Merge remote-tracking branch 'upstream/master' into dmlc-sparse-squash

R-package/R/io.R

@@ -1,14 +1,12 @@
-is.MXDataIter <- function(x) {
-  inherits(x, "Rcpp_MXNativeDataIter") ||
-  inherits(x, "Rcpp_MXArrayDataIter")
-}
-
 #' Judge if an object is mx.dataiter
 #'
 #' @return Logical indicator
 #'
 #' @export
-is.mx.dataiter <- is.MXDataIter
+is.mx.dataiter <- function(x) {
+  inherits(x, "Rcpp_MXNativeDataIter") ||
+  inherits(x, "Rcpp_MXArrayDataIter")
+}
 
 #' Extract a certain field from DataIter.
 #'

R-package/R/metric.R

@@ -78,3 +78,13 @@ mx.metric.rmsle <- mx.metric.custom("rmsle", function(label, pred) {
   return(res)
 })
 
+#' Perplexity metric for language model
+#'
+#' @export
+mx.metric.Perplexity <- mx.metric.custom("Perplexity", function(label, pred) {
+  label_probs <- as.array(mx.nd.choose.element.0index(pred, label))
+  batch <- length(label_probs)
+  NLL <- -sum(log(pmax(1e-15, as.array(label_probs)))) / batch
+  Perplexity <- exp(NLL)
+  return(Perplexity)
+})

R-package/R/model.R

@@ -279,7 +279,13 @@ mx.model.train <- function(symbol, ctx, input.shape, output.shape,
   return(model)
 }
 
-# Initialize parameters
+#' Parameter initialization
+#' @param symbol The symbolic configuration of the neural network.
+#' @param input.shape The shape of the input for the neural network.
+#' @param output.shape The shape of the output for the neural network. It can be NULL.
+#' @param initializer, initializer object. The initialization scheme for parameters.
+#' @param ctx mx.context. The devices used to perform initialization.
+#' @export
 mx.model.init.params <- function(symbol, input.shape, output.shape, initializer, ctx) {
   if (!is.MXSymbol(symbol)) stop("symbol need to be MXSymbol")
 
@@ -296,7 +302,7 @@ mx.model.init.params <- function(symbol, input.shape, output.shape, initializer,
 
 # Initialize the data iter
 mx.model.init.iter <- function(X, y, batch.size, is.train) {
-  if (is.MXDataIter(X)) return(X)
+  if (is.mx.dataiter(X)) return(X)
   if (is.null(y)) {
     if (is.train) stop("Need to provide parameter y for training with R arrays.")
     shape <- dim(X)

R-package/R/rnn_model.R

@@ -3,19 +3,9 @@ is.param.name <- function(name) {
            grepl('gamma$', name) || grepl('beta$', name) )
 }
 
-# Initialize parameters
-mx.model.init.params.rnn <- function(symbol, input.shape, initializer, ctx) {
-  if (!is.mx.symbol(symbol)) stop("symbol need to be MXSymbol")
-  slist <- symbol$infer.shape(input.shape)
-  if (is.null(slist)) stop("Not enough information to get shapes")
-  arg.params <- mx.init.create(initializer, slist$arg.shapes, ctx, skip.unknown=TRUE)
-  aux.params <- mx.init.create(initializer, slist$aux.shapes, ctx, skip.unknown=FALSE)
-  return(list(arg.params=arg.params, aux.params=aux.params))
-}
-
 # Initialize the data iter
 mx.model.init.iter.rnn <- function(X, y, batch.size, is.train) {
-  if (is.MXDataIter(X)) return(X)
+  if (is.mx.dataiter(X)) return(X)
   shape <- dim(X)
   if (is.null(shape)) {
     num.data <- length(X)
@@ -56,11 +46,11 @@ setup.rnn.model <- function(rnn.sym, ctx,
             }
         }
     }
-    params <- mx.model.init.params.rnn(rnn.sym, input.shapes, initializer, mx.cpu())
+    params <- mx.model.init.params(rnn.sym, input.shapes, NULL, initializer, mx.cpu())
     args <- input.shapes
     args$symbol <- rnn.sym
     args$ctx <- ctx
-    args$grad.req <- "add"
+    args$grad.req <- "write"
     rnn.exec <- do.call(mx.simple.bind, args)
 
     mx.exec.update.arg.arrays(rnn.exec, params$arg.params, match.name=TRUE)

R-package/tests/testthat/test_model.R

@@ -127,11 +127,11 @@ test_that("Fine-tune", {
 
   new_fc <- mx.symbol.FullyConnected(data = flatten, num_hidden = 2, name = "fc1")
   new_soft <- mx.symbol.SoftmaxOutput(data = new_fc, name = "softmax")
-  arg_params_new <- mxnet:::mx.model.init.params(symbol = new_soft,
-                                                 input.shape = list("data" = c(224, 224, 3, 8)),
-                                                 output.shape = NULL,
-                                                 initializer = mx.init.uniform(0.1),
-                                                 ctx = mx.cpu())$arg.params
+  arg_params_new <- mx.model.init.params(symbol = new_soft,
+                                         input.shape = list("data" = c(224, 224, 3, 8)),
+                                         output.shape = NULL,
+                                         initializer = mx.init.uniform(0.1),
+                                         ctx = mx.cpu())$arg.params
   fc1_weights_new <- arg_params_new[["fc1_weight"]]
   fc1_bias_new <- arg_params_new[["fc1_bias"]]
 

amalgamation/amalgamation.py

@@ -8,7 +8,8 @@
     'kvstore_dist.h', 'mach/clock.h', 'mach/mach.h',
     'malloc.h', 'mkl.h', 'mkl_cblas.h', 'mkl_vsl.h', 'mkl_vsl_functions.h',
     'nvml.h', 'opencv2/opencv.hpp', 'sys/stat.h', 'sys/types.h', 'cuda.h', 'cuda_fp16.h',
-    'omp.h', 'execinfo.h', 'packet/sse-inl.h', 'emmintrin.h', 'thrust/device_vector.h'
+    'omp.h', 'execinfo.h', 'packet/sse-inl.h', 'emmintrin.h', 'thrust/device_vector.h', 
+    'cusolverDn.h'
     ]
 
 minimum = int(sys.argv[6]) if len(sys.argv) > 5 else 0

amalgamation/mxnet_predict0.cc

@@ -26,6 +26,7 @@
 
 
 #include "src/ndarray/ndarray_function.cc"
+#include "src/ndarray/autograd.cc"
 #include "src/ndarray/ndarray.cc"
 
 #include "src/engine/engine.cc"