[R] An easy-to-use interface for mlp training

R-package/DESCRIPTION

@@ -20,6 +20,6 @@ Imports:
     magrittr,
     stringr
 Suggests:
-    testthat
+    testthat, mlbench
 LinkingTo: Rcpp
 RoxygenNote: 5.0.0

R-package/NAMESPACE

@@ -37,6 +37,7 @@ export(mx.metric.custom)
 export(mx.metric.mae)
 export(mx.metric.rmse)
 export(mx.metric.rmsle)
+export(mx.mlp)
 export(mx.model.FeedForward.create)
 export(mx.model.load)
 export(mx.model.save)
@@ -65,6 +66,7 @@ export(mx.symbol.BatchNorm)
 export(mx.symbol.BlockGrad)
 export(mx.symbol.Concat)
 export(mx.symbol.Convolution)
+export(mx.symbol.Deconvolution)
 export(mx.symbol.Dropout)
 export(mx.symbol.ElementWiseSum)
 export(mx.symbol.Flatten)
@@ -78,6 +80,7 @@ export(mx.symbol.Pooling)
 export(mx.symbol.Reshape)
 export(mx.symbol.SliceChannel)
 export(mx.symbol.Softmax)
+export(mx.symbol.SoftmaxOutput)
 export(mx.symbol.Variable)
 export(mx.symbol.exp)
 export(mx.symbol.infer.shape)

R-package/R/mlp.R

@@ -0,0 +1,69 @@
+#' Convenience interface for multiple layer perceptron
+#' 
+#' @param data the input matrix.
+#' @param label the training label.
+#' @param hidden_node a vector containing number of hidden nodes on each hidden layer as well as the output layer.
+#' @param out_node the number of nodes on the output layer.
+#' @param dropout a number in [0,1) containing the dropout ratio from the last hidden layer to the output layer.
+#' @param activation either a single string or a vector containing the names of the activation functions.
+#' @param out_activation a single string containing the name of the output activation function.
+#' @param device whether train on cpu (default) or gpu.
+#' @param eval_metric the evaluation metric/
+#' @param ... other parameters passing to \code{mx.model.FeedForward.create}/
+#' 
+#' @examples
+#' 
+#' require(mlbench)
+#' data(Sonar, package="mlbench")
+#' Sonar[,61] = as.numeric(Sonar[,61])-1
+#' train.ind = c(1:50, 100:150)
+#' train.x = data.matrix(Sonar[train.ind, 1:60])
+#' train.y = Sonar[train.ind, 61]
+#' test.x = data.matrix(Sonar[-train.ind, 1:60])
+#' test.y = Sonar[-train.ind, 61]
+#' model = mx.mlp(train.x, train.y, hidden_node = 10, out_node = 2, out_activation = "softmax", 
+#'                learning.rate = 0.1)
+#' preds = predict(model, test.x)
+#' 
+#' @export
+mx.mlp <- function(data, label, hidden_node = 1, out_node, dropout = NULL, 
+                   activation = "tanh", out_activation = "softmax",
+                   device=mx.ctx.default(), ...) {
+
+  m <- length(hidden_node)
+  if (!is.null(dropout)) {
+    if (length(dropout) != 1) {
+      stop("only accept dropout ratio of length 1.")
+    }
+    dropout = max(0,min(dropout, 1-1e-7))
+  }
+
+  # symbol construction
+  act <- mx.symbol.Variable("data")
+  if (length(activation) == 1) {
+    activation <- rep(activation, m)
+  } else {
+    if (length(activation) != m) {
+      stop(paste("Length of activation should be",m))
+    }
+  }
+  for (i in 1:m) {
+    fc <- mx.symbol.FullyConnected(act, num_hidden=hidden_node[i])
+    act <- mx.symbol.Activation(fc, act_type=activation[i])
+    if (i == m && !is.null(dropout)) {
+      act <- mx.symbol.Dropout(act, p = dropout)
+    }
+  }
+  fc <- mx.symbol.FullyConnected(act, num_hidden=out_node)
+  if (out_activation == "rmse") {
+    out <- mx.symbol.LinearRegressionOutput(fc)
+  } else if (out_activation == "softmax") {
+    out <- mx.symbol.SoftmaxOutput(fc)
+  } else if (out_activation == "logistic") {
+    out <- mx.symbol.LogisticRegressionOutput(fc)
+  } else {
+    stop("Not supported yet.")
+  }
+  model <- mx.model.FeedForward.create(out, X=data, y=label, ctx=device, ...)
+  return(model)
+}

R-package/R/mxnet_generated.R

@@ -295,6 +295,37 @@ mx.symbol.Convolution <- function(...) {
   mx.varg.symbol.Convolution(list(...))
 }
 
+#' Apply deconvolution to input then add a bias.
+#' 
+#' @param data  Symbol
+#'     Input data to the DeconvolutionOp.
+#' @param weight  Symbol
+#'     Weight matrix.
+#' @param bias  Symbol
+#'     Bias parameter.
+#' @param kernel  Shape(tuple), required
+#'     deconvolution kernel size: (y, x)
+#' @param stride  Shape(tuple), optional, default=(1, 1)
+#'     deconvolution stride: (y, x)
+#' @param pad  Shape(tuple), optional, default=(0, 0)
+#'     pad for deconvolution: (y, x)
+#' @param num.filter  int (non-negative), required
+#'     deconvolution filter(channel) number
+#' @param num.group  int (non-negative), optional, default=1
+#'     number of groups partition
+#' @param workspace  long (non-negative), optional, default=512
+#'     Tmp workspace for deconvolution (MB)
+#' @param no.bias  boolean, optional, default=True
+#'     Whether to disable bias parameter.
+#' @param name  string, optional
+#'     Name of the resulting symbol.
+#' @return out The result mx.symbol
+#' 
+#' @export
+mx.symbol.Deconvolution <- function(...) {
+  mx.varg.symbol.Deconvolution(list(...))
+}
+
 #' Apply dropout to input
 #' 
 #' @param data  Symbol
@@ -479,7 +510,7 @@ mx.symbol.SliceChannel <- function(...) {
   mx.varg.symbol.SliceChannel(list(...))
 }
 
-#' Perform a softmax transformation on input.
+#' DEPRECATED: Perform a softmax transformation on input. Please use SoftmaxOutput
 #' 
 #' @param data  Symbol
 #'     Input data to softmax.
@@ -496,6 +527,23 @@ mx.symbol.Softmax <- function(...) {
   mx.varg.symbol.Softmax(list(...))
 }
 
+#' Perform a softmax transformation on input, backprop with logloss.
+#' 
+#' @param data  Symbol
+#'     Input data to softmax.
+#' @param grad.scale  float, optional, default=1
+#'     Scale the gradient by a float factor
+#' @param multi.output  boolean, optional, default=False
+#'     If set to true, for a (n,k,x_1,..,x_n) dimensionalinput tensor, softmax will generate n*x_1*...*x_n output, eachhas k classes
+#' @param name  string, optional
+#'     Name of the resulting symbol.
+#' @return out The result mx.symbol
+#' 
+#' @export
+mx.symbol.SoftmaxOutput <- function(...) {
+  mx.varg.symbol.SoftmaxOutput(list(...))
+}
+
 #' Take exp of the src
 #' 
 #' @param src  Symbol