Dlobd: a Comprehensive Study of Deep Learning Over Big Data Stacks on HPC Clusters Xiaoyi Lu , Haiyang Shi, Rajarshi Biswas, M

Total Page:16

File Type:pdf, Size:1020Kb

Dlobd: a Comprehensive Study of Deep Learning Over Big Data Stacks on HPC Clusters Xiaoyi Lu , Haiyang Shi, Rajarshi Biswas, M DLoBD: A Comprehensive Study of Deep Learning over Big Data Stacks on HPC Clusters Xiaoyi Lu , Haiyang Shi, Rajarshi Biswas, M. Haseeb Javed , and Dhabaleswar K. Panda 1 What is Big Data? ● "Big data" is a field that treats ways to analyze, systematically extract information from, or otherwise deal with data sets that are too large or complex to be dealt with by traditional data-processing application software. (Source: Wikipedia) ● What does “Big” mean? ○ According to Forbes, we generate 2.5 quintillion bytes of data each day. ● Properties of Big Data ○ Variety ○ Velocity ○ Volume 2 How is Big Data different from large data stored in relational databases? How is Big Data different from large data stored in relational databases? 3 Uses of Big Data Predictive analytics: Utilizing a variety of statistical techniques ● Data mining(discovering patterns in large data sets), ● Predictive modeling(uses statistics to predict outcomes), and ● Machine learning User Behaviour Analytics (UBA) : analyze human behavior patterns 4 Big Data Technology stack Source: https://www.slideshare.net/Khalid-Imran/big-data-technology-stack-nutshell 5 Big Data Frameworks 6 Hadoop vs Spark ● Spark is 100 times faster than Hadoop. Why? ○ In-memory computation ● Spark does not provide it’s own distributed storage. ● Hence, Big Data projects involve installing Spark on top of Hadoop. ○ Spark applications can use Hadoop Distributed File System (HDFS). 7 Batch Processing ● Processing happens of blocks of data that have already been stored over a period of time. ● Example - calculating monthly payroll summaries. 8 source - https://medium.com/@gowthamy/big-data-battle-batch-processing-vs-stream-processing-5d94600d8103 Stream Processing ● Process data in real time as they arrive. ● Example - determining if a bank transaction is fraudulent or not. 9 source - https://medium.com/@gowthamy/big-data-battle-batch-processing-vs-stream-processing-5d94600d8103 MapReduce ● Processing technique and a program model for distributed computing. ● Key idea - take the input data and divide it into many parts. Each part is then sent to a different machine to be processed and finally aggregated. ● Based upon horizontal scaling. ● Under the MapReduce model, the data processing primitives are called mappers and reducers. 10 MapReduce source - https://towardsdatascience.com/big-data-analysis-spark-and-hadoop-a11ba591c057 11 DLoBD 12 Introduction ➢ What is DLoBD? ○ Running Deep Learning libraries over Big Data stacks, like Apache Hadoop and Spark. ➢ Advantages - enable distributed Deep Learning on Big Data analytics clusters. ➢ Contributions of paper: ○ Extensive performance evaluations of representative DLoBD stacks. ○ Characterization based on performance, scalability, accuracy, and resource utilization. 13 Why DL over Big Data? 1. Better data locality. 2. Efficient resource sharing & cost effective. 3. Easy integration of DL on Big Data processing components. source: https://www.kdnuggets.com/2016/02/yahoo-caffe-spark-distributed-deep-learning.html 14 source: https://www.slideshare.net/JenAman/caffeonspark-deep-learning-on-spark-cluster 15 Examples of DLoBD stacks ● CaffeOnSpark ● SparkNet ● TensorFlowOnSpark ● TensorFrame ● DeepLearning4J ● BigDL ● MMLSpark (or CNTKOnSpark) source:https://www.slideshare.net/databricks/dlobd-an-emerging-paradigm-of- deep-learning-over-big-data-stacks-with-dhabaleswar-k-panda-and-xiaoyi-lu 16 Convergence of DL, Big Data, and HPC ● How much performance benefit we can achieve for end deep learning applications? ● What is the impact of these advanced hardware and the associated efficient building blocks on various Deep Learning aspects? ● How much performance overhead is brought due to the heavy layers of DLoBD stacks? 