Submanifold Sparse Convolutional Networks for Sparse, Locally Dense - - PowerPoint PPT Presentation

▶
submanifold sparse convolutional networks for sparse
SMART_READER_LITE
LIVE PREVIEW

Submanifold Sparse Convolutional Networks for Sparse, Locally Dense - - PowerPoint PPT Presentation

Oct 02, 2023 471 likes •804 views

Submanifold Sparse Convolutional Networks for Sparse, Locally Dense Particle Image Analysis Laura Domine (Stanford / SLAC) Kazuhiro Terao (SLAC) 2018 CPAD Instrumentation Frontier Workshop 1 Outline 1. Particle image analysis &

slide-1
SLIDE 1

Submanifold Sparse Convolutional Networks for Sparse, Locally Dense Particle Image Analysis

Laura Domine (Stanford / SLAC) Kazuhiro Terao (SLAC)

2018 CPAD Instrumentation Frontier Workshop

1

slide-2
SLIDE 2

2018 CPAD / L.Domine and K.Terao

Outline

  • 1. Particle image analysis & Convolutional networks
  • 2. Submanifold Sparse Convolutions
  • 3. Comparison study between a dense and sparse network

2

slide-3
SLIDE 3

2018 CPAD / L.Domine and K.Terao

Outline

  • 1. Particle image analysis & Convolutional networks
  • 2. Submanifold Sparse Convolutions
  • 3. Comparison study between a dense and sparse network

3

e

slide-4
SLIDE 4

2018 CPAD / L.Domine and K.Terao

Particle Image Analysis with LArTPCs

Liquid Argon Time Projection Chamber (LArTPC) = particle imaging detector ~3mm resolution

4 Cosmic rays in a 3D LArTPC charge readout (arxiv:1808.02969) @ LBNL Neutrino interaction candidate from MicroBooNE experiment @ Fermilab

Wire LArTPC (2D projections) Pixel LArTPC (native 3D)

slide-5
SLIDE 5

2018 CPAD / L.Domine and K.Terao

Particle Image Analysis with LArTPCs for neutrinos

5

Neutrino detectors & LArTPCs Goal: Extract flavor + energy

slide-6
SLIDE 6

2018 CPAD / L.Domine and K.Terao

Particle Image Analysis with LArTPCs for neutrinos

6

Neutrino detectors & LArTPCs Goal: Extract flavor + energy

slide-7
SLIDE 7

2018 CPAD / L.Domine and K.Terao

Convolutional Neural Networks

Semantic segmentation Object detection & classification

7

Now state-of-the art technique in computer vision for complex image analysis tasks:

slide-8
SLIDE 8

2018 CPAD / L.Domine and K.Terao

Sparse, locally dense data

Less than 1% of voxels are nonzero in LArTPC images

% of nonzero voxels:

  • ~0.05% for 192px^3
  • ~0.01% for 512px^3

But CNNs rely on dense matrix multiplications!

Dense Sparse (but locally dense)

8

slide-9
SLIDE 9

2018 CPAD / L.Domine and K.Terao

Outline

  • 1. Particle image analysis & Convolutional networks
  • 2. Submanifold Sparse Convolutions
  • 3. Comparison study between a dense and sparse network

9

slide-10
SLIDE 10

2018 CPAD / L.Domine and K.Terao

Submanifold Sparse Convolutions

Many possible definitions and implementations of ‘sparse convolutions’... Submanifold Sparse Convolutions (arxiv:1711.10275, CVPR2018): https://github.com/facebookresearch/SparseConvNet State-of-the-art on ShapeNet challenge (3D part segmentation)

10

slide-11
SLIDE 11

2018 CPAD / L.Domine and K.Terao

Submanifold Sparse Convolutions

Submanifold = “input data with lower effective dimension than the space in which it lives” Ex: 1D curve in 2+D space, 2D surface in 3+D space Our case: the worst! 1D curve in 3D space...

11

slide-12
SLIDE 12

2018 CPAD / L.Domine and K.Terao

Submanifold Sparse Convolutions

3D Semantic Segmentation with Submanifold Sparse Convolutional Networks (arxiv: 1711.10275)

Dilation problem

  • 1 nonzero site leads to 3d nonzero sites after 1 convolution
  • How to keep the same level of sparsity throughout the network?

12

slide-13
SLIDE 13

2018 CPAD / L.Domine and K.Terao

Submanifold Sparse Convolutions

13

2-classes (particle track vs electromagnetic shower ) pixel-level segmentation on 512px 3D images.

