+ All Categories
Home > Documents > Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep...

Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep...

Date post: 06-Aug-2020
Category:
Upload: others
View: 3 times
Download: 0 times
Share this document with a friend
43
Source: NVIDIA HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge
Transcript
Page 1: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Source: NVIDIA

HP Workstations, CGG, HP Labs 3D PrintMarch 2018

Deep Learning at the Edge

Page 2: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Today’s Presenters

BRUCE BLAHO

Fellow

Workstations Chief Technologist

HP Inc.

STEVE DOMINGUEZ

Team Lead

Seismic Interpretation Software

CGG

HE LUAN

Research Scientist, APhD2B

HP LABS 3D Print

HP Inc.

DR. JUN ZENG

Principal Investigator

HP LABS 3D Print

HP Inc.

Page 3: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Deep Learning Edge Development Platforms

HP Z8WorkstationUp to:• 56 CPU cores• 3 TB RAM• 48 TB Storage• 3-7 GPU’s • 9 PCI-e slots• 1125 - 1700 Watt

Power Supply• 3 year warranty

NVIDIAQuadro GV100• 32 GB HBM2• 5120 CUDA Cores• NVLink• 118.5 TFLOPS

NVIDIA GPU CloudContainers

Page 4: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

After the presentations…

Page 5: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Machine Learning at the Edge

for Seismic Interpretation Workflows

Steve Dominguez, CGG GeoSoftware

Page 6: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Machine Learning at the Edge - Talk Outline

•Deep Learning Applications in Seismic Interpretation– Recognize plausible applications

– Noise Filtering

– Object Classification / Image Segregation

•On-going R&D Overview– Neural Network Designs

– Training Data Sets

•R&D Workflow – Powered by HP Z workstations– Machine Learning at the Edge: Scalable R&D workflow for assured success

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 20186

Page 7: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Deep Learning in Seismic

7

Page 8: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Deep Learning Concept Overview

• Computers– great at repetitive tasks, raw calculations

– Slow and clunky at pattern recognition

• People– Great at pattern recognition

– Slow and clunky at repetitive calculation

• How would we design software to excel at pattern recognition?

8 Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 9: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

• “Hierarchical Learning Model”– Algorithms model high-level data abstractions using complex structures of non-linear transformations

– Enables software to identify data patterns without being explicitly programmed

• Involve *simple* math circuits in great numbers to model complex problems

• “Train” the network of simple circuits with isolated individual circuit adjustments to optimize output

9

Deep Learning Applications in Seismic Interpretation

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 10: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

• Salt Body Interpretation

10

Deep Learning Applications in Seismic Interpretation

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 11: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

• Salt Body Interpretation

11

Deep Learning Applications in Seismic Interpretation

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 12: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

• Acquisition Footprint / Noise

12

Deep Learning Applications in Seismic Interpretation

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 13: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

• Acquisition Footprint / Noise

13

Deep Learning Applications in Seismic Interpretation

Page 14: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

• Acquisition Footprint / Noise

14

Deep Learning Applications in Seismic Interpretation

Page 15: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Deep Learning Applications in Seismic Interpretation

• Fault Imaging

15Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 16: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Deep Learning Applications in Seismic Interpretation

• *Pattern Recognition*– Noise Mitigation

– Fault Imaging

– Geobody Detection

– Well Ties

• InsightEarth state-of-the-art approach– Take mundane repetitive tasks of ‘picking’ and accelerate

them through computer automation

– “Interpreter-Guided” automation

• The human must set appropriate parameters based on their geoscience learning and experience to guide the algorithms

16

• Deep Learning approach– Interpreter Guided ‘teaching’

– Developers must work with interpreters to guide the neural network with appropriate teaching data to facilitate good ‘learned’ behavior

– Training initially happens in a controlled environment (development lab)

– Released as ‘commercially viable’ with reasonable accuracy expectations

– Should include tuning parameters to refine behavior for various common scenarios

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 17: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

17

On-Going R&D Overview

Page 18: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Deep Learning Applications in Seismic Interpretation

18

• How much of the ‘middle man’ processes could we cut?• How much faster could this approach run over the current workflow…?!?

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 19: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

“NeUral Network Recognition for Faults - NURF” (working title) ☺

19

• Automatic Fault Extraction

• Current state-of-art:• 15x in-memory working copies of data, 12 passes through data

• Machine Learning Approach:• 1x in-memory active copy, 1 pass through data!