17 CaffeOnSpark Overview ➢ Designed by Yahoo! ➢ Inherits features from Caffe like computing on: ○ CPU ○ GPU ○ GPU + cuDNN ➢ Enables Deep Learning training and testing with Caffe embedded inside Spark applications. ➢ Major benefit? ○ Eliminates unnecessary data movement 18 Background ● Resilient Distributed Datasets (RDDs) ○ Fundamental data structure of Spark ○ Read-only ○ Each dataset in RDD is divided into logical partitions, which may be computed on different nodes of the cluster. ● Hadoop YARN (Yet Another Resource Negotiator) ○ Job scheduling and cluster resource management ○ Two components — ■ ResourceManager : manages the global assignment of compute resources to applications, e.g., memory, cpu, disk, network, etc. ■ NodeManager : tracks its own local resources and communicates its resource configuration to the ResourceManager. 19 Background - Spark Cluster Mode ● Master/slave architecture ● The driver converts the user program into tasks and after that it schedules the tasks on the executors. ● Executors are worker nodes in charge of running individual tasks in a given Spark job. Once they have run the task they send the results to the driver. 20 CaffeOnSpark Architecture 21 CaffeOnSpark Advantages Disadvantages ➢ Integration of Deep Learning, Big Data, ➢ Parameter exchange phase is and HPC. implemented in a out-of-band fashion. ○ Cannot take advantage of optimizations for ➢ Scaling out of Caffe with the help from Hadoop YARN and Spark frameworks. defaut Spark. Dedicated communication channels (either ➢ By using HDFS: ➢ ○ Data-sharing RDMA or TCP/IP based) in model ○ Fault tolerance synchronizers. ○ Need extra effort to maintain them. 22 TensorFlowOnSpark ● Again designed by Yahoo! ● Allows running DL jobs using TensorFlow over Big Data clusters. ● Parameter Server approach ○ Storing and updating the model’s parameter. ○ Async training - all workers are independently training over the input data and updating variables asynchronously. 23 MMLSpark ( or CNTKOnSpark) ● Similar architecture as the previous DLoBD stacks. ● Out-of-band communication via MPI. ○ Users can specify the channel (TCP/IP or RDMA) based on MPI library chosen. ○ In this paper, MMLSpark is complied with OpenMPI library. 24 BigDL ➢ Proposed by Intel. ➢ Allows to import pre-trained models in Caffe and Torch into Spark. ➢ Written using Intel’s Math Kernel Library (MKL). ○ Optimizes math functions like Linear Algebra, FFTs. ➢ Advantage: ○ No separate communication channel required. 25 Characterization Methodology 26 Data Sets 27 Selected Deep Learning Models and Algorithms 28 Performance Evaluation 29 Experimental Setup 1. OSU RI2 Cluster (Cluster A): a. 20 nodes i. 2 Intel Broadwell 14-core processors ii. 128 GB RAM iii. 128 GB local HDD iv. Nvidia Tesla K80 GPU b. Connected with Mellanox single port InifiniBand EDR (100 Gbps) HCA. 2. SDSC Comet Cluster (Cluster B): a. 1984 nodes (used only 17 nodes) i. Intel Haswell dual 12-core processors ii. 128 GB RAM iii. 320 GB local HDD b. Connected with Infiniband FDR (56 Gbps) 3. Default - # nodes x batch size = 128 30 Evaluation on CPU vs. GPU (Cluster A) 31 Evaluation on IPoIB vs. RDMA (Cluster A) 32 Evaluation on Performance & Accuracy 33 Epoch-level evaluation VGG model using BigDL on default Spark CPU cores - 192 and Batch Size - 768 34 Scalability Evaluation VGG model on Cluster B using BigDL Figure shows accumulative time taken to finish 18th epoch 2.7x speedup 35 Evaluation on Resource Utilization Cifar-10 Quick model on Cifar-10 dataset with CaffeOnSpark (Cluster A) Utilization results generated with avg resource consumption per 60 seconds. (During training) 36 TensorFlowOnSpark 37 Performance Overhead on