GTC 2018 Sony's deep learning software Neural...

21
Sony's deep learning software "Neural Network Libraries/Console“ and its use cases in Sony Yoshiyuki Kobayashi – Senior Machine Learning Researcher Sony Network Communications Inc. / Sony Corporation GTC 2018

Transcript of GTC 2018 Sony's deep learning software Neural...

Sony's deep learning software "Neural Network Libraries/Console“

and its use cases in Sony

Yoshiyuki Kobayashi – Senior Machine Learning Researcher Sony Network Communications Inc. / Sony Corporation

GTC 2018

2

Neural Network Libraries Neural Network Console

nnabla.org dl.sony.com

Windows version is now available for free. Open source (Apache 2.0 license)

Deep learning framework with Python API GUI based deep learning IDE

Sony's deep learning software

3

Sony real estate AR Effect Xperia Ear aibo Xperia Hello

• Used since 2011 and more than 1,000 users in Sony group. • Just released in 2017

• Already utilized in many products and services.

2011~ 1st gen framework

2013~ 2nd gen framework

2016~ 3rd gen framework Neural Network Libraries

2015~ GUI Tool Neural Network Console

Open Sourced at Jul. 2017

Released Windows Version at Aug. 2017 for Free

4

Motivation Deep learning engineer is lacking worldwide

Deep Learning is great and demand is explosively growing

• Need to make deep learning research & development more efficient • Also, human resource development needs to be accelerated

SOLD

5

Introduction of Neural Network Console https://www.youtube.com/watch?v=-lXjnaUSEtM

6

Application Example #1:Sony Real Estate Real Estate Price Estimate Neural Network Libraries is used in Real Estate Price Estimate Engine of Sony Real Estate Corporation. the Library realizes the solution that statistically estimates signed price in buying and selling real estate, analyzing massive data with unique algorism developed based on evaluation know-how and knowledge of Sony Real Estate Corporation. The solution is utilized in various businesses of Sony Real Estate Corporation such as ”Ouchi Direct”, “Real Estate Property Search Map” and “Automatic Evaluation”.

Input Bunch of features about a real estate Total floor area Floor plan Street address …

Output Price of a real estate

7

Application Example #2:Xperia Ear Gesture Sensitivity The Library is used in an intuitive gesture sensitivity function of Sony Mobile Communications “Xperia Ear”. Based on data from several sensors embedded in Xperia Ear, you can just use a nod of the head to confirm a command - answering Yes/No, answering/declining the phone call, cancelling text-to-speech reading of notifications, skipping/rewinding a song track.

Input Sensor data

Output Head gestures such as Yes and No

8

Application Example #3:aibo Image Recognition (User Identification, Face Tracking, etc.) The Library is used to realize image recognition of Sony’s Entertainment Robot "aibo" 『ERS-1000』. In the image recognition through fish-eye cameras installed at the nose, the Library is actively utilized for user identification, face tracking, charge stand recognition, generic object recognition, etc. These features and various inbuilt sensors enable its adaptable behavior.

入力 Image

Output Face, object, Charge stand, etc…

9

Application of DL spreading depending on input / output

… Input Output

Function Input Output

Image recognition Image Category of image

Image filtering Source image target image

Speech recognition Speech String

Machine translation (En->Ja) English word string Japanese word string

Chat-bot Input word string String of response

Sensor error detection Sensor signal Abnormality level

Robot control Robot's sensor Robot actuator

The possibilities of application are infinite

Function

10

Feature 1. Ideal for deep learning beginners

Easy to setup • Just unzip the downloaded file and run the application • Select “GPU” in Setup window to use NVIDIA GPU

You can learn deep learning visually Trial and error in a short time

Provides shortcut for skill improvement on deep learning

Logistic regression

Multi layer perceptron

Convolutional Neural Networks

11

Feature 2. Realize very efficient development

Easy debugging • Instantly locate the error on GUI

Easy to deploy

• Created model can be deployed immediately by using Neural Network Libraries • NNL offers both Python and C++ API

