8

Does anybody know where can I find an example of LSTM via MXNet (R package)? The basic task is prediction of x[t + 1] value by x[1 ... t] sequence.

sermal
  • 93
  • 1
  • 6

1 Answers1

3

If you are still looking for it, there is this very nice blog explaining how to use RNN in mxnet with R bindings: https://www.r-bloggers.com/recurrent-models-and-examples-with-mxnetr/.

geoalgo
  • 678
  • 3
  • 11
  • There is still no tutorial on application of R MXNet to the numeric time-series. The link posted and other tutorials work with sequences of words and perform a classification task. If someone would create one this would be very **helpful.** – Alexey Burnakov Nov 22 '17 at 16:56
  • True but I think it should be relatively straightforward to replace the softmax output from this example by a mean squared error starting to adapt to numeric time-series. – geoalgo Nov 23 '17 at 09:26
  • I thought so, but it is more complicated. There are some other symbols that for example describe the dimension of output (which is the number of classes) and embedding size: **output_dim=num.embed**. Not sure how to convert that to a regression of one value ahead. – Alexey Burnakov Nov 23 '17 at 09:40
  • 1
    mx.symbol.SoftmaxOutput should be replacable by LinearRegressionOutput. For the input, you can remove the embedding. – geoalgo Nov 24 '17 at 12:05