Download as PDF, PPTX
![backend interface
13
from neon.backends import gen_backend
be = gen_backend(backend=‘gpu') # backend creation
x0 = be.ones((2, 2), name=‘x0') # buffer allocation
x1 = be.ones((2, 2), name=‘x1')
f_val = be.empty((2, 2))
f = x0 + x1 # op-tree creation
f_val[:] = f # execution
• numpy-like
• consistent cpu/gpu interface
• deferred execution op-tree](https://image.slidesharecdn.com/meetup03032016-160304205016/75/Introduction-to-Deep-Learning-with-Will-Constable-13-2048.jpg)
Uploaded byIntel Nervana
Introduction to Deep Learning with Will Constable
This document discusses deep learning frameworks Neon and Nervana Cloud. It provides an introduction to deep learning, describing how it uses learned features from multiple levels to make predictions rather than designed features. The document outlines Neon's model design, training workflow, model zoo, Caffe compatibility, and backend interface. It also describes the Nervana Cloud for importing, building, training, and deploying models using Neon on AWS hardware and S3 storage. A demo of the tools is proposed.
More Related Content
![[252] 증분 처리 플랫폼 cana 개발기](https://cdn.slidesharecdn.com/ss_thumbnails/225cana-150915052201-lva1-app6892-thumbnail.jpg?width=640&height=640&fit=bounds)
![[251] implementing deep learning using cu dnn](https://cdn.slidesharecdn.com/ss_thumbnails/215implementingdeeplearningusingcudnn-150915052020-lva1-app6892-thumbnail.jpg?width=640&height=640&fit=bounds)
Similar to Introduction to Deep Learning with Will Constable
Introduction to Deep Learning with Will Constable
- 1. Proprietary and confidential.Do not distribute. Deep Learning, neon, and the Nervana Cloud Will Constable March 3 2016 MAKING MACHINES SMARTER.™
- 2. Agenda 2 Neon Nervana Cloud Intro toDeep Learning Demo
- 3.
- 4. what is deeplearning? 4 Before deep learning: • Input → designed features → output • Input → designed features → SVM → output • Input → learned features → SVM → output • Input → levels of learned features → output
- 5. what is deeplearning? 5 (Zeiler and Fergus, 2013)
- 6. what is deeplearning? 6 ++ +
- 7. Proprietary and confidential.Do not distribute. Deep learning techniques imagnet error rate 7 Source: ImageNet 1: ImageNet top 5 error rate Error rate1 human performance
- 8. what is deeplearning? 8 No free lunch: • lots of data • model design needs intuition • slow experiments
- 9. agenda 9 Neon Nervana Cloud Intro toDeep Learning Demo
- 10. model Convolution BatchNorm ReLu Activation training workflow 10 FullyConnected Pooling Softmax cost optimizer Train Set Validation Set
- 11. model zoo 11 • github.com/nervanazoo/NervanaModelZoo •model files, parameters GoogLeNetAlexnet VGG Deep Residual Net bAbI Q&A imdb Sentiment Analysis Video Activity Detection Deep Reinforcement Learning LSTM Image Captioning Fast-RCNN Object Localization AllCNN
- 12. caffe compatibility 12 • github.com/NervanaSystems/caffe2neon.git •Convert caffe model to neon format > ./decaffeinate.py caffe_model.prototxt caffe_params -o neon_model
- 13. backend interface 13 from neon.backendsimport gen_backend be = gen_backend(backend=‘gpu') # backend creation x0 = be.ones((2, 2), name=‘x0') # buffer allocation x1 = be.ones((2, 2), name=‘x1') f_val = be.empty((2, 2)) f = x0 + x1 # op-tree creation f_val[:] = f # execution • numpy-like • consistent cpu/gpu interface • deferred execution op-tree
- 14. full stack optimization 14 •multithreaded data loader • hand-assembled kernels • compound kernel generator
- 15.
- 16. winograd convolution 16 • 2xfaster convolutionAlgorithmicSpeedup 0 0.5 1 1.5 2 winograd cuDNN v4 Efficient Convolution
- 17. Agenda 17 Neon Nervana Cloud Intro toDeep Learning Demo
- 18. nervana cloud 18 Import BuildTrain Deploy
- 19. nervana cloud 19 neon deep learning framework S3NervanaHardwareAWS S3TitanX Cluster
- 20.
- 21.
- 22. Proprietary and confidential.Do not distribute. Summary 22 • neon 1.3 (download today) • cloud - contact us (info@nervanasys.com) • We’re hiring! (nervanasys.com/careers) will@nervanasys.com
- 23. Agenda 23 Neon Nervana Cloud Intro toDeep Learning Demo