Automatic structure search • search for better neural network

structure automatically

12

More features

Functions Management of trial and error history Visualization of weights Export python code

Supports …

Various data as well as images, matrices and vectors Various objectives such as classification, detection,

signal processing, regression, etc. Multiple inputs and outputs for multi-modal application Training of huge neural networks like ResNet-152 Recurrent neural networks (RNN) Generative Adversarial Networks (GAN) Semi-supervised learning Transfer learning

etc…

13

Neural Network Libraries Lightweight C++ core library with Python API

• Flexible for arbitrary NNs • Cross-platform & Cross-device • Easy to add a new NN operator • Portable to C++

Sophisticated • Easy-to-write & easy-to-read • Fast training and inference with CUDA/cuDNN • Supports both static and dynamic NNs • Supports distributed training • Supports fp16 and Tensor Cores

• Resources • NNabla-Examles

https://github.com/sony/nnabla-examples/ ResNet, GANs, CapsNet, Quantized Nets, …

• Jupyter Notebook Tutorials https://github.com/sony/nnabla/tree/master/tutorial

https://github/sony/nnabla/

pip install nnabla

import nnabla as nn import nnabla.functions as F import nnabla.parametric_functions as PF x = nnabla.Variable((batch_size, 1, 28, 28)) c1 = PF.convolution(x, 16, (5, 5), name='c1') c1 = F.relu(F.max_pooling(c1, (2, 2))) c2 = PF.convolution(c1, 16, (5, 5), name='c2') c2 = F.relu(F.max_pooling(c2, (2, 2))) f3 = F.relu(PF.affine(c2, 50, name='f3')) y = PF.affine(f3, 50, name='f4') t = Variable((batch_size, 1)) loss = F.mean(F.softmax_cross_entropy(y, t)) …

Suitable for both research and production use

14

Neural Network Libraries Neural Network Console

Sony is providing an integrated development environment with hopes that a wider range of developers and researchers will build on its programs,

and with the aim of contributing to the development of society.

nnabla.org dl.sony.com

Windows version is now available for free. Open source (Apache 2.0 license)

Deep learning framework with Python API GUI based deep learning IDE

Sony's deep learning software

16

Appendix

17

Tutorial Image classification (hand-writing digits recognition)

… …

Classifier (Neural Network)

Input : image Output : Classification result

Take a typical application “Hand-writing digits recognition” as an example

2

Number of input neuron is size of input data

28x28 pixel

784 Number of output neural is number of categories of target data 10

18

Work required to create a image classifier

… …

Classifier (Neural Network)

Input : Image Output : Classification result

2

1. Prepare dataset Prepare as much data as possible Each data includes image and its class index 2. Design neural network structure You can use sample project 3. Train designed neural network using prepared dataset

Basically you can create image classifier with these three steps

… 「0」 … 「1」 … 「2」 … 「3」 … 「4」 … 「5」 … 「6」 … 「7」 … 「8」 … 「9」

19

1. Prepare dataset

• Create dataset CSV files according to a predetermined format

x:image y:label ./training/5/0.png 5 ./training/0/1.png 0 ./training/4/2.png 4 ./training/1/3.png 1 ./training/9/4.png 9 ./training/2/5.png 2 ./training/1/6.png 1 …

File name of hand-writing digit class Index

It can be created with Excel or simple script

20

2. Design neural network structure • Convolutional Neural Networks • Put layers as it is by drag & drop,

and edit properties of layers

Add layers by drag & drop

Edit property of layer

You can design the network structure very easily and intuitively

5 x 5 Convolution

Activation

2 x 2 Pooling

3 x 3 Convolution

Activation

2 x 2 Pooling

Affine

Activation

Fully connected

Softmax

Categorical Loss

21

3. Train designed neural network using prepared dataset

• Training starts just by pushing the training button

When the training is complete, an image classifier is obtained

loss

training iteration