TensorFlowOnSpark ● TensorFlowOnSpark doesnot involve Spark drivers in tensor communication. ● Experimental Setup: ○ SoftMax Regression Model ○ MNIST datset ○ 4 node parameter server ○ Batch size - 128 ○ CPU training ● Between ~15 - 19% of execution time spent in YARN and Spark each. ● This overhead can be amortized for long running DL jobs. 38 Impact of RDMA on TensorFlow ❖ Experimental Setup: ➢ ResNet50 model ➢ TF CNN benchmarks ➢ Batch size - 64 ➢ GPU based training ➢ 2 GPUs/node ➢ Parameter server mode (1 PS and rest workers) ➢ MPI libraries – ➢ MVAPICH2-2.3b ➢ Intel-MPI-2018 39 Summary ● Detailed architectural overview of 4 representative DLoBD stacks: ○ CaffeOnSpark ○ TensorFlowOnSpark ○ MMLSpark ○ BigDL ● RDMA scheme benefits DL workloads. ○ 2.7x performance speedup with RDMA vs IPoIB. ● Mostly, GPU-based DL designs outperform CPU-based ones. 40 BigDL: A Distributed Deep Learning Framework for Big Data 41 Introduction ❏ Distributed deep learning framework for Big Data platforms and workflows. ❏ Implemented on top of Apache Spark. ❏ Allows to write DL applications as standard Spark programs. source: https://www.slideshare.net/SparkSummit/bigdl-a-distributed-deep-learning-library-on-spark-spark-summit-east-talk-by-yiheng-wang 42 Why BigDL? source: https://www.slideshare.net/SparkSummit/bigdl-a-distributed-deep-learning-library-on-spark-spark-summit-east-talk-by-yiheng-wang 43 Why BigDL? ❏ Rich Deep learning support ❏ Neural network operations ❏ Layers ❏ Losses ❏ Optimizers ❏ Directly run existing models defined in other frameworks (such as TensorFlow, Keras, Caffe and Torch). 44 Programming Model Transform text to list of words 45 Programming Model Data transformation 46 Programming Model Model construction 47 Programming Model Model Training 48 Programming Model Model Prediction 49 Execution Model ➔ Single driver and multiple workers. ➔ Driver - scheduling and dispatching tasks to worker nodes. ➔ Worker - actual computation and physical data storage. ➔ How does BigDL support efficient and scalable distributed training on top of Spark? 50 Data-parallel training ● Synchronous mini-batch SGD. ● Distributed
Recommended publications
  • Intel® Optimized AI Frameworks
    Intel® optimized AI frameworks Dr. Fabio Baruffa & Shailen Sobhee Technical Consulting Engineers, Intel IAGS Visit: www.intel.ai/technology Speed up development using open AI software Machine learning Deep learning TOOLKITS App Open source platform for building E2E Analytics & Deep learning inference deployment Open source, scalable, and developers AI applications on Apache Spark* with distributed on CPU/GPU/FPGA/VPU for Caffe*, extensible distributed deep learning TensorFlow*, Keras*, BigDL TensorFlow*, MXNet*, ONNX*, Kaldi* platform built on Kubernetes (BETA) Python R Distributed Intel-optimized Frameworks libraries * • Scikit- • Cart • MlLib (on Spark) * * And more framework Data learn • Random • Mahout optimizations underway • Pandas Forest including PaddlePaddle*, scientists * * • NumPy • e1071 Chainer*, CNTK* & others Intel® Intel® Data Analytics Intel® Math Kernel Library Kernels Distribution Acceleration Library Library for Deep Neural Networks for Python* (Intel® DAAL) (Intel® MKL-DNN) developers Intel distribution High performance machine Open source compiler for deep learning optimized for learning & data analytics Open source DNN functions for model computations optimized for multiple machine learning library CPU / integrated graphics devices (CPU, GPU, NNP) from multiple frameworks (TF, MXNet, ONNX) 2 Visit: www.intel.ai/technology Speed up development using open AI software Machine learning Deep learning TOOLKITS App Open source platform for building E2E Analytics & Deep learning inference deployment Open source, scalable,
    [Show full text]
  • Lecture 6 Learned Feedforward Visual Processing Neural Networks, Deep Learning, Convnets
