求调试

nnet_vs_mxnet.R

## nnet
library(nnet)
n = nnet(Species ~ ., data = iris, size=2, maxit = 500, abstol=1e-6)
mean(predict(n, iris[, 1:4], type='class') == iris[, 5])
# 0.66 or 0.99
library(devtools)
source_url('https://gist.githubusercontent.com/fawda123/7471137/raw/466c1474d0a505ff044412703516c34f1a4684a5/nnet_plot_update.r')
plot.nnet(n)

## mxnet
library(mxnet)
index = sample(1:150, 150)
# x_train = data.matrix(iris[index , 1:4])
x_train = scale(data.matrix(iris[index , 1:4])) 
y_train = as.numeric(iris[index, 5])

d = mx.symbol.Variable('data')
h1 = mx.symbol.FullyConnected(data=d, num.hidden=2)
h1 = mx.symbol.Activation(data=h1, act.type='sigmoid') # use 'relu' or 'tanh' does not help
h2 = mx.symbol.FullyConnected(data=h1, num.hidden=3)
res = mx.symbol.SoftmaxOutput(h2)

mx.set.seed(0)
model <- mx.model.FeedForward.create(res, 
                                     X=x_train, y=y_train, 
                                     ctx=mx.cpu(), 
                                     num.round=500, array.batch.size=150, 
                                     learning.rate=1e-6, momentum=0.9, 
                                     eval.metric=mx.metric.accuracy)

mean(max.col(t(predict(model, x_train))) == y_train)
# always 0.33