linear_reg_classification.py

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

#  1-dimensional values. 
# Numbers close to 5 will be given the label [0], and 
# Numbers close to 2 will be given the label [1].

x_label0 = np.random.normal(5, 1, 10)
# outliers: 
# x_label0 = np.append(np.random.normal(5, 1,9), 20)

x_label1 = np.random.normal(2, 1, 10)
xs = np.append(x_label0, x_label1)
labels = [0.] * len(x_label0) + [1.] * len(x_label1)

plt.scatter(xs, labels)

# Define the hyper-parameters, placeholders, and variables:
learning_rate = 0.001
training_epochs = 1000

X = tf.placeholder("float")
Y = tf.placeholder("float")

w = tf.Variable([0., 0.], name="parameters")


def model(X, w):
    return tf.add(tf.multiply(w[1], tf.pow(X, 1)),
                  tf.multiply(w[0], tf.pow(X, 0)))

# Define the cost function
y_model = model(X, w)
cost = tf.reduce_sum(tf.square(Y-y_model))

# Set up the training op, and also introduce a couple ops to calculate some metrics, such as accuracy:
train_op = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
# When the model response is greater than 0.5, it should be a positive label, and vice versa.
correct_prediction = tf.equal(Y, tf.to_float(tf.greater(y_model, 0.5)))
accuracy = tf.reduce_mean(tf.to_float(correct_prediction))

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

# Run learning operation multiple times.
for epoch in range(training_epochs):
    sess.run(train_op, feed_dict={X: xs, Y: labels})
    current_cost = sess.run(cost, feed_dict={X: xs, Y: labels})
    if epoch % 100 == 0:
        print(epoch, current_cost)

# Show some final metrics/results:
w_val = sess.run(w)
print('learned parameters', w_val)

print('accuracy', sess.run(accuracy, feed_dict={X: xs, Y: labels}))

sess.close()

# Plot the learned function to show the best-fit line
all_xs = np.linspace(0, 10, 100)
plt.plot(all_xs, all_xs*w_val[1] + w_val[0])
plt.scatter(xs, labels)
plt.show()