    William T. Freeman, Antonio Torralba, 2017 Lecture 6 Learned feedforward visual processing Neural Networks, Deep learning, ConvNets Some slides modified from R. Fergus We need translation invariance Lots of useful linear filters… Laplacian Gaussian derivative Gaussian Gabor And many more… High order Gaussian derivatives We need translation and scale invariance Lots of image pyramids… Gaussian Pyr Laplacian Pyr And many more: QMF, steerable, … We need … What is the best representation? • All the previous representation are manually constructed. • Could they be learnt from data? A brief history of Neural Networks enthusiasm time Perceptrons, 1958 Rosenblatt http://www.ecse.rpi.edu/homepages/nagy/PDF_chrono/2011_Na gy_Pace_FR.pdf. Photo by George Nagy 9 http://www.manhattanrarebooks-science.com/rosenblatt.htm Perceptrons, 1958 10 Perceptrons, 1958 enthusiasm time Minsky and Papert, Perceptrons, 1972 12 Perceptrons, 1958 enthusiasm Minsky and Papert, 1972 time Parallel Distributed Processing (PDP), 1986 14 XOR problem Inputs Output 0 0 0 1 0 1 0 1 1 0 1 1 1 0 0 1 PDP authors pointed to the backpropagation algorithm as a breakthrough, allowing multi-layer neural networks to be trained. Among the functions that a multi-layer network can represent but a single-layer network cannot: the XOR function. 15 Perceptrons, PDP book, 1958 1986 enthusiasm Minsky and Papert, 1972 time LeCun conv nets, 1998 Demos: http://yann.lecun.com/exdb/lenet/index.html 17 18 Neural networks to recognize handwritten digits? yes Neural networks for tougher problems? not really http://pub.clement.farabet.net/ecvw09.pdf 19 NIPS 2000 • NIPS, Neural Information Processing Systems, is the premier conference on machine learning.
    [Show full text]
  • Comparative Study of Deep Learning Software Frameworks
    Comparative Study of Deep Learning Software Frameworks Soheil Bahrampour, Naveen Ramakrishnan, Lukas Schott, Mohak Shah Research and Technology Center, Robert Bosch LLC {Soheil.Bahrampour, Naveen.Ramakrishnan, fixed-term.Lukas.Schott, Mohak.Shah}@us.bosch.com ABSTRACT such as dropout and weight decay [2]. As the popular- Deep learning methods have resulted in significant perfor- ity of the deep learning methods have increased over the mance improvements in several application domains and as last few years, several deep learning software frameworks such several software frameworks have been developed to have appeared to enable efficient development and imple- facilitate their implementation. This paper presents a com- mentation of these methods. The list of available frame- parative study of five deep learning frameworks, namely works includes, but is not limited to, Caffe, DeepLearning4J, Caffe, Neon, TensorFlow, Theano, and Torch, on three as- deepmat, Eblearn, Neon, PyLearn, TensorFlow, Theano, pects: extensibility, hardware utilization, and speed. The Torch, etc. Different frameworks try to optimize different as- study is performed on several types of deep learning ar- pects of training or deployment of a deep learning algorithm. chitectures and we evaluate the performance of the above For instance, Caffe emphasises ease of use where standard frameworks when employed on a single machine for both layers can be easily configured without hard-coding while (multi-threaded) CPU and GPU (Nvidia Titan X) settings. Theano provides automatic differentiation capabilities which The speed performance metrics used here include the gradi- facilitates flexibility to modify architecture for research and ent computation time, which is important during the train- development. Several of these frameworks have received ing phase of deep networks, and the forward time, which wide attention from the research community and are well- is important from the deployment perspective of trained developed allowing efficient training of deep networks with networks.