Input Predictions

slide-14
SLIDE 14

2018 CPAD / L.Domine and K.Terao

Outline

  • 1. Particle image analysis & Convolutional networks
  • 2. Submanifold Sparse Convolutions
  • 3. Comparison study between a dense and sparse network

14

slide-15
SLIDE 15

2018 CPAD / L.Domine and K.Terao

Outline

  • 1. Particle image analysis & Convolutional networks
  • 2. Submanifold Sparse Convolutions
  • 3. Comparison study between a dense and sparse network

a. Dataset b. Task c. Metrics d. Network architecture

15

slide-16
SLIDE 16

2018 CPAD / L.Domine and K.Terao

  • 1. Dataset & 2. Task

Total: 100,000 simulated 3D events Spatial size: 192px / 512px / 768px (~3mm/pix) Semantic segmentation with 5 classes

  • Protons
  • Minimum ionizing particles

(muons and pions)

  • Electromagnetic shower
  • Delta rays
  • Michel electrons

Publicly available: https://osf.io/vruzp/

16

slide-17
SLIDE 17

2018 CPAD / L.Domine and K.Terao

  • 3. Metrics
  • Nonzero accuracy: fraction of correctly labeled pixels, i.e.

# nonzero voxels whose predicted label is correct / # nonzero voxels

  • GPU memory (hardware limitation)
  • Computation time

17

slide-18
SLIDE 18

2018 CPAD / L.Domine and K.Terao

  • 4. Network architecture: UResNet

Encoder Decoder

UResNet = U-Net + ResNet (residual connections)

18

Residual connections

input conv conv-s2 dconv-s2 linear softmax

Concatenation

slide-19
SLIDE 19

2018 CPAD / L.Domine and K.Terao

  • 4. Network architecture: UResNet

Encoder Decoder

UResNet = U-Net + ResNet (residual connections)

19

Residual connections

input conv conv-s2 dconv-s2 linear softmax

Concatenation

slide-20
SLIDE 20

2018 CPAD / L.Domine and K.Terao

  • 4. Network architecture: UResNet

Encoder Decoder

UResNet = U-Net + ResNet (residual connections)

20

Residual connections

input conv conv-s2 dconv-s2 linear softmax

Concatenation

slide-21
SLIDE 21

2018 CPAD / L.Domine and K.Terao

  • 4. Network architecture: UResNet

Encoder Decoder

UResNet = U-Net + ResNet (residual connections)

21

Residual connections

input conv conv-s2 dconv-s2 linear softmax

Concatenation

Concatenation

slide-22
SLIDE 22

2018 CPAD / L.Domine and K.Terao

  • 4. Network architecture: UResNet

Encoder Decoder

UResNet = U-Net + ResNet (residual connections)

22

Residual connections

input conv conv-s2 dconv-s2 linear softmax

Concatenation

Residual connections

slide-23
SLIDE 23

2018 CPAD / L.Domine and K.Terao

Semantic Segmentation with UResNet: it works.

A Deep Neural Network for Pixel-Level Electromagnetic Particle Identification in the MicroBooNE Liquid Argon Time Projection Chamber. (arxiv:1808.07269)

Data Network’s output

23

slide-24
SLIDE 24

2018 CPAD / L.Domine and K.Terao

Dense vs Sparse UResNet

24

Sparse = 99.3% Dense = 92% Nonzero Accuracy (training) vs Iterations Dense & Sparse both trained with 80k events

slide-25
SLIDE 25

2018 CPAD / L.Domine and K.Terao

Dense vs Sparse UResNet

25

Sparse = 19h Dense = 11 days Nonzero Accuracy (training) vs Wall Time Dense & Sparse both trained with 80k events

slide-26
SLIDE 26

2018 CPAD / L.Domine and K.Terao

Dense vs Sparse UResNet

26

Sparse Dense Input Spatial Size 192px 512px 768px 192px Final nonzero accuracy 98% 98.8% 98.9% 92%* GPU memory usage (Gb) 0.066 0.57 1.0 4.6 Forward computation time (s) 0.058 2.6 3.6 0.68

Performance for different input spatial size

*Training time accuracy.

slide-27
SLIDE 27

2018 CPAD / L.Domine and K.Terao

Dense vs Sparse UResNet

27

Sparse = Almost linear... Dense = Power? Exponential?!

slide-28
SLIDE 28

2018 CPAD / L.Domine and K.Terao

Learning from mistakes: the case of Michel electrons

28

Mean % of nonzero voxels in an event Nonzero accuracy per class HIP 12% 98.4% MIP 43% 99.5% EM shower 42% 99.1% Delta rays 2% 87.5% Michel electrons 1% 62.8%

Nonzero accuracy per class = # correctly predicted voxels in this class / # voxels in this class

slide-29
SLIDE 29

2018 CPAD / L.Domine and K.Terao

Learning from mistakes: the case of Michel electrons

29

Predictions

slide-30
SLIDE 30

2018 CPAD / L.Domine and K.Terao

Learning from mistakes: the case of Michel electrons

30

True Labels

slide-31
SLIDE 31

2018 CPAD / L.Domine and K.Terao

Summary

Submanifold sparse convolutions...

  • Run faster
  • Use less GPU memory
  • … and outperform standard convolutions.

Better performance and better scalability! Reproduce our results / start using SSCN:

  • Open dataset
  • Software containers available

31