MXNET

library(Hmisc)    # useful for multi histogram
hist.data.frame(full)  # histogram of each variable
pairs(full)            # pairwise scatter plots


data = mx.symbol.Variable('data')
fc1 = mx.symbol.FullyConnected(data=data, num_hidden=7)
sigm1 = mx.symbol.Activation(data=fc1, act_type="relu")
fc2 = mx.symbol.FullyConnected(data=sigm1, num_hidden=2)
net = mx.symbol.SoftmaxOutput(data=fc2)


device = mx.cpu();
model <- mx.model.FeedForward.create(
  X                  = as.matrix(train.x),
  y                  = as.numeric(train.y)-1,
  ctx                = device,    # device is either the cpu or gpu (graphical processor unit)
  symbol             = net,       # this is the network structure
  eval.metric        = mx.metric.accuracy,
  num.round          = 100,        # how many batches to work with
  learning.rate      = 0.01,      # 0.01 is a good start
  momentum           = 0.9,       # using second derivative
  wd                 = 0.0001,    # what is this for?
  #initializer        = mx.init.normal(1/sqrt(dim(train)[1])),   # the standard devation is scaled with the number of
  #observations to prevent slow learning if by chance all weights are large or small
  #initializer        = mx.init.uniform(0.1),   # the standard devation is scaled with the number of 
  array.batch.size   = 100,
  #epoch.end.callback = mx.callback.save.checkpoint("titanic"),
  #batch.end.callback = mx.callback.log.train.metric(100),
  array.layout="rowmajor"
);

mx.model.save(model, 'titanic100', 100)

model4 <- mx.model.FeedForward.create(
  symbol             = model3$symbol,
  X                  = as.matrix(train.x),
  y                  = as.numeric(train.y)-1,
  # eval.data          = list("data"=g_test,"label"=y_test),
  ctx                = device,    # device is either the cpu or gpu (graphical processor unit)
  eval.metric        = mx.metric.accuracy,
  num.round          = 1500,        # how many batches to work with
  learning.rate      = 0.0005,      # 0.01 is a good start
  momentum           = 0.9,       # using second derivative
  wd                 = 0.0001,    # what is this for?
  array.batch.size   = 100,
  arg.params=model3$arg.params,
  aux.params=model3$aux.params,
  verbose=FALSE,
  #epoch.end.callback = mx.callback.save.checkpoint("titanic"),
  #batch.end.callback = mx.callback.log.train.metric(100),
  array.layout="rowmajor"
);

mx.model.save(model4, 'titanic31000', 1500)