    [Show full text]
  • Reinforcement Learning in Videogames
    Reinforcement Learning in Videogames Alex` Os´esLaza Final Project Director: Javier B´ejarAlonso FIB • UPC 20 de junio de 2017 2 Acknowledgements First I would like and appreciate the work and effort that my director, Javier B´ejar, has put into me. Either solving a ton of questions that I had or guiding me throughout the whole project, he has helped me a lot to make a good final project. I would also love to thank my family and friends, that supported me throughout the entirety of the career and specially during this project. 3 4 Abstract (English) While there are still a lot of projects and papers focused on: given a game, discover and measure which is the best algorithm for it, I decided to twist things around and decided to focus on two algorithms and its parameters be able to tell which games will be best approachable with it. To do this, I will be implementing both algorithms Q-Learning and SARSA, helping myself with Neural Networks to be able to represent the vast state space that the games have. The idea is to implement the algorithms as general as possible.This way in case someone wanted to use my algorithms for their game, it would take the less amount of time possible to adapt the game for the algorithm. I will be using some games that are used to make Artificial Intelligence competitions so I have a base to work with, having more time to focus on the actual algorithm implementation and results comparison. 5 6 Abstract (Catal`a) Mentre ja existeixen molts projectes i estudis centrats en: donat un joc, descobrir i mesurar quin es el millor algoritme per aquell joc, he decidit donar-li la volta i centrar-me en donat dos algorismes i els seus par`ametres,ser capa¸cde trobar quin tipus de jocs es beneficien m´esde la configuraci´odonada.
    [Show full text]
  • Distributed Negative Sampling for Word Embeddings Stergios Stergiou Zygimantas Straznickas Yahoo Research MIT [email protected] [email protected]
    Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI-17) Distributed Negative Sampling for Word Embeddings Stergios Stergiou Zygimantas Straznickas Yahoo Research MIT [email protected] [email protected] Rolina Wu Kostas Tsioutsiouliklis University of Waterloo Yahoo Research [email protected] [email protected] Abstract Training for such large vocabularies presents several chal- lenges. Word2Vec needs to maintain two d-dimensional vec- Word2Vec recently popularized dense vector word represen- tors of single-precision floating point numbers for each vo- tations as fixed-length features for machine learning algo- cabulary word. All vectors need to be kept in main memory rithms and is in widespread use today. In this paper we in- to achieve acceptable training latency, which is impractical vestigate one of its core components, Negative Sampling, and propose efficient distributed algorithms that allow us to scale for contemporary commodity servers. As a concrete exam- to vocabulary sizes of more than 1 billion unique words and ple, Ordentlich et al. (Ordentlich et al. 2016) describe an corpus sizes of more than 1 trillion words. application in which they use word embeddings to match search queries to online ads. Their dictionary comprises ≈ 200 million composite words which implies a main mem- Introduction ory requirement of ≈ 800 GB for d = 500. A complemen- tary challenge is that corpora sizes themselves increase. The Recently, Mikolov et al (Mikolov et al. 2013; Mikolov and largest reported dataset that has been used was 100 billion Dean 2013) introduced Word2Vec, a suite of algorithms words (Mikolov and Dean 2013) which required training for unsupervised training of dense vector representations of time in the order of days.
