For the full video of this presentation, please visit:
https://www.embedded-vision.com/platinum-members/embedded-vision-alliance/embedded-vision-training/videos/pages/may-2017-embedded-vision-summit-bordoloi
For more information about embedded vision, please visit:
http://www.embedded-vision.com
Unmesh Bordoloi, Senior Researcher at General Motors, presents the "Collaboratively Benchmarking and Optimizing Deep Learning Implementations" tutorial at the May 2017 Embedded Vision Summit.
For car manufacturers and other OEMs, selecting the right processors to run deep learning inference for embedded vision applications is a critical but daunting task. One challenge is the vast number of options in terms of neural network models, frameworks (such as Caffe, TensorFlow, Torch), and libraries such as CUDA and OpenCL. Another challenge is the large number of network parameters that can affect the computation requirements, such the choice of training data sets, precision, and batch size. These challenges also complicate efforts to optimize implementations of deep learning algorithms for deployment.
In this talk, Bordoloi presents a methodology and open-source software framework for collaborative and reproducible benchmarking and optimization of convolutional neural networks. General Motors' software framework, CK-Caffe, is based on the Collective Knowledge framework and the Caffe framework. GM invites the community to collaboratively evaluate, design and optimize convolutional neural networks to meet the performance, accuracy and cost requirements of a variety of applications – from sensors to self-driving cars.
The Ultimate Guide to Choosing WordPress Pros and Cons
"Collaboratively Benchmarking and Optimizing Deep Learning Implementations," a Presentation from General Motors
1. General Motors 1
Collaboratively Benchmarking and Optimizing
Deep Learning Implementations
Unmesh D. Bordoloi1, Massimo Osella1, Anton Lokhmotov2, Grigori Fursin2
May 2017
General Motors1 dividiti2
2.
3. General Motors 3
• Towards global availability everywhere
• Towards affordability by everyone
Opportunities for optimization
Among others, the high performance
computing requirements is a factor
4. General Motors 4
Major components of autonomous car
Radar Data
Lidar Data
Processing
Sensor Fusion
Camera Image
Processing
Scene
Understanding
Control
Algorithms
Deep learning has been identified as a key enabler for at
least some of these functions
Planning
6. General Motors 6
Chaotic universe of deep learning
DSP
TensorFlow
clBLAS AlexNet
Many-
core
CPU
Squeez
eNet
GoogleNet
FPGA
GM‘s world
famous secret
network
CLBlast
OpenBLAS
cuBLAS
(BVLC)
cuDNN
(BVLC)
cuBLAS
(NVIDIA)
cuDNN
(NVIDIA) cuBLAS
fp16
(NVIDIA)
cuDNN
fp16
(NVIDIA)
viennaCL
libDNN-
clBLAS
libDNN-
CLBlast
libDNN-
cuBLAS
libDNN-
viennaCL
TensorRT
TensorRT
fp16
YOLO
SSD
Fast
RCNN
Mask
RCNN
RCNN
VGG
ResNet
Caffe
Torch
Theano
pyTorch
CNTK
GPU ASIC/ASIP
Vendor 1 Vendor 2 Vendor n
7. General Motors 7
• The Computational Challenge for Self-Driving Cars
• Collective Knowledge (CK)
• Results
• Outlook
Outline of this talk
8. General Motors 8
Libraries
Goal: CK (Collective Knowledge) to address the
problem of benchmarking
Hardware
DSP
Many-
core
CPU
FPGA
GPU
CK
clBLAS
CLBlast
Open
BLAS
cuBLAS
(BVLC)
cuDNN
(BVLC)
cuBLAS
fp16
(NVIDIA)
cuDNN
fp16
(NVIDIA)
viennaCL
libDNN
clBLAS
libDNN-
CLBlast
libDNN-
cuBLAS
libDNN-
viennaCL
TensorRT
TensorRT
fp16
TensorFlowCaffe
Torch Theano
pyTorchCNTK
Framework
AlexNet
Squeeze
Net
GoogleNet
GM‘s world
famous
secret
network
YOLO
SSD
Fast
RCNN
Mask
RCNN
RCNN
VGG ResNet
Neural Net
Model
RESULTS
KITTIImagenet GM dataset
Dataset
9. General Motors 9
TECHNOLOGY
• A technology for creating, sharing and re-
using research artifacts such as
workloads, datasets, tools, experimental
results, experimental workflows, predictive
models, etc.
• Cross-platform
• Customizable/extensible
• Open-source
• Lightweight Python2/3 package
CK (Collective Knowledge) approach
METHODOLOGY
• Enables systematic and reproducible
experimentation
• Encourages artifact share and re-use
• Involves the community to collaboratively
find and explain unexpected behaviour
• Crowdsourcing: benchmarking, design
space exploration, optimization…
10. General Motors 10
• CK-Caffe is a CK based wrapper around the Caffe framework from
Berkeley as well as the libraries associated with Caffe
• Cross-platform: Linux, windows, android
• Customizable: extensible with new models, datasets, forks of Caffe…
• Open-source: https://github.com/dividiti/ck-caffe
• Lightweight: Python2/3 package
What is CK-Caffe?
11. General Motors 11
• GM, dividiti, and suppliers are using CK-Caffe for performance evaluation and
design space exploration
• We welcome the community to use and contribute
• Go to https://github.com/dividiti/ck-caffe
• (Or type in “CK-Caffe” or “cknowledge” in your search engine)
• Follow the steps outlined in the guide, e.g.
CK Caffe: Getting started
12. General Motors 12
• The Computational Challenge for Self-Driving Cars
• Collective Knowledge (CK)
• Results
• Outlook
Outline of this talk
20. General Motors 20
• The Computational Challenge for Self-Driving Cars
• Collective Knowledge (CK)
• Results
• Outlook
Outline of this talk
21. General Motors 21
• Collaborate in benchmarking/optimization: IP providers, chipmakers &
OEMs can use CK-Caffe
• GM, dividiti, several chip vendors
• Provides validation to numbers shown in marketing slides
• Results can be easily reproduced
• And growing…
• Student/community competitions
• Academic community (University of Michigan)
CK-Caffe for collaborative
benchmarking/optimization
22. General Motors 22
• For applications/workloads like Deep Learning/CNN
• What is the right kind of abstraction and models to represent
workloads for real-time schedulability analysis?
• For resources like GPU, FPGA, DSP
• How can we formally analyze worst-case execution time (WCET),
schedulability, utilization on such devices for deep learning networks?
Open challenges for design automation / real-
time systems community: examples
23. General Motors 23
• Links
• http://cknowledge.org/ai.html
• https://github.com/dividiti/ck-caffe
• https://twitter.com/cruise
Resources and Thanks!