#!/usr/bin/python3
 import pandas as pd
 import tensorflow as tf
 import numpy
 import os

 os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

 rnd             = numpy.random
 learning_rate   = .01
 training_epochs = 2000
 chkpoint        = 250

 train_df = pd.read_csv("odometer.csv")

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

 W = tf.Variable(rnd.randn(), name="weight")
 b = tf.Variable(rnd.randn(), name="bias")

   # model: Y = X*W + b
 pred = tf.add(tf.multiply(X, W), b)

   # mean squared error
 total = len(train_df.index)
 cost  = tf.reduce_sum(
     tf.pow(pred-Y, 2))/(2*total)

   # normalize training set
 nn_offset=int(train_df[['date']].min())
 nn_div=int(train_df[['date']].max() -
        train_df[['date']].min())
 print("norm_off=", nn_offset)
 print("norm_mult=", nn_div)
 train_df['date'] -= nn_offset
 train_df['date'] /= nn_div

 opt = tf.train.GradientDescentOptimizer(
     learning_rate).minimize(cost)

 init = tf.global_variables_initializer()

   # tensorflow session
 with tf.Session() as sess:
     sess.run(init)
     for epoch in range(training_epochs):
         for ix, row in train_df.iterrows():
             sess.run(opt, feed_dict={
                 X: row['date'],
                 Y: row['miles']})
         if epoch % chkpoint == 0:
             c=sess.run(cost, feed_dict={
                 X: train_df['date'],
                 Y: train_df['miles']})
             print("W=", sess.run(W),
                   "b=", sess.run(b),
                   "cost=", c)