1

I am trying to implement a simple Multi-layer feed forward neural network using "neuralnet" package available in R for the "iris" dataset.

The code that I am using is as follows-

library(neuralnet)
data(iris)

D <- data.frame(iris, stringsAsFactors=TRUE)

# create formula-
f <- as.formula(Species ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width)

# convert qualitative variables to dummy (binary) variables-
m <- model.matrix(f, data = D)

# create neural network-
iris_nn <- neuralnet(f, data = m, hidden = 4, learningrate = 0.3)

I have two questions at this point of time-

1.) How do I use the "hidden" parameter? According to the manual pages, its saying-

hidden: a vector of integers specifying the number of hidden neurons (vertices) in each layer

How should I supply the vector of integer? Say if I wanted to have 1 hidden layer of 4 neurons/perceptrons in each layer Or if I wanted to have 3 hidden layers of 5 neurons in each layer.

2.) The last line of code gives me the error-

Error in eval(predvars, data, env) : object 'Species' not found

If I remove the "hidden" parameter, this error still persists.

What am I doing wrong here?

Edit: after adding the line-

m <- model.matrix(f, data = D)

The matrix 'm' no longer contains "Species" variable/attribute which I am trying to predict.

Output of 

str(D)

str(D) 'data.frame': 150 obs. of 5 variables: $ Sepal.Length: num 5.1 4.9 4.7 4.6 5 5.4 4.6 5 4.4 4.9 ... $ Sepal.Width : num 3.5 3 3.2 3.1 3.6 3.9 3.4 3.4 2.9 3.1 ... $ Petal.Length: num 1.4 1.4 1.3 1.5 1.4 1.7 1.4 1.5 1.4 1.5 ... $ Petal.Width : num 0.2 0.2 0.2 0.2 0.2 0.4 0.3 0.2 0.2 0.1 ... $ Species : Factor w/ 3 levels "setosa","versicolor",..: 1 1 1 1 1 1 1 1 1 1 ...

I have coded this with "nnet" successfully. Posting my code for reference-

data(iris)
library(nnet)

# create formula-
f <- as.formula(Species ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width)

# create a NN with hidden layer having 4 neurons/node and
# maximum number of iterations = 3
iris_nn <- nnet(f, data = iris, size = 4, maxit = 3)

# create a test data-
new_obs <- data.frame(Sepal.Length = 5.5, Sepal.Width = 3.1, Petal.Length = 1.4, Petal.Width = 0.4)

# make prediction-
predict(iris_nn, new_obs)   # gives percentage of which class it may belong
predict(iris_nn, new_obs, type = "class")   # gives the class instead of percentages of which 'class' this data type may belong to


# create a 'confusion matrix' to measure accuracy of model-
# rows are actual values and columns are predicted values-
# table(iris$Species, predict(iris_nn, iris[, 1:4], type = "class"))
cat("\n\nConfusion Matrix for # of iters = 3\n")
print(table(iris$Species, predict(iris_nn, iris[, 1:4], type = "class")))
cat("\n\n")

rm(iris_nn)

# setting 'maxit' to 1000, makes the model coverge-
iris_nn <- nnet(f, data = iris, size = 4, maxit = 1000)

# create a new confusion matrix to check model accuracy again-
cat("\n\nConfusion Matrix for # of iters = 1000\n")
print(table(iris$Species, predict(iris_nn, iris[, 1:4], type = "class")))
# table(iris$Species, predict(iris_nn, iris[, 1:4], type = "class"))


# to plot 'iris_nn' trained NN-
# library("NeuralNetTools")
# plotnet(iris_nn)

Thanks!!

