>TensorFlow and deep learning_

without a PhD

deep deep
Science ! Code ...

#Tensorflow @martin_gorner
 Hello World: handwritten digits classification - MNIST

?
MNIST = Mixed National Institute of Standards and Technology - Download the dataset at http://yann.lecun.com/exdb/mnist/
Very simple model: softmax classification

784 pixels
28x28
pixels
... weighted sum of all
pixels + bias

softmax ...

0 1 2 9
neuron outputs

Data & Analytics 3

 In matrix notation, 100 images at a time
10 columns
w0,0 w0,1 w0,2 w0,3 … w0,9
w1,0 w1,1 w1,2 w1,3 … w1,9 broadcast
w2,0 w2,1 w2,2 w2,3 … w2,9
w3,0 w3,1 w3,2 w3,3 … w3,9

784 lines
w4,0 w4,1 w4,2 w4,3 … w4,9
x w5,0 w5,1 w5,2 w5,3 … w5,9
x w6,0 w6,1 w6,2 w6,3 … w6,9
x
X : 100 images, x
x
w7,0 w7,1 w7,2 w7,3 … w7,9
w8,0 w8,1 w8,2 w8,3 … w8,9
one per line, x …
x
flattened x
w783,0 w783,1 w783,2 … w783,9

L0,0 L0,1 L0,2 L0,3 ……L0,9 + b 0 b1 b2 b3 … b9

L1,0 L1,1 L1,2 L1,3 … L1,9
L2,0 L2,1 L2,2 L2,3 … L2,9
L3,0 L3,1 L3,2 L3,3 … L3,9 + Same biases
L4,0 L4,1 L4,2 L4,3 … L4,9
… on all lines
L99,0 L99,1 L99,2 … L99,9
784 pixels
 Softmax, on a batch of images

Predictions Images Weights Biases

Y[100, 10] X[100, 748] W[748,10] b[10]

broadcast
applied line matrix multiply
on all lines
by line
tensor shapes in [ ]

Data & Analytics 5

Now in TensorFlow (Python)

tensor shapes: X[100, 748] W[748,10] b[10]

Y = tf.nn.softmax(tf.matmul(X, W) + b)

broadcast
matrix multiply
on all lines

Data & Analytics 6

Success ?
0 1 2 3 4 5 6 7 8 9
0 0 0 0 0 0 1 0 0 0

actual probabilities, “one-hot” encoded

Cross entropy:
this is a “6”
computed probabilities
0.1 0.2 0.1 0.3 0.2 0.1 0.9 0.2 0.1 0.1
0 1 2 3 4 5 6 7 8 9

Data & Analytics 7

Demo
92%
TensorFlow - initialisation

import tensorflow as tf this will become the batch size, 100

X = tf.placeholder(tf.float32, [None, 28, 28, 1])

W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10])) 28 x 28 grayscale images

init = tf.initialize_all_variables()

Training = computing variables W and b

Data & Analytics 10

TensorFlow - success metrics
flattening images
# model
Y = tf.nn.softmax(tf.matmul(tf.reshape(X, [-1, 784]), W) + b)
# placeholder for correct answers
Y_ = tf.placeholder(tf.float32, [None, 10])
“one-hot” encoded
# loss function
cross_entropy = -tf.reduce_sum(Y_ * tf.log(Y))
“one-hot” decoding
# % of correct answers found in batch
is_correct = tf.equal(tf.argmax(Y,1), tf.argmax(Y_,1))
accuracy = tf.reduce_mean(tf.cast(is_correct, tf.float32))

Data & Analytics 11

TensorFlow - training

learning rate

optimizer = tf.train.GradientDescentOptimizer(0.003)
train_step = optimizer.minimize(cross_entropy)

loss function

Data & Analytics 12

 TensorFlow - run !
sess = tf.Session() running a Tensorflow
sess.run(init) computation, feeding
placeholders
for i in range(1000):
# load batch of images and correct answers
batch_X, batch_Y = mnist.train.next_batch(100)
train_data={X: batch_X, Y_: batch_Y}

# train
sess.run(train_step, feed_dict=train_data)

# success ?
Tip: a,c = sess.run([accuracy, cross_entropy], feed_dict=train_data)
do this
every 100 # success on test data ?
iterations test_data={X: mnist.test.images, Y_: mnist.test.labels}
a,c = sess.run([accuracy, cross_entropy, It], feed=test_data)
 TensorFlow - full python code
training step

import tensorflow as tf
initialisation optimizer = tf.train.GradientDescentOptimizer(0.003)
train_step = optimizer.minimize(cross_entropy)

X = tf.placeholder(tf.float32, [None, 28, 28, 1]) sess = tf.Session()

W = tf.Variable(tf.zeros([784, 10])) sess.run(init)
b = tf.Variable(tf.zeros([10]))
init = tf.initialize_all_variables() model for i in range(10000):
# load batch of images and correct answers
# model batch_X, batch_Y = mnist.train.next_batch(100)
Y=tf.nn.softmax(tf.matmul(tf.reshape(X,[-1, 784]), W) + b) train_data={X: batch_X, Y_: batch_Y}

