This is an example of using TensorFlow for simple Linear Regression of 1M points

tfRegExample.py

# imports
import tensorflow as tf
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

# Create linear spaced million points
xData = np.linspace(0.0, 10.0, 1000000, dtype=float)
# Create million random points
noise = np.random.randn(len(xData))

# y =mx + b
b = 5

# First order polynomial
yTrueData = 0.5 * xData + b + noise

# Labeling Data with headers
x_df = pd.DataFrame(data=xData, columns=['X Data'])
y_df = pd.DataFrame(data=yTrueData, columns=['Y'])

# Joining input and output as columns
myData = pd.concat([x_df, y_df], axis=1)

# Get random initial values for b and m
rValue = np.random.randn(2)
m = tf.Variable(rValue[0])
b = tf.Variable(rValue[1])

# Number of points per batch
batchSize = 8
xph = tf.placeholder(dtype=tf.float64, shape=[batchSize])
yph = tf.placeholder(dtype=tf.float64, shape=[batchSize])

# Create the first order model
y_model = m * xph + b

# Loss function
error = tf.reduce_sum(tf.square(yph - y_model))

# Optimizer
optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
train = optimizer.minimize(error)

# Create a session
sess = tf.Session()

# Initialize all variables
init = tf.global_variables_initializer()
sess.run(init)

# Start optimization
numberOfBatches = 1000
for i in range(numberOfBatches):
    randIndices = np.random.randint(0, len(xData), size=batchSize)
    feedDictionary = {xph: xData[randIndices], yph: yTrueData[randIndices]}
    sess.run(train, feed_dict=feedDictionary)

# Get optimal m and b
[mModel, bModel] = sess.run([m, b])
print('m=')
print(mModel)
print('\n')
print('b=')
print(bModel)

# Plotting result
yModel = mModel * xData + bModel

# plotting sample of the generated data
plt.figure()
myData.sample(250).plot(kind='scatter', x='X Data', y='Y')
# plt.plot(xData, yTrueData, '*')
plt.plot(xData, yModel, 'r')
plt.show()

Code result using pyCharm