Autoenconders for R package

[mateuszr]

Hi! Is there a plan to include a slightly more high-level API for the R package to setup autoencoders? I've noticed the python example in example/autoencoder but unfortunately there is no equivalent for R package. H2O provides some sort of simple API for autoencoders construction, which could be some example to mimick:

model = h2o.deeplearning(
  x = features,
  training_frame = MDIG,
  hidden = c(400, 200, 2, 200, 400 ),
  epochs = 600,
  activation = 'Tanh',
  autoencoder = TRUE
)

Thanks!

[thirdwing]

Hi, @mateuszr

Actually I spent some time on autoencoder in R https://github.com/thirdwing/mxnet/blob/autoencoder/example/autoencoder/autoencoder.R

But I don't have enough time to finish it.

Besides, I am also thinking about a question: does R user prefer the high-level API, such as in h2o, or examples building networks step by step?

[mateuszr]

Hello! That looks very good! Btw. I'm getting "Error: not compatible with requested type" when testing it with MNIST dataset. Do you have some running example of invocation for the AE_setup()?

Regarding the API question: I can speak only for myself, but I suppose that some clean high-level API (but with enough parameters for some hyperparameter optimisation) probably solves most use cases and provides a very good starting point for experimentation. For other custom autoencoders, people would probably construct their own networks... Best,

[terrytangyuan]

@thirdwing Any progress on this?

R users definitely prefer high-level API. I suspect that many people are willing to build networks by themselves.

[thirdwing]

I just found a good blog on autoencoder in R using MXNet from http://qiita.com/mwmsnn/items/6ca48a3b8292211c1acd

I will invite the author to write a vignette.

[ankkhedia]

@sandeep-krishnamurthy Could you please tag this issue as feature request