Import the usual libraries:
In [1]:
%matplotlib inline
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
Let's say we have numbers that we want to classify. They'll just be 1-dimensional values. Numbers close to 5 will be given the label [0], and numbers close to 2 will be given the label [1], as designed here:
In [2]:
x_label0 = np.random.normal(5, 1, 10)
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)
Out[2]:
Define the hyper-parameters, placeholders, and variables:
In [3]:
learning_rate = 0.001
training_epochs = 1000
X = tf.placeholder("float")
Y = tf.placeholder("float")
w = tf.Variable([0., 0.], name="parameters")
Define the model:
In [4]:
def model(X, w):
return tf.add(tf.multiply(w[1], tf.pow(X, 1)),
tf.multiply(w[0], tf.pow(X, 0)))
Given a model, define the cost function:
In [5]:
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:
In [6]:
train_op = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
correct_prediction = tf.equal(Y, tf.to_float(tf.greater(y_model, 0.5)))
accuracy = tf.reduce_mean(tf.to_float(correct_prediction))
Prepare the session:
In [7]:
sess = tf.Session()
init = tf.global_variables_initializer()
sess.run(init)
Run the training op multiple times on the input data:
In [8]:
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:
In [9]:
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
In [10]:
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()