Deep Learning in your Browser: powered by WebGL

Deep Learning and WebGL
WebGL Meetup 03/27/2018
GTC SJ Convention Center
Oswald Campesato
oswald@perceptrons.io
ocampesato@yahoo.com

Overview
intro to AI/ML/DL
linear regression
activation/cost functions
gradient descent
back propagation
hyper-parameters
what are CNNs
JS/WebGL and DL

Gartner 2017: Deep Learning (YES!)

The Official Start of AI (1956)

Neural Network with 3 Hidden Layers

AI/ML/DL: How They Differ
Traditional AI (20th century):
based on collections of rules
Led to expert systems in the 1980s
The era of LISP and Prolog

Machine Learning:
Started in the 1950s (approximate)
Alan Turing and “learning machines”
Data-driven (not rule-based)
Many types of algorithms
Involves optimization

Deep Learning:
Started in the 1950s (approximate)
The “perceptron” (basis of NNs)
Data-driven (not rule-based)
large (even massive) data sets
Involves neural networks (CNNs: ~1970s)
Lots of heuristics
Heavily based on empirical results

AI/ML/DL: Commonality
All of them involve a model
A model represents a system
Goal: a good predictive model
The model is based on:
Many rules (for AI)
data and algorithms (for ML)
large sets of data (for DL)

A Basic Model in Machine Learning
Let’s perform the following steps:
1) Start with a simple model (2 variables)
2) Generalize that model (n variables)
3) See how it might apply to a NN

Linear Regression
One of the simplest models in ML
Fits a line (y = m*x + b) to data in 2D
Finds best line by minimizing MSE:
m = average of x values (“mean”)
b also has a closed form solution

Linear Regression in 2D: example

Linear Regression: example #1
One feature (independent variable):
X = number of square feet
Predicted value (dependent variable):
Y = cost of a house
A very “coarse grained” model
We can devise a much better model

Linear Regression: example #2
Multiple features:
X1 = # of square feet
X2 = # of bedrooms
X3 = # of bathrooms (dependency?)
X4 = age of house
X5 = cost of nearby houses
X6 = corner lot (or not): Boolean
a much better model (6 features)

Linear Multivariate Analysis
General form of multivariate equation:
Y = w1*x1 + w2*x2 + . . . + wn*xn + b
w1, w2, . . . , wn are numeric values
x1, x2, . . . , xn are variables (features)
Properties of variables:
Can be independent (Naïve Bayes)
weak/strong dependencies can exist

Neural Networks: equations
Node “values” in first hidden layer:
N1 = w11*x1+w21*x2+…+wn1*xn
N2 = w12*x1+w22*x2+…+wn2*xn
N3 = w13*x1+w23*x2+…+wn3*xn
. . .
Nn = w1n*x1+w2n*x2+…+wnn*xn
Similar equations for other pairs of layers

Neural Networks: Matrices
From inputs to first hidden layer:
Y1 = W1*X + B1 (X/Y1/B1: vectors; W1: matrix)
From first to second hidden layers:
From second to third hidden layers:
 Apply an “activation function” to y values

Neural Networks (general)
Multiple hidden layers:
Layer composition is your decision
Activation functions: sigmoid, tanh, RELU
https://en.wikipedia.org/wiki/Activation_function
Back propagation (1980s)
https://en.wikipedia.org/wiki/Backpropagation
=> Initial weights: small random numbers

The sigmoid Activation Function

What’s the “Best” Activation Function?
Initially: sigmoid was popular
Then: tanh became popular
Now: RELU is preferred (better results)
Softmax: for FC (fully connected) layers
NB: sigmoid and tanh are used in LSTMs

CNNs versus RNNs
CNNs (Convolutional NNs):
Good for image processing
2000: CNNs processed 10-20% of all checks
=> Approximately 60% of all NNs
RNNs (Recurrent NNs):
Good for NLP and audio

CNNs: Convolution Calculations
https://docs.gimp.org/en/plug-in-
convmatrix.html

CNNs: Convolution Matrices (examples)
Sharpen:
Blur:

CNNs: Convolution Matrices (examples)
Edge detect:
Emboss:

How do Filters Work?
 compute “inner product” with an image
The “padding” parameter
Filter size: 3x3, 5x5, or 1x1 are common
Stride values: 1 or 2 (possibly greater)
“parallelize” multiple filters then merge

