Ch 03: Concept 01

Linear regression

Import TensorFlow for the learning algorithm. We'll need NumPy to set up the initial data. And we'll use matplotlib to visualize our data.

library(tensorflow)

Define some constants used by the learning algorithm. There are called hyper-parameters.

learning_rate <- 0.01
training_epochs <- 100

Set up fake data that we will use to to find a best fit line

x_train <- seq.int(-1, 1, length.out = 101)
y_train <- 2 * x_train + rnorm(length(x_train), mean = 0, sd=1) * 0.33

Plot the raw data

plot(x_train, y_train)

Set up the input and output nodes as placeholders since the value will be injected by x_train and y_train.

X <- tf$placeholder("float")
Y <- tf$placeholder("float")

Define the model as y = w'*x

model <- function(X, w){
    return(tf$multiply(X, w))
}

Set up the weights variable

w <- tf$Variable(0.0, name="weights")

Define the cost function as the mean squared error

y_model <- model(X, w)
cost <- tf$reduce_mean(tf$square(Y-y_model))

Define the operation that will be called on each iteration of the learning algorithm

train_op <- tf$train$GradientDescentOptimizer(learning_rate)$minimize(cost)

Initialize all variables

sess = tf$Session()
init = tf$global_variables_initializer()
sess$run(init)

Train on each (x, y) pair multiple times

for(epoch in 1:training_epochs){
    for(i in 1:length(x_train)){
        sess$run(train_op, feed_dict=dict(X=x_train[i], Y= y_train[i]))
    }
}

Fetch the value of the learned parameter

w_val <- sess$run(w)
sess$close()

Visualize the best fit curve

plot(x_train, y_train)
y_learned <- x_train*w_val
lines(x_train, y_learned, col='red')