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 20: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

“Neural Network Noise Negation – N4” (working title) ☺

• Footprint Removal / Noise Conditioning

• Current state-of-art:

• Algorithm orients to match 1st derivative directional vectors at each point of data sampling

• Calculates a 2D operator “pane” of medians, takes the mean of the medians and adjusts the center location value

• GPU accelerated, but the orientation and interpolation is still time consuming…

• Machine Learning Approach:

• 1x in-memory active copy, 1 pass through data

20Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 21: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Neural Network Geobody Identification(don’t even have a working title yet)

• Salt / Stratigraphic Interpretation

• Workflow is HIGHLY subjective – X passes through data, dependent upon image quality, attributes, calculations required to “see” the geobodies

• Machine Learning Approach…

• Salt - 1 pass through the data!

• This is the easy one…

• Stratigraphy – in R&D stages determining applicability• Not entirely convinced you can even “see” these objects without a “domain transformation”

21Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 22: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

“Teach Your Children Well”

22

“Strange bird lost at sea”

• Network has never been trained with any fish before• Knows over a 1000 species of birds,

dogs, cats, etc.• But no fish….

• Good guess, all things considered!• Imagine what a young child might

have called this thing if they had never been taught the word ‘fish’ in any context…

Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 23: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

“Teach Your Children Well”

• Proper training data sets governs behavior!

• Deeper networks can learn more content, more accurately

• Top layer and bottom layer govern input constraints and output behavior

23Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 24: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Next Steps

• Deepen network layers for increased accuracy– Improved accuracy… more layers can learn more detections with improved results

• Integrate network layers for 3D processing vs. 2D flat imaging– Interpretation is a 3D process! Why would we want to restrict the available information to 2D slices? No reason the algorithms

can’t function in 3D…

• Improve training data

– Consider edge-stack vs. fault-enhance

– Consider wider variety of structural deformation examples for training: can’t expect it to identify a thrust fault if it’s never seen one before…

• Abstract output layer for multi-item detections

– Will eventually want to consider geo-body detection of any type

• Convert for ‘heat-map’ output instead of explicit bounding-box output– Better ‘probability’ model outputs / more accurate object bounding in this manner

24Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 25: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

25

R&D Workflow – Powered by HP Z Workstations

Machine Learning at the Edge! Scalable R&D workflow for assured success

Page 26: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Recipe for Successful R&D

• Agile Development Process

– … but “Mile-High Agile” is a totally different conference…

– Tangible / measurable milestones, clearly stated and understood objectives

– Why ‘train in the cloud’ or ‘train on the cluster’ before you’re reasonably sure of success??

• Expensive and time consuming!

• Start small, and scale!

– Laptop R&D PC R&D Workstation R&D …

26Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 27: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

R&D Path

• Retrained ‘LeNET’ deep learning model for fault detection

– Training conducted attempted with ‘fault / no-fault’ categories

– Training conducted with varying fault degree categories (0%, 25%, 50%, 75%, 100% certainties)

– Done on a laptop during a flight home from GTC

• Trained accuracy then tested on other Gulf area data sets

– Results around 83% accuracy when measured vs. expected behaviors from corresponding available Fault Enhance volumes

• Trained accuracy then tested on different structural regime data; North Sea

– Results around 70% accuracy when measured vs. expected behaviors from corresponding available Fault Enhance volumes

27Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 28: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

R&D Path

• Extending other convolutional networks with a deconv end layer into 3D models

– Training conducted arbitrary probability score floating point accuracy

– STILL ON A LAPTOP!

• Scale up into additional data sets

– Wider variety of geology involved

– Still using size-constrained input data sets…

– Scaled exact same dev / training environment up to a Z4, then Z6

28Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 29: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

R&D Path

• Grow into full-sized data sets, covering global expanse of Geological Examples

– Conduct initial training on Z8 workstation at full scale / full speed

• Multi GPU acceleration on GV100!

– 80% prediction of success can be measured from preliminary training – first few training epochs provide good indication of “correct” or “incorrect” direction

• Convergence or divergence seen in learning models

• Could then scale up across multiple Z8 “nodes”, or choose to go “cloud” depending on time, resources, and cost calculations…

– Study your learning curves vs. time / compute, and decide a) is more time training needed and b) are the cost-benefit returns present for ‘cloud’ investments

29Machine Learning at the Edge; Seismic Interpretation Applications – March 28, 2018

Page 30: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Apply Deep Learning to HP’s 3D Printing Fusing Science Research

