You are on page 1of 18

TensorFlow and

Keras
Day 2
Communication & Questions

Mics are disabled but chat is open


Post questions at anytime
Responses will be during a logical break
MNIST Dataset
MNIST Dataset

 Modified National Institute of Standards and Technology database


 Handwritten digits
 Black and white images
 28x28 pixel (Min 0, Max 255)
 60,000 training images
 10,000 testing images

DEMO
Activation Functions

 Step functions – 0 or 1
 Sigmoid – Between 0 and 1
Sick

Healthy
Sick

Healthy
Sick

Healthy
Hyperbolic Tangent / tanh
Rectified linear unit (ReLU)
f(x) = max(0,x)
Loss and Optimizer

 Loss
 how poorly the model is performing
 Objective is to minimize loss
 Optimizer
 The specific algorithm to minimize the loss
Loss

 mse – Good for regression


 sparse_categorical_crossentropy
 When each sample belongs exactly to one class
 categorical_crossentropy
when one sample can have multiple classes
Optimizer

 Adam – most common


SGD

RMSprop

Adadelta

Adagrad

Adamax

Nadam

Ftrl

You might also like