Hello World in Deep Learning / Keras

1. Artificial Intelligence
for everyone
Build your 1st solution on
Deep Learning
https://sites.google.com/view/AIforEveryone
AI for Good Workshop
March 2019
Session 4

2. Topics in this video
1. Define a purpose
2. Apply the Strategy to jump start
3. Hands-on : Create your 1st Deep Learning solution
Jump start your AI journey !

3. The inspiration
Francois Chollet
Artificial Intelligence Researcher,
Google
ISBN-13: 978-1617294433
Deep Learning with Python
https://www.manning.com/books/deep-learning-with-python

4. Design of Deep Learning experiment
Define problem What is the purpose
Prepare the data
Collect data, spilit data,
vectorize, reshape,
normalize, OHE
Define the network
architecture
architecture
Define loss,
optimizers
Define the metrics,
optimizers, loss function,
and Compile the model
Train the network Train the network
Improve power to
generalize
Evaluate the network,
Hypertune
Test Test, Predict
Deploy Deploy
1
2
3
4
5
6
7
8AIforEveryone.org

5. Define the idea
• App idea: Recognize your hand gesture to
unlock your smartphone (Secure unlock)
• Problem: Recognize hand draw gestures to
unlock your phone
Define your
larger purpose
1

7. visual recognition – image classification

8. Method: 7 steps to develop your Deep Learning model
1: Define problem
Collect data representing
the purpose
2: Format the data
Spilt data, vectorize, reshape,
normalize, OHE
3: Design the network Design the Neural Network
4: Define loss Think of what to optimize for
5: Train the network
Allow the network to learn
patterns in the data
6: Validate & Improve Power of generalization?
7: Predict Predict
Adjust model’s
capacity to learn
“just the patterns”
Develop model
that overfits
Develop 1st model

10. A  B
•

11. X  Y
X = image
Y = Category of image

12. Simplify your goal
Problem: Users draws a hand drawn gestures, your app identifies it.
Simplified Problem: User can draw any number between 0 to 9
Define your
problem
1

15. 1. Define
problem
Collect the data
representing the
world you what
to predict
Load the data Visualize data
•
Labeled dataset

16. Method: 7 steps to develop your Deep Learning model
1: Define problem
Collect data representing
the purpose
2: Format the data
Spilt data, vectorize, reshape,
normalize, OHE
3: Design the network Design the Neural Network
4: Define loss Think of what to optimize for
5: Train the network
Allow the network to learn
patterns in the data
6: Validate & Improve Power of generalization?
7: Predict Predict
Adjust model’s
capacity to learn
“just the patterns”
Develop model
that overfits
Develop 1st model

17. 1. Define
problem
Collect the data
representing the
world you what
to predict
Load the data Visualize data
from keras.datasets import mnist
#1: Define problem, Collect data, Visualize it
# Recognize handwritten numbers
(trainX, trainY), (testX, testY) = mnist.load_data(
from keras.datasets import mnist
(trainX, trainY), (testX, testY) = mnist. load_data()
len(trainX)
keras. datasets. mnist. load_data()
type ( myObject )

18. 1. Define
problem
Collect the data
representing the
world you what
to predict
Load the data Visualize data
from keras.datasets import mnist
Visualize it
exampleImage = trainY [2]
pyplot. imshow (exampleImage )
pyplot. show()
from matplotlib import pyplot
matplotlib. pyplot. imshow ( myArrayImage )

19. Method: 7 steps to develop your Deep Learning model
1: Define problem
Collect data representing
the purpose
2: Format the data
Spilt data, vectorize, reshape,
normalize, OHE
3: Design the network Design the Neural Network
4: Define loss Think of what to optimize for
5: Train the network
Allow the network to learn
patterns in the data
6: Validate & Improve Power of generalization?
7: Predict Predict
Adjust model’s
capacity to learn
“just the patterns”
Develop model
that overfits
Develop 1st model

20. 2. Prepare data
2. Prepare
the data
Spilt data in
training &
testing
Vectortize
as tensors
Reshape Normalize
One Hot
encode
AIforEveryone.org

21. 2. Prepare
the data
Spilt data in
training &
testing
Vectortize
as tensors
Reshape Normalize
One Hot
encode
No of training records = 60000
No of test records = 10000
Shape of training image = (60000, 28, 28)
60000

22. 2. Prepare
the data
Spilt data in
training &
testing
Vectortize
as tensors
Reshape Normalize
One Hot
encode
trainXready = trainX. reshape (800, 28*28)
trainX.shape
Shape of trainX: (800, 28, 28)
Shape of trainXready: (800, 784)
newArray = myArray . reshape (28 * 28)

