jjallaire here I would like to pass a list of tensors to the lambda layer, but

layer_lambda(list(z_mean, z_log_var)) was giving an error Error: Invalid input to layer function (must be a model or a tensor)

So I concatenated the values and separated inside the function. Do you see a better approach?

[jjallaire]

This is the code which receives the layer function input:

https://github.com/rstudio/keras/blob/master/R/layers-core.R#L420-L440

All I can think to do is to check for a list and just call layer_concatenate under the hood. However, I'm not sure if that results in an improvement in the code (as I think it's somewhat nice to see the explicit concatenation operation in the code).

[dfalbel]

How does layer_concatenate works? It´s very similar to passing a list to layer_lambda. The python equivalent code uses a list in this example:

https://github.com/rstudio/keras/blob/master/vignettes/examples/variational_autoencoder.py#L29-L36

We could add a compose_layer.default checking for a listof tensors too. What do you think?

[jjallaire]

After reviewing the equivalent Python code, I think this should be able to work exactly as you are hoping. Could you provide a version of the code which exhibits the Error: Invalid input and I'll debug it from there?

[dfalbel]

Here's a minimal example:

x <- layer_input(c(100,2))
y <- layer_input(c(100,2))

my_fun <- function(args){
  a <- args[[1]]
  b <- args[[2]]
  a + b
}

layer_lambda(list(x, y), my_fun)

In the context of the example I would like to do:

sampling <- function(arg){
  z_mean <- arg[[1]]
  z_log_var <- arg[[2]]
  
  epsilon <- K$random_normal(
    shape = c(batch_size, latent_dim), 
    mean=0.,
    stddev=epsilon_std
  )
  
  z_mean + K$exp(z_log_var/2)*epsilon
}

# note that "output_shape" isn't necessary with the TensorFlow backend
z <- layer_lambda(list(z_mean, z_log_var), sampling)

[jjallaire]

I think what you want is this then:

library(keras)

x <- layer_input(c(100,2))
y <- layer_input(c(100,2))

my_fun <- function(args){
  a <- args[[1]]
  b <- args[[2]]
  a + b
}

z <- layer_lambda(f = my_fun, arguments = list(x, y))

The reason you need the explicit f argument is that R keras layers always take a model (or tensor) as their first argument (for magrittr chaining). This isn't required though, but if you want evade this requirement you need to explicitly name your other arguments (that's why in our docs/examples we always name the first argument e.g. units in layer_dense) so that users become accustomed to always naming layer arguments.

[dfalbel]

That's exactly what I needed. I will submit a PR to fix this in the example.