Arun
  • 2,222
  • 7
  • 43
  • 78
  • 1
    Possible duplicate of [Working with neuralnet in R for the first time: get "requires numeric/complex matrix/vector arguments"](https://stackoverflow.com/questions/17457028/working-with-neuralnet-in-r-for-the-first-time-get-requires-numeric-complex-ma) – SamFlynn Dec 16 '17 at 16:17
  • @SamFlynn I have edited my post to include the matrix 'm'. But now the variable which I am trying to predict "Species" is gone! Hence the last line of code is giving error that "Species" could not be found! Any ideas? – Arun Dec 16 '17 at 16:29
  • I too tried that, couldn't figure it out kept getting some error. Add the output of `str(d)` in the question. What I did was change all factorial columns manually to dummy variables and it worked. – SamFlynn Dec 16 '17 at 16:33
  • Will normalization of the attributes help? – Arun Dec 16 '17 at 16:35

2 Answers2

1

No clue how NN runs and what's the best way to run it. Don't know much about the iris dataset as well.

Just pointing out why its not running - the column Species 

str(d)
'data.frame':   150 obs. of  5 variables:
 $ Sepal.Length: num  5.1 4.9 4.7 4.6 5 5.4 4.6 5 4.4 4.9 ...
 $ Sepal.Width : num  3.5 3 3.2 3.1 3.6 3.9 3.4 3.4 2.9 3.1 ...
 $ Petal.Length: num  1.4 1.4 1.3 1.5 1.4 1.7 1.4 1.5 1.4 1.5 ...
 $ Petal.Width : num  0.2 0.2 0.2 0.2 0.2 0.4 0.3 0.2 0.2 0.1 ...
 $ Species     : Factor w/ 3 levels "setosa","versicolor",..: 1 1 1 1 1 1 1 1 1 1 ...

Species is a factor NN doesnt take factors.

Convert to dummy varibles - 

d$set <-0
d$set[d$Species == "setosa"] <- 1

d$versi <-0 
d$versi[d$Species == "versicolor"] <- 1



f <- as.formula(set+versi ~ Sepal.Length + Sepal.Width + Petal.Length + Petal.Width)

iris_nn <- neuralnet(f, data = d, hidden = 4, learningrate = 0.3)

EDIT:

So when you say hidden = c(5,3) then the neural network diagram would have your input nodes, 5 side by side hidden nodes(a layer), 3 side by side hidden nodes(another layer), output node/nodes

No clue how they impact the accuracy.

The compute for neuralnet is like predict for all other machine learning models. 

library(neuralnet)
library(caret) #has the confusionmatrix function in it
#for some reason compute needs to be called like that, calling normally was producing some error
nnans <- neuralnet::compute(NN, test)
confusionMatrix(nnans, test_labels))
SamFlynn
  • 369
  • 7
  • 20
  • Thanks for the code! It's working! But I still have two questions- 1.) How to use "hidden" paramter (2.) For 'neuralnet', how shoudl I use the "compute()" function. I understand you don't know much about NNs. Just asking the others out there. – Arun Dec 16 '17 at 17:01
  • @Arun made an edit. Also close the question by accepting it. – SamFlynn Dec 16 '17 at 17:24
  • Two of my questions are still unanswered – Arun Dec 16 '17 at 17:45
  • @Arun hidden and? – SamFlynn Dec 16 '17 at 17:53
  • and using "compute()" function of "neuralnet" package since "predict()" will not work on an object of "neuralnet" – Arun Dec 16 '17 at 17:57
  • @Arun What do you not understand about compute? There is no predict for nueralnet object. You run compute on the model, and test set it gives you results. Just saying, you'll have to be specific about what the problem is with compute, because I just showed how compute is used. – SamFlynn Dec 16 '17 at 18:05
  • you solved it!! Any online tutorial on where I can read more about it apart from doing "?neuralnet" in R? Thanks a lot for your help! – Arun Dec 16 '17 at 19:34
0

1.) Referring to your question how to use the "hidden" parameter, here are some examples.

neuralnet(f, data = m, hidden = c(2,3,2) , learningrate = 0.3)
or
neuralnet(f, data = m, hidden = c(2,2) , learningrate = 0.3)
SBMVNO
  • 582
  • 3
  • 13