tf_linreg.py

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

## 1. definicija računskog grafa
# podatci i parametri

X = tf.placeholder(tf.float32, [None])
Y_ = tf.placeholder(tf.float32, [None])
a = tf.Variable(0.0, name="a")
b = tf.Variable(0.0, name="b")

# afini regresijski model
Y = a * X + b

# kvadratni gubitak
loss = (Y-Y_)**2

# optimizacijski postupak: gradijentni spust
trainer = tf.train.GradientDescentOptimizer(0.1)
train_op = trainer.minimize(loss)
grad = trainer.compute_gradients(loss)

debug = []
for tens, var in grad:
    debug.append(tf.Print(tens, [tens], var.name))
train_op = tf.group(*debug, trainer.apply_gradients(grad))

## 2. inicijalizacija parametara
sess = tf.Session()
sess.run(tf.initialize_all_variables())

## 3. učenje
# neka igre počnu!
for i in range(100):
  val_loss, _, val_a,val_b = sess.run([loss, train_op, a,b],
      feed_dict={X: [1,2], Y_: [3,5]})
  print(i,val_loss, val_a,val_b)