CNN in Python/Keras (fragment)
 from keras.models import Sequential
 from keras.layers.core import Dense, Dropout, Flatten, Activation
 from keras.layers.convolutional import Conv2D, MaxPooling2D
 from keras.optimizers import Adadelta
 input_shape = (3, 32, 32)
 nb_classes = 10
 model = Sequential()
 model.add(Conv2D(32, (3, 3), padding='same’,
input_shape=input_shape))
 model.add(Activation('relu'))
 model.add(Conv2D(32, (3, 3)))
 model.add(Activation('relu'))
 model.add(MaxPooling2D(pool_size=(2, 2)))
 model.add(Dropout(0.25))

What is TensorFlow?
An open source framework for ML and DL
A “computation” graph
Created by Google (released 11/2015)
Evolved from Google Brain
Linux and Mac OS X support (VM for Windows)
TF home page: https://www.tensorflow.org/

What is TensorFlow?
Support for Python, Java, C++
Desktop, server, mobile device (TensorFlow Lite)
CPU/GPU/TPU support
Visualization via TensorBoard
Can be embedded in Python scripts
Installation: pip install tensorflow
TensorFlow cluster:
https://www.tensorflow.org/deploy/distributed

TensorFlow Use Cases (Generic)
Image recognition
Computer vision
Voice/sound recognition
Time series analysis
Language detection
Language translation
Text-based processing
Handwriting Recognition

What is TensorFlow?
Graph: graph of operations (DAG)
Sessions: contains Graph(s)
lazy execution (default)
operations in parallel (default)
Nodes: operators/variables/constants
Edges: tensors
=> graphs are split into subgraphs and
executed in parallel (or multiple CPUs)

TensorFlow Graph Execution
Execute statements in a tf.Session() object
Invoke the “run” method of that object
“eager” execution is now possible
Not part of the mainline yet
Installation: pip install tf-nightly

What is a Tensor?
TF tensors are n-dimensional arrays
TF tensors are very similar to numpy ndarrays
scalar number: a zeroth-order tensor
vector: a first-order tensor
matrix: a second-order tensor
3-dimensional array: a 3rd order tensor
https://dzone.com/articles/tensorflow-simplified-
examples

TensorFlow “primitive types”
 tf.constant: initialized immediately
 tf.placeholder (a function):
+ initial value is not required
+ assigned value via feed_dict at run time
+ are not modified during training
 tf.Variable (a class):
+ initial value is required
+ updated during training
+ in-memory buffer (saved/restored from disk)
+ can be shared between works (distributed env)

TensorFlow: constants (immutable)
 import tensorflow as tf # tf-const.py
 aconst = tf.constant(3.0)
 print(aconst)
# output: Tensor("Const:0", shape=(), dtype=float32)
 sess = tf.Session()
 print(sess.run(aconst))
# output: 3.0
 sess.close()
 # => there's a better way…

TensorFlow: constants
import tensorflow as tf # tf-const2.py
aconst = tf.constant(3.0)
print(aconst)
Automatically close “sess”
with tf.Session() as sess:
 print(sess.run(aconst))

TensorFlow Arithmetic
import tensorflow as tf # basic1.py
a = tf.add(4, 2)
b = tf.subtract(8, 6)
c = tf.multiply(a, 3)
d = tf.div(a, 6)
with tf.Session() as sess:
print(sess.run(a)) # 6
print(sess.run(b)) # 2
print(sess.run(c)) # 18
print(sess.run(d)) # 1

TensorFlow Arithmetic Methods
import tensorflow as tf #tf-math-ops.py
PI = 3.141592
sess = tf.Session()
print(sess.run(tf.div(12,8)))
print(sess.run(tf.floordiv(20.0,8.0)))
print(sess.run(tf.sin(PI)))
print(sess.run(tf.cos(PI)))
print(sess.run(tf.div(tf.sin(PI/4.), tf.cos(PI/4.))))