    [Show full text]
  • Comparative Study of Caffe, Neon, Theano, and Torch
    Workshop track - ICLR 2016 COMPARATIVE STUDY OF CAFFE,NEON,THEANO, AND TORCH FOR DEEP LEARNING Soheil Bahrampour, Naveen Ramakrishnan, Lukas Schott, Mohak Shah Bosch Research and Technology Center fSoheil.Bahrampour,Naveen.Ramakrishnan, fixed-term.Lukas.Schott,[email protected] ABSTRACT Deep learning methods have resulted in significant performance improvements in several application domains and as such several software frameworks have been developed to facilitate their implementation. This paper presents a comparative study of four deep learning frameworks, namely Caffe, Neon, Theano, and Torch, on three aspects: extensibility, hardware utilization, and speed. The study is per- formed on several types of deep learning architectures and we evaluate the per- formance of the above frameworks when employed on a single machine for both (multi-threaded) CPU and GPU (Nvidia Titan X) settings. The speed performance metrics used here include the gradient computation time, which is important dur- ing the training phase of deep networks, and the forward time, which is important from the deployment perspective of trained networks. For convolutional networks, we also report how each of these frameworks support various convolutional algo- rithms and their corresponding performance. From our experiments, we observe that Theano and Torch are the most easily extensible frameworks. We observe that Torch is best suited for any deep architecture on CPU, followed by Theano. It also achieves the best performance on the GPU for large convolutional and fully connected networks, followed closely by Neon. Theano achieves the best perfor- mance on GPU for training and deployment of LSTM networks. Finally Caffe is the easiest for evaluating the performance of standard deep architectures.
    [Show full text]
  • BUSEM at Semeval-2017 Task 4A Sentiment Analysis with Word
    BUSEM at SemEval-2017 Task 4 Sentiment Analysis with Word Embedding and Long Short Term Memory RNN Approaches Deger Ayata1, Murat Saraclar 1, Arzucan Ozgur2 1Electrical & Electronical Engineering Department, Bogaziçi University 2Computer Engineering Department, Bogaziçi University Istanbul , Turkey {deger.ayata, murat.saraclar, arzucan.ozgur}@boun.edu.tr Abstract uses word embeddings for feature representation and Support Vector Machine This paper describes our approach for (SVM), Random Forest (RF) and Naive Bayes SemEval-2017 Task 4: Sentiment Analysis in (NB) algorithms for classification Twitter Twitter. We have participated in Subtask A: messages into negative, neutral and positive Message Polarity Classification subtask and polarity. The second system is based on Long developed two systems. The first system Short Term Memory Recurrent Neural uses word embeddings for feature representation and Support Vector Machine, Networks (LSTM) and uses word indexes as Random Forest and Naive Bayes algorithms sequence of inputs for feature representation. for the classification of Twitter messages into The remainder of this article is structured as negative, neutral and positive polarity. The follows: Section 2 contains information about second system is based on Long Short Term the system description and Section 3 explains Memory Recurrent Neural Networks and methods, models, tools and software packages uses word indexes as sequence of inputs for used in this work. Test cases and datasets are feature representation. explained in Section 4. Results are given in Section 5 with discussions. Finally, section 6 summarizes the conclusions and future work. 1 Introduction 2 System Description Sentiment analysis is extracting subjective information from source materials, via natural We have developed two independent language processing, computational linguistics, systems.
    [Show full text]
  • Intel® Software Template Overview
    data analytics on spark Software & Services Group Intel® Confidential — INTERNAL USE ONLY Spark platform for big data analytics What is Apache Spark? • Apache Spark is an open-source cluster-computing framework • Spark has a developed ecosystem • Designed for massively distributed apps • Fault tolerance • Dynamic resource sharing https://software.intel.com/ai 2 INTEL and Spark community • BigDL – deep learning lib for Apache Spark • Intel® Data Analytics Acceleration Library (Intel DAAL) – machine learning lib includes support for Apache Spark • Performance Optimizations across Apache Hadoop and Apache Spark open source projects • Storage and File System: Ceph, Tachyon, and HDFS • Security: Sentry and integration into various Hadoop & Spark modules • Benchmarks Contributions: Big Bench, Hi Bench, Cloud Sort 3 Taxonomy Artificial Intelligence (AI) Machines that can sense, reason, act without explicit programming Machine Learning (ML), a key tool for AI, is the development, and application of algorithms that improve their performance at some task based on experience (previous iterations) BigDL Deep Learning Classic Machine Learning DAAL focus focus Algorithms where multiple layers of neurons Algorithms based on statistical or other learn successively complex representations techniques for estimating functions from examples Dimension Classifi Clusterin Regres- CNN RNN RBM … -ality -cation g sion Reduction Training: Build a mathematical model based on a data set Inference: Use trained model to make predictions about new data 4 BigDLand Intel DAAL BigDL and Intel DAAL are machine learning and data analytics libraries natively integrated into Apache Spark ecosystem 5 WhATis bigdl? • BigDL is a distributed deep learning library for Apache Spark • Allows to write deep learning applications as standard Spark programs • Runs on top of existing Spark or Hadoop/Hive clusters • Feature parity with popular DL frameworks.