He Luan1, Jun Zeng1, Sam Stodder2, Jordi Roca3, David Murphy1, Thomas Paula4

1HP Labs, 2HP/3D Printing/Fusing Sciences, 3HP/3D Printing/Software, 4HP/CTO

GTC 2018 Silicon Valley

March 28th 2018

Page 31: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

31

Polymer 3D Printing Multi Jet Fusion - MJF

Basics

Build Method Lamp Energy Melt Polymer Powder

Build Rate (mm/hr) High

Build Dimensions Medium

Typical Materials Nylon, other SC thermoplastics

Material Constraints No thermosets, amorphous

Resolution (Microns) 80 micron typical

Supports none

Post-process Decake, sandblast

Surface Finish Medium

Printer Cost

Material Cost Med, good reuse

Power Med-High

Parts/Print Multiple Parts in bed

Printer Companies HP

Additional Voxel-level control

Finished Parts still in print bed

MJF Part

Page 32: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief
Page 33: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Voxel thermal historyT

emp

erat

ure

(deg

C)

Powder Melt In Process Cool Post Process Cool

Fast cool

Slow cool

Lower CrystallinityMore Ductile

Higher CrystallinityMore Brittle

Time

hrs

High amount of Gamma CrystalStructures

Melt Temperature must be held long enough to completely break down Crystal Structure

Page 34: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Problem Statement

Energy InLosses

• Multi Jet Fusion technology relies on precisely projecting thermal energy at voxel level to ensure end-part quality.

• Problem addressed: create thermal prediction per layer at voxel level to enable closed-loop voxel thermal control.

• Why difficult:

– energy at voxel level is affected by voxel energy absorption/loss, in-layer thermal diffusion (spatial) and cross-layer thermal diffusion (tempo).

– Material properties are not only anisotropic but also phase dependent.

– Hi-resolution physical sensing is challenging.

• Our approach: apply deep learning to create a 3-stage deep neutral network model (DL4FS) that predicts thermal energy at voxel level based on digital print pipeline outputs.

Page 35: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Deep neural network architecture

Architectural innovation

Decouple principal spatial voxel energy driver and principal tempo voxel energy driver, and then synthesize both components as the final prediction.

learn heat map generated by voxel energy map (fusing/detailing agents).

learn the layer heat transferred from previous layers simulating heat transfer.

learn the contribution of above two components and synthesize them.

Current layer contonedetailing & fusing

Previous layers heat image

Current layer heat prediction

Current layer heat prediction

Current layer ultimate prediction

Spatial (CNN)

Spatiotemporal(Conv-LSTM)

Synthesis (CNN)

Page 36: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Spatial correlationlocal features: spatial correlation, learn possible spatial correlations by extracting multiple feature maps

One convolution layer: learn feature maps.

multiple convolution layer: learn the feature of features, explore deeper and more complex correlations!

shape Boundary thermal diffusionPart-part Thermal coupling

Page 37: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

How RNN capture sequential influence?

• CNN is layer independent

• RNN could capture the information transferred form previous layer. And this information is spatial.

•This is how heat transfers!

Contone change

CNN infers from current layer only, therefore lose important sequential information

Page 38: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Model network structure in Tensorflow

Page 39: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Thermal prediction layer by layer: inputs, predictions vs. ground truth

Data collected with a HP Jet Fusion 3D 4200 printer running at our R&D facility

Page 40: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Prediction error “heat-map” layer by layer

Page 41: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

Computing cost: Training & Predictions

Dataset

Sec.

/epoc

h

Num.

epoche

s

Training

time (hrs)

Accuracy

(MSE/MSSIM)

Prediction time

(sec.)Hardware

Printer #1

(HP/San Diego)Patch level 38.5 9.2K 125.7 2.97/0.87 0.13/layer 2 x Nvidia/M6000 (12GBx2)

Build-bed level 37 10K 110.4 2.00/0.94 0.05/layer 1 x Tesla/K80 (12GB)

Printer #2

(HP/Vancouver) Patch level 22 22.7K 170.3 2.43/0.88 0.12/layer 2 x Nvidia/M6000 (12GBx2)

Build-bed level 13.5 11.5K 43.1 2.22/0.91 0.04/layer 1 x Tesla/K80 (12GB)

Per-layer production time is in order of 10 seconds. Current per-layer prediction time cost shown here gives us reasonable hope that we may be able to integrate this as a run-time prediction-correction step.

Page 42: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

42

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Thank you!

MB1@Jabil

Page 43: Deep Learning at the Edge - NVIDIA€¦ · HP Workstations, CGG, HP Labs 3D Print March 2018 Deep Learning at the Edge. Today’s Presenters BRUCE BLAHO Fellow Workstations Chief

And the winner is…


Recommended