TensorFlow Arithmetic Methods
Output from tf-math-ops.py:
1
2.0
6.27833e-07
-1.0
1.0

TF placeholders and feed_dict
import tensorflow as tf # tf-var-multiply.py
a = tf.placeholder("float")
b = tf.placeholder("float")
c = tf.multiply(a,b)
# initialize a and b:
feed_dict = {a:2, b:3}
# multiply a and b:
with tf.Session() as sess:
print(sess.run(c, feed_dict))

Deep Learning JavaScript Toolkits
Tensorfire ( https://tenso.rs )
deeplearn.js ( https://deeplearnjs.org )
 keras-js ( https://github.com/transcranial/keras-js )
MXNetJS ( https://github.com/dmlc/mxnet.js )
ML.js ( https://github.com/mljs/ml )
Jeeliz ( https://jeeliz.com )

What is TensorFire?
uses WebGL to run NNs in browsers
TensorFire has two parts:
1) a GLSL-based language to write massively parallel
WebGL shaders that operate on 4D tensors
2) a high-level library for importing models trained with
Keras or TensorFlow

What is TensorFire?
imitates the structure of TensorFlow & numpy
uses WebGL JavaScript API for 2D/3D graphics
can also perform computations on the GPU
hence it offers significant performance

What is TensorFire?
it supports low-precision quantized tensors
supports browsers that don't fully support the
OES_texture_float extension
https://petewarden.com/2015/05/23/why-are-eight-
bits-enough-for-deep-neural-networks

What is TensorFire?
works on any GPU (with/without CUDA)
on computers with AMD graphics
the latest 2016 Retina Macbook Pro
https://github.com/tensorfire

What is TensorFire?
the Tensorfire demo app:
1) runs the style-transfer neural network
2) renders in your browser
3) as fast as CPU TensorFlow on a desktop
https://tenso.rs/demos/fast-neural-style/
https://tenso.rs/demos/rock-paper-scissors/

TensorFire Code Sample
Go to ‘demos’ directory
Go to backprop subdirectory
Open backprop.html
https://github.com/tensorfire/tensorfire/blob/master/demos/bac
kprop/backprop.html

What is deeplearn.js?
an open-source library
developed by the Google Brain PAIR team
API closely mirrors TensorFlow eager API
can be used with TypeScript (preferred)
can be used with plain JavaScript

train NNs in a browser OR
run pre-trained models (inference mode)
A hardware-accelerated machine intelligence
library for the web

install via yarn/npm:
yarn add deeplearn or npm install deeplearn
available as a global variable named dl
Use a version (e.g. 0.5.0) instead of @latest:
<script src="https://cdn.jsdelivr.net/npm/deeplearn@0.5.0">
</script>
OR
<script src="https://unpkg.com/deeplearn@latest"></script>

 import * as dl from 'deeplearn';
 const a = dl.tensor1d([1, 2, 3]);
 const b = dl.scalar(2);
 const result = a.add(b);
 //Float(32Array([3,4,5])
 result.data().then(data => console.log(data));
 // Alternatively can use a blocking call:
 //Float(32Array([3,4,5])
 console.log(result.dataSync());

deeplearn.js Code Sample
Go to ‘demos’ directory
Go to mnist_eager subdirectory
Open model.ts
https://github.com/PAIR-
code/deeplearnjs/blob/master/demos/mnist_eager/model.ts

deeplearn.js Useful Links
port TensorFlow models to deeplearn.js:
https://deeplearnjs.org/demos/mnist/mnist.html
API reference:
https://deeplearnjs.org/docs/api/index.html
Other samples:
https://deeplearnjs.org/index.html#demos

Deep Learning, JS, and GPU Support
Matrix comparison of JS toolkits:
 https://towardsdatascience.com/gpu-accelerated-neural-
networks-in-javascript-195d6f8e69ef

About Me: Recent Books
1) HTML5 Canvas and CSS3 Graphics (2013)
2) jQuery, CSS3, and HTML5 for Mobile (2013)
3) HTML5 Pocket Primer (2013)
4) jQuery Pocket Primer (2013)
5) HTML5 Mobile Pocket Primer (2014)
6) D3 Pocket Primer (2015)
7) Python Pocket Primer (2015)
8) SVG Pocket Primer (2016)
9) CSS3 Pocket Primer (2016)
10) Android Pocket Primer (2017)
11) Angular Pocket Primer (2017)
12) Data Cleaning Pocket Primer (2018)
13) RegEx Pocket Primer (2018)

About Me: Training
=> Deep Learning. Keras, and TensorFlow:
http://codeavision.io/training/deep-learning-workshop
=> Instructor at UCSC (May/2018):
Deep Learning with TensorFlow
=> Mobile and TensorFlow Lite (WIP)
=> R and Deep Learning (WIP)
=> Android for Beginners

Deep Learning in your Browser: powered by WebGL

More Related Content

What's hot (20)

Similar to Deep Learning in your Browser: powered by WebGL (20)

Recently uploaded (20)

Deep Learning in your Browser: powered by WebGL