TensorFlow

TensorFlow is an open-source software library for dataflow programming across a range of tasks. It is a symbolic math library, and is also used for machine learning applications such as neural networks.

Introduction

Tensors

• N-dimensional arrays
• Measured by "rank"
• All elements are same datatype

# Rank 0:
[1]
# Rank 1:
[1][2][3]
# Rank 2:
[1][2][3]
[4][5][6]
# Rank 3 (3D):
[1][2][3]
 [4][5][6]
  [7][8][9]

Tensor operations

• Addition and subtraction
• Multiplication and Division
• Matrix multiplication
• Dot product
• Transpose

[1 2  3  4 ]   [1 5 9 ]T
|5 6  7  8 | = |2 6 10|
[9 10 11 12]   |3 7 11|
               [4 8 12]

TensorFlow building blocks

• Lower level
  • Tensors
  • Operations
  • Graphs and sessions
• Higher level
  • Loss functions
  • Optimizers
  • Layers
  • Estimators

Loss functions

• Differentiatable functions that measure differences/error between true and predicted values
• Common types:
  • Mean squared error (MSE)
  • Log loss
  • Cosine distance
  • Cross entropy

Optimizers

• Optimizers are algorithms that minimize the loss (or error) of a model
• Local minimum vs. global minimum
• Built-in optimizers inherit from the Optimizer class
• Common types:
  • Gradient descent
  • Adam
  • RMSProp
  • Adagrad
  • Momentum
  • Adadelta

Layers

• What are they?
  • Composed of tensors and operations forming the model
  • Generally connected in series
  • Pre-made functions for creating layers in a model
• Common types:
  • Input
  • Convolutional (1d, 2d, 3d)
  • Pooling
  • Dropout
  • Dense

Estimators

• Training
• Evaluation
• Prediction
• Build Graph

Examples

• Basic #1

import tensorflow as tf

m = tf.constant(3.0, name='m')
b = tf.constant(1.5, name='b')
x = tf.placeholder(dtype='float32', name='x')
y = m*x + b

sess = tf.Session()

y.eval({x: 2}, session=sess)  # => 7.5

• Basic #2

import tensorflow as tf

M = tf.constant([[1,2], [3,4]], dtype='float32')
v = tf.constant([5,6], dtype='float32')
sess = tf.Session()

sess.run(M + v)
# array([[ 6.,  8.],
#       [ 8., 10.]], dtype=float32)

sess.run(M * v)
# array([[ 5., 12.],
#        [15., 24.]], dtype=float32)

sess.run(tf.matmul(M, tf.reshape(v, [2, 1])))
# array([[17.],
#        [39.]], dtype=float32)

Machine learning

Machine learning lifecycle

1. Define objective
2. Collect data
3. Data cleaning
4. Exploratory Data Analysis (EDA)
5. Data processing
6. Train/evaluate models
7. Deploy
8. Monitor results

Machine learning lifecycle example

• 1. Define objective

Infer how IQ, years experince, and age affects income using linear model

• 2. Collect data

import tensorflow as tf
import numpy as np

import pandas as pd
from pandas import DataFrame as DF

# Create dataset
np.random.seed(555)
X1 = np.random.normal(100, 15, 200).astype(int) # IQ
X2 = np.random.normal(10, 4.5, 200) # years experience
X3 = np.random.normal(32, 4, 200).astype(int) # age
dob = np.datetime64('2017-10-31') - 365*X3
b = 5
er = np.random.normal(0, 1.5, 200) # noise/error

Y = np.array([0.3*x1 + 1.5*x2 + 0.83*x3 + b + e for x1,x2,x3,e in zip(X1,X2,X3,er)])

• 3. Data Cleaning

cols = ['iq', 'years_experience', 'dob']
df = DF(list(zip(X1,X2,dob)), columns=cols)
df['income'] = Y
df.info()
# <class 'pandas.core.frame.DataFrame'>
# RangeIndex: 200 entries, 0 to 199
# Data columns (total 4 columns):
# iq                  200 non-null int64
# years_experience    200 non-null float64
# dob                 200 non-null datetime64[ns]
# income              200 non-null float64
# dtypes: datetime64[ns](1), float64(2), int64(1)
# memory usage: 6.3 KB

df.describe()
# Remove any negative values for years of experience
df = df[df.years_experience >= 0]
df.describe() # no more negative values

• 4. EDA

df.describe(include=['datetime64'])

import matplotlib.pyplot as plt
%matplotlib inline

pd.plotting.scatter_matrix(df, figsize=(16,9));

External links

• Official website
• TensorFlow on GitHub