    [Show full text]
  • Tensorflow, Theano, Keras, Torch, Caffe Vicky Kalogeiton, Stéphane Lathuilière, Pauline Luc, Thomas Lucas, Konstantin Shmelkov Introduction
    TensorFlow, Theano, Keras, Torch, Caffe Vicky Kalogeiton, Stéphane Lathuilière, Pauline Luc, Thomas Lucas, Konstantin Shmelkov Introduction TensorFlow Google Brain, 2015 (rewritten DistBelief) Theano University of Montréal, 2009 Keras François Chollet, 2015 (now at Google) Torch Facebook AI Research, Twitter, Google DeepMind Caffe Berkeley Vision and Learning Center (BVLC), 2013 Outline 1. Introduction of each framework a. TensorFlow b. Theano c. Keras d. Torch e. Caffe 2. Further comparison a. Code + models b. Community and documentation c. Performance d. Model deployment e. Extra features 3. Which framework to choose when ..? Introduction of each framework TensorFlow architecture 1) Low-level core (C++/CUDA) 2) Simple Python API to define the computational graph 3) High-level API (TF-Learn, TF-Slim, soon Keras…) TensorFlow computational graph - auto-differentiation! - easy multi-GPU/multi-node - native C++ multithreading - device-efficient implementation for most ops - whole pipeline in the graph: data loading, preprocessing, prefetching... TensorBoard TensorFlow development + bleeding edge (GitHub yay!) + division in core and contrib => very quick merging of new hotness + a lot of new related API: CRF, BayesFlow, SparseTensor, audio IO, CTC, seq2seq + so it can easily handle images, videos, audio, text... + if you really need a new native op, you can load a dynamic lib - sometimes contrib stuff disappears or moves - recently introduced bells and whistles are barely documented Presentation of Theano: - Maintained by Montréal University group. - Pioneered the use of a computational graph. - General machine learning tool -> Use of Lasagne and Keras. - Very popular in the research community, but not elsewhere. Falling behind. What is it like to start using Theano? - Read tutorials until you no longer can, then keep going.
    [Show full text]
  • DIY Deep Learning for Vision: the Caffe Framework
    DIY Deep Learning for Vision: the Caffe framework caffe.berkeleyvision.org github.com/BVLC/caffe Evan Shelhamer adapted from the Caffe tutorial with Jeff Donahue, Yangqing Jia, and Ross Girshick. Why Deep Learning? The Unreasonable Effectiveness of Deep Features Classes separate in the deep representations and transfer to many tasks. [DeCAF] [Zeiler-Fergus] Why Deep Learning? The Unreasonable Effectiveness of Deep Features Maximal activations of pool5 units [R-CNN] conv5 DeConv visualization Rich visual structure of features deep in hierarchy. [Zeiler-Fergus] Why Deep Learning? The Unreasonable Effectiveness of Deep Features 1st layer filters image patches that strongly activate 1st layer filters [Zeiler-Fergus] What is Deep Learning? Compositional Models Learned End-to-End What is Deep Learning? Compositional Models Learned End-to-End Hierarchy of Representations - vision: pixel, motif, part, object - text: character, word, clause, sentence - speech: audio, band, phone, word concrete abstract learning What is Deep Learning? Compositional Models Learned End-to-End figure credit Yann LeCun, ICML ‘13 tutorial What is Deep Learning? Compositional Models Learned End-to-End Back-propagation: take the gradient of the model layer-by-layer by the chain rule to yield the gradient of all the parameters. figure credit Yann LeCun, ICML ‘13 tutorial What is Deep Learning? Vast space of models! Caffe models are loss-driven: - supervised - unsupervised slide credit Marc’aurelio Ranzato, CVPR ‘14 tutorial. Convolutional Neural Nets (CNNs): 1989 LeNet: a layered model composed of convolution and subsampling operations followed by a holistic representation and ultimately a classifier for handwritten digits. [ LeNet ] Convolutional Nets: 2012 AlexNet: a layered model composed of convolution, + data subsampling, and further operations followed by a holistic + gpu representation and all-in-all a landmark classifier on + non-saturating nonlinearity ILSVRC12.