# placeholder for correct answers # train

Y_ = tf.placeholder(tf.float32, [None, 10]) sess.run(train_step, feed_dict=train_data) Run
# loss function
success metrics # success ? add code to print it
cross_entropy = -tf.reduce_sum(Y_ * tf.log(Y)) a,c = sess.run([accuracy, cross_entropy], feed=train_data)

# % of correct answers found in batch # success on test data ?

is_correct = tf.equal(tf.argmax(Y,1), tf.argmax(Y_,1)) test_data={X:mnist.test.images, Y_:mnist.test.labels}
accuracy = tf.reduce_mean(tf.cast(is_correct,tf.float32)) a,c = sess.run([accuracy, cross_entropy], feed=test_data)

Data & Analytics 14

Cookbook

Softmax
Cross-entropy
Mini-batch
ML
 e e p !

|
Go d

|
|
|
 Let’s try 5 fully-connected layers !
overkill
;-)

784

200

100
sigmoïd function
60

30

10 softmax
0 1 2 ... 9

Data & Analytics 17

TensorFlow - initialisation

K = 200
L = 100 weights initialised
M = 60 with random values
N = 30

W1 = tf.Variable(tf.truncated_normal([28*28, K] ,stddev=0.1))
B1 = tf.Variable(tf.zeros([K])

W2 = tf.Variable(tf.truncated_normal([K, L], stddev=0.1))

B2 = tf.Variable(tf.zeros([L])
W3 = tf.Variable(tf.truncated_normal([L, M], stddev=0.1))
B3 = tf.Variable(tf.zeros([M])
W4 = tf.Variable(tf.truncated_normal([M, N], stddev=0.1))
B4 = tf.Variable(tf.zeros([N])
W5 = tf.Variable(tf.truncated_normal([N, 10], stddev=0.1))
B5 = tf.Variable(tf.zeros([10]))

Data & Analytics 18

TensorFlow - the model

weights and biases

X = tf.reshape(X, [-1, 28*28])

Y1 = tf.nn.sigmoid(tf.matmul(X, W1) + B1)

Y2 = tf.nn.sigmoid(tf.matmul(Y1, W2) + B2)
Y3 = tf.nn.sigmoid(tf.matmul(Y2, W3) + B3)
Y4 = tf.nn.sigmoid(tf.matmul(Y3, W4) + B4)
Y = tf.nn.softmax(tf.matmul(Y4, W5) + B5)

Data & Analytics 19

Demo - slow start ?

Data & Analytics 20

Relu !
RELU
RELU = Rectified Linear Unit

Y = tf.nn.relu(tf.matmul(X, W) + b)
Data & Analytics 22
RELU

Data & Analytics 23

Demo - noisy accuracy curve ?

yuck!
Slow down . . .
Learning
rate decay
Learning rate decay

Learning rate 0.003 at start then dropping exponentially to 0.0001

Data & Analytics 26
Demo - dying neurons

Dy
ing
...

Data & Analytics 27

Dropout
 Dropout

pkeep =
tf.placeholder(tf.float32) TRAINING EVALUATION
pkeep=0.75 pkeep=1

Yf = tf.nn.relu(tf.matmul(X, W) + B)
Y = tf.nn.dropout(Yf, pkeep)
Data & Analytics 29
Dropout

Dy
ing
...

D
ea
d
with dropout
Data & Analytics 30
Demo
 98%
Data & Analytics 32
All the party tricks

98.2%
97.9%
sustained
peak

Sigmoid,decaying
RELU, learning
learningrate
learning
rate= =0.003
0.003
rate 0.003 -> 0.0001 and dropout 0.75

Data & Analytics 33

Overfitting
Cross-entropy loss

Overfitting

Data & Analytics 34

Overfitting Too many
neurons
?!?

BAD
Network
Not
enough
DATA
 Convolutional layer
convolutional
subsampling

+padding convolutional
subsampling
convolutional
subsampling
stride

W1[4, 4, 3]

W2[4, 4, 3] W[4, 4, 3, 2]
filter input output
size channels channels

Data & Analytics 36

Hacker’s tip

ALL
Convolu-
tional
Convolutional neural network
+ biases on
all layers
28x28x1
convolutional layer, 4 channels
W1[5, 5, 1, 4] stride 1
28x28x4
convolutional layer, 8 channels
14x14x8 W2[4, 4, 4, 8] stride 2

convolutional layer, 12 channels

7x7x12
W3[4, 4, 8, 12] stride 2

200 fully connected layer W4[7x7x12, 200]

10 softmax readout layer W5[200, 10]
Tensorflow - initialisation
filter input output
K=4
size channels channels
L=8
M=12

W1 = tf.Variable(tf.truncated_normal([5, 5, 1, K] ,stddev=0.1))
B1 = tf.Variable(tf.ones([K])/10)
W2 = tf.Variable(tf.truncated_normal([5, 5, K, L] ,stddev=0.1))
B2 = tf.Variable(tf.ones([L])/10)
W3 = tf.Variable(tf.truncated_normal([4, 4, L, M] ,stddev=0.1))
B3 = tf.Variable(tf.ones([M])/10)
weights initialised
N=200 with random values