23. Multi dimensional Array == Tensor
What_is_size_of_array = myArray.shape

24. Reshape() an array

25. 2. Prepare
the data
Spilt data in
training &
testing
Vectortize
as tensors
Reshape Normalize
One Hot
encode
trainXready = temptrainX.astype('float32') / 255
newArray = myArray / 255
afloat = aInteger. astype(‘float32’)

26. One Hot Encoding
Data:
[2]
Data after One Hot Encoding:
[[ 0. 0. 1. 0. 0. 0. 0. 0. 0. 0.]]
yNew = to_categorical (y)

27. 2. Prepare
the data
Spilt data in
training &
testing
Vectortize
as tensors
Reshape Normalize
One Hot
encode2. Prepare
the data
Spilt data in
training &
testing
Vectortize
as tensors
Reshape Normalize
One Hot
encode
myArray = keras.utils. to_categorical ( aInteger)
Data: [2]
Data after OHE : [[ 0. 0. 1. 0. 0. 0. 0. 0. 0. 0.]]

28. 3. Design the network
AIforEveryone.org
Adjust model’s capacity
to learn
“just the patterns”
Develop model that
overfits
Develop 1st model
Data
RepresentationofData

29. Method: 7 steps to develop your Deep Learning model
1: Define problem
Collect data representing
the purpose
2: Format the data
Spilt data, vectorize, reshape,
normalize, OHE
3: Design the network Design the Neural Network
4: Define loss Think of what to optimize for
5: Train the network
Allow the network to learn
patterns in the data
6: Validate & Improve Power of generalization?
7: Predict Predict
Adjust model’s
capacity to learn
“just the patterns”
Develop model
that overfits
Develop 1st model

31. 3. Design the network
#Step 3: Define the network architecture
network = models.Sequential()
layer1 = layers.Dense(512, input_shape=(28*28,), activation='relu')
network.add(layer1)
layer2 = layers.Dense(128, activation='relu')
network.add(layer2)
layer3 = layers.Dense(10, activation='softmax')
network.add(layer3)

32. 3. Design the network
#Step 3: Define the network architecture
network = models.Sequential()
layer1 = layers. Dense (512, input_shape=(28*28,))
network.add(layer1)
layer2 = layers. Dense (128)
network.add(layer2)
layer3 = layers. Dense (10)
network.add(layer3)

34. Method: 7 steps to develop your Deep Learning model
1: Define problem
Collect data representing
the purpose
2: Format the data
Spilt data, vectorize, reshape,
normalize, OHE
3: Design the network Design the Neural Network
4: Define loss Think of what to optimize for
5: Train the network
Allow the network to learn
patterns in the data
6: Validate & Improve Power of generalization?
7: Predict Predict
Adjust model’s
capacity to learn
“just the patterns”
Develop model
that overfits
Develop 1st model

35. 5. Train
4 4 4

36. myModel.fit(X,Y)

37. X  Y
X = image
Y = 5
Sample image
Label of the image

39. 7. Predict
Hand
drawing
samples
 0
 1
 .
 9
Classifier
Input Predicted output

40. y = myModel . Predict(X)

41. X  Y
•

42. 7. Predict
Problem: Users draws a hand drawn gestures , your app predicts it.
predictedResults = network.predict(testXready[1:3])
Predicted Results:
[ 1.35398892e-08 3.91871922e-08 9.99876499e-01 1.22501457e-04
1.77109036e-10 1.30937892e-07 2.99046292e-07 2.58880729e-11
4.58700157e-07 2.09212196e-11]
Ground Truth OHE : [[ 0. 0. 1. 0. 0. 0. 0. 0. 0. 0.]]
Ground Truth : [2]
AIforEveryone.org

43. Let’s it together
• App idea: Recognize your hand gesture to
unlock your smartphone (Secure unlock)
• Biz outcome: Recognize hand draw gestures
to unlock your phone
Test accuracy is: 0.9157
#Step 6: Evaluate the network
metrics_test_loss, metrics_test_accuracy = network.evaluate(testXready, testYready)

44. Type this URL
github.com/
rajagopalmotivate1/
DeepLearningLab
/Play_3_MNIST_Predict_wit
h_Dense.ipynb

45. Design of Deep Learning experiment
Define problem What is the purpose
Prepare the data
Collect data, spilit data,
vectorize, reshape,
normalize, OHE
Define the network
architecture
architecture
Define loss,
optimizers
Define the metrics,
optimizers, loss function,
and Compile the model
Train the network Train the network
Improve power to
generalize
Evaluate the network,
Hypertune
Test Test, Predict
Deploy Deploy
1
2
3
4
5
6
7
8