    [Show full text]
  • Online Power Management for Multi-Cores: a Reinforcement Learning Based Approach
    1 Online Power Management for Multi-cores: A Reinforcement Learning Based Approach Yiming Wang, Weizhe Zhang, Senior Member, IEEE, Meng Hao, Zheng Wang, Member, IEEE Abstract—Power and energy is the first-class design constraint for multi-core processors and is a limiting factor for future-generation supercomputers. While modern processor design provides a wide range of mechanisms for power and energy optimization, it remains unclear how software can make the best use of them. This paper presents a novel approach for runtime power optimization on modern multi-core systems. Our policy combines power capping and uncore frequency scaling to match the hardware power profile to the dynamically changing program behavior at runtime. We achieve this by employing reinforcement learning (RL) to automatically explore the energy-performance optimization space from training programs, learning the subtle relationships between the hardware power profile, the program characteristics, power consumption and program running times. Our RL framework then uses the learned knowledge to adapt the chip’s power budget and uncore frequency to match the changing program phases for any new, previously unseen program. We evaluate our approach on two computing clusters by applying our techniques to 11 parallel programs that were not seen by our RL framework at the training stage. Experimental results show that our approach can reduce the system-level energy consumption by 12%, on average, with less than 3% of slowdown on the application performance. By lowering the uncore frequency to leave more energy budget to allow the processor cores to run at a higher frequency, our approach can reduce the energy consumption by up to 17% while improving the application performance by 5% for specific workloads.
    [Show full text]
  • March 2012 Version 2
    WWRF VIP WG CONNECTED VEHICLES White Paper Connected Vehicles: The Role of Emerging Standards, Security and Privacy, and Machine Learning Editor: SESHADRI MOHAN, CHAIR CONNECTED VEHICLES WORKING GROUP PROFESSOR, SYSTEMS ENGINEERING DEPARTMENT UA LITTLE ROCK, AR 72204, USA Project website address: www.wwrf.ch This publication is partly based on work performed in the framework of the WWRF. It represents the views of the authors(s) and not necessarily those of the WWRF. EXECUTIVE SUMMARY The Internet of Vehicles (IoV) is an emerging technology that provides secure vehicle-to-vehicle (V2V) communication and safety for drivers and their passengers. It stands at the confluence of many evolving disciplines, including: evolving wireless technologies, V2X standards, the Internet of Things (IoT) consisting of a multitude of sensors that are housed in a vehicle, on the roadside, and in the devices worn by pedestrians, the radio technology along with the protocols that can establish an ad-hoc vehicular network, the cloud technology, the field of Big Data, and machine intelligence tools. WWRF is presenting this white paper inspired by the developments that have taken place in recent years in standards organizations such as IEEE and 3GPP and industry consortia efforts as well as current research in academia. This white paper provides insights into the state-of-the-art regarding 3GPP C-V2X as well as security and privacy of ETSI ITS, IEEE DSRC WAVE, 3GPP C-V2X. The White Paper further discusses spectrum allocation worldwide for ITS applications and connected vehicles. A section is devoted to a discussion on providing connected vehicles communication over a heterogonous set of wireless access technologies as it is imperative that the connectivity of vehicles be maintained even when the vehicles are out of coverage and/or the need to maintain vehicular connectivity as a vehicle traverses multiple wireless access technology for access to V2X applications.
    [Show full text]