W4 = tf.Variable(tf.truncated_normal([7*7*M, N] ,stddev=0.1))
B4 = tf.Variable(tf.ones([N])/10)
W5 = tf.Variable(tf.truncated_normal([N, 10] ,stddev=0.1))
B5 = tf.Variable(tf.zeros([10])/10)
Tensorflow - the model

input image batch weights stride biases

X[100, 28, 28, 1]

Y1 = tf.nn.relu(tf.nn.conv2d(X, W1, strides=[1, 1, 1, 1], padding='SAME') + B1)

Y2 = tf.nn.relu(tf.nn.conv2d(Y1, W2, strides=[1, 2, 2, 1], padding='SAME') + B2)
Y3 = tf.nn.relu(tf.nn.conv2d(Y2, W3, strides=[1, 2, 2, 1], padding='SAME') + B3)

YY = tf.reshape(Y3, shape=[-1, 7 * 7 * M])

Y3 [100, 7, 7, 12]
flatten all values for
Y4 = tf.nn.relu(tf.matmul(YY, W4) + B4) fully connected layer
YY [100, 7x7x12]
Y = tf.nn.softmax(tf.matmul(Y4, W5) + B5)

Data & Analytics 40

Demo
 98.9%

Data & Analytics 42

WTFH ???

???

Data & Analytics 43

Bigger convolutional network + dropout
+ biases on
all layers
28x28x1
convolutional layer, 6 channels
28x28x6 W1[6, 6, 1, 6] stride 1

convolutional layer, 12 channels

14x14x12 W2[5, 5, 6, 12] stride 2

7x7x24 convolutional layer, 24 channels

W3[4, 4, 12, 24] stride 2
+DROPOUT
200 fully connected layer W4[7x7x24, 200]
p=0.75
10 softmax readout layer W5[200, 10]
Demo
 99.3%

Data & Analytics 46

YEAH !

with dropout
Data & Analytics 47
 Learning
Relu !
rate decay

Dropout

Softmax
Cross-entropy ALL
Mini-batch Convolu-
Go deep

Overfitting Too many

tional ?!?
neurons
|
|
BAD
Network
Not
enough
DATA
Cartoon images copyright: alexpokusay / 123RF stock photos
 Have fun !
- cloud.google.com

All code snippets are on

Cloud ML ALPHA GitHub:
github.com/martin-
your TensorFlow models gorner/tensorflow-
trained in Google’s cloud, mnist-tutorial
fast.
This presentation:

Pre-trained models:
goo.gl/pHeXe7

tensorflow.org Cloud Vision API

Cloud Speech API ALPHA

Martin Görner
Google Developer relations
That’s all
Google Translate API
@martin_gorner folks...
Data & Analytics 49
plus.google.com/+MartinGorner
Workshop

Keyboard shortcuts for the

visualisation GUI:

1 ......... display 1st graph only

2 ......... display 2nd graph only
3 ......... display 3rd graph only
4 ......... display 4th graph only
5 ......... display 5th graph only
6 ......... display 6th graph only
7 ......... display graphs 1 and 2
8 ......... display graphs 4 and 5
9 ......... display graphs 3 and 6
ESC or 0 .. back to displaying all graphs

SPACE ..... pause/resume

O ......... box zoom mode (then use mouse)
H ......... reset all zooms
Ctrl-S .... save current image

Data & Analytics 50

 Starter code and solutions:
Workshop github.com/martin-gorner/tensorflow-mnist-tutorial
1. Theory (sit back and listen) 6. Practice
Replace all your sigmoids with RELUs. Test. Then add
Softmax classifier, mini-batch, cross-entropy and how to learning rate decay from 0.003 to 0.0001 using the
implement them in Tensorflow (slides 1-14) formula lr = lrmin+(lrmax-lrmin)*exp(-i/2000).
Solution in: mnist_2.1_five_layers_relu_lrdecay.py

2. Practice 7. Practice (if time allows)

Open file: mnist_1.0_softmax.py
Run it, play with the visualisations (see instructions on Add dropout on all layers using a value between 0.5 and
previous slide), read and understand the code as well as 0.8 for pkeep.
the basic structure of a Tensorflow program. Solution in: mnist_2.2_five_layers_relu_lrdecay_dropout.py

3. Theory (sit back and listen) 8. Theory (sit back and listen)

Hidden layers, sigmoid activation function (slides 16-19) Convolutional networks (slides 36-42)

4. Practice 9. Practice
Start from the file you have and add one or two hidden
layers. Use cross_entropy_with_logits to avoid Replace your model with a convolutional network,
numerical instabilities with log(0). without dropout.
Solution in: mnist_2.0_five_layers_sigmoid.py Solution in: mnist_3.0_convolutional.py

5. Theory (sit back and listen) 10. Practice (if time allows)

The neural network toolbox: RELUs, learning rate decay, Try a bigger neural network (good hyperparameters on
dropout, overfitting (slides 20-35) slide 44) and add dropout on the last layer.
Solution in: mnist_3.0_convolutional_bigger_dropout.py