Deep Learning to Big Data Analytics on Apache Spark Using BigDL with Xianyan Jia and Yuhao Yang

Intel® Confidential — INTERNAL USE ONLY Deeplearningtobigdataanalyticsonapache spark*usingbigdl Yuhao Yang (yuaho.yang@intel.com) Xianyan Jia (Xianyan.jia@intel.com) Big Data Technology, Software and Service Group, Intel

2 Outline BigDL § Apache Spark* + High Performance + Deep Learning Speech recognition: § Deep Speech 2 on BigDL: ML Pipeline + BigDL Object detection: § Faster RCNN and SSD on BigDL

3 What is BigDL? BigDL is a distributed deep learning library for Apache Spark*

4 BigDL: Deep learning on Apache Spark* BigDL open sourced on Dec 30, 2016 § Apache Spark*, MKL Acceleration, High performance Rich function § AlexNet, GoogleNet, VGG, Faster R-CNN, SSD, Deep Speech, Recommendation… § Scala/Java + Python § AWS EC2, TensorBoard, Notebook, caffe/torch load/export… Popularity § Support from Cloud: Microsoft, Amazon, Cloudera, Databricks… § Community. 1700+ stars

5 Basic Component Tensor: § ND-array data structure § Generic data type § Rich and fast math operations (powered by Intel MKL) Layers § 113+ layers (Conv, 3D Conv, Pooling, 3D Pooling, FC …) Criterion § 23+ criterions (DiceCoefficient, ClassNLL, CrossEntropy …) Optimization § SGD, Adagrad, LBFGS § Community contribution: Adam, Adadelta, RMSprop, Adamx

6 Run as standard Apache Spark* jobs

8 Speech Recognition Challenges § Audio à text § Speaker variability, Channel variability, Different languages Solutions: o Hybrid system: – DNNs, Hidden Markov Models (HMMs), context-dependent phone models, Lexicon models, GMM. – Domain expertise and multi-stage ü DNN end to end: – DNN. Much easier – More data, better model

9 Deep Speech 2 for Speech Recognition • “The Deep Speech 2 ASR pipeline approaches or exceeds the accuracy of Amazon Mechanical Turk human workers on several benchmarks, works in multiple languages with little modification, and is deployable in a production setting.” • “Table 13 shows that the DS2 system outperforms humans in 3 out of the 4 test sets and is competitive on the fourth. Given this result, we suspect that there is little room for a generic speech system to further improve on clean read speech without further domain adaptation.” https://arxiv.org/pdf/1512.02595.pdf

10 Deep Speech 2 on BigDL feature extraction DNN model decoder CAB Languag e Model

11 Deep Speech 2 on BigDL: Feature transformers Apache Spark* ML Pipeline flac/wav file reader DFTSpecgram Segmenter MelFreqFilterBankTranscriptVectorizer windower

12 Deep Speech 2 on BigDL: Model 9 layers biRNN: >50 Million parameters ! ... conv biRNN 1 biRNN 2 biRNN k affine softmax CTC

13 Deep Speech 2 on BigDL: CTC Loss Connectionist Temporal Classification § a loss function useful for performing supervised learning on sequence data, without needing an alignment between input data and labels. (Alex Graves etc. 2006) § Raw waveforms and text transcription BigDL developed first open source CTC on Java/Scala § Loss/Gradient in consistency with baidu/warp-ctc § JNI version about 3X faster than Scala version, but CTC only takes 0.2% of the training time. http://www.cs.toronto.edu/~graves/icml_2006.pdf “…BigDL help users rundeep learning on Spark…”

14 Deep Speech 2 on BigDL: Model training 5% 75% 5% 0%0% 15% Training time batchNormalization Recurrent Linear CTC Convolution Other With libriSpeech, 5 RNN layer, 30 seconds uttLength, 30 epoches.

15 Deep Speech 2 on BigDL: Decoder Existing decoder: § BestPathDecoder (argmax) wer = 27% § VocabDecoder wer = 22% Contribution welcome § decoder with Language Model & beam search, expect wer < 10% t1 t2 t3 t4 … A 0.01 0.1 0.7 0.03 B 0.05 0.01 0.2 0.5 C 0.01 0.1 0.09 0.01 D 0.8 0.05 0.01 0.01 … … … … … Blank 0.1 0.6 0.01 0.4 D - A B

16 Deep Speech 2 with AN4 data • Deep Speech 2 (8 layer, 5 RNN), uttLength 8 seconds • Word Error Rate with hold-out validation dataset wer(without LM) Deep Speech on Tensorflow 12.4% BigDL < 5%

17 Deep Speech 2 with LibriSpeech • Deep Speech 2 (12 layers, 9 RNN), uttLength 30 seconds • Word Error Rate with hold-out validation dataset • Still under further tuning and optimization. • More training data • Optimizer (Adam, SGD, nesterov ) cer wer(without LM) Hannun, et al. (2014) 10.7 35.8 Graves-Jaitly (ICML 2014) 9.2 30.1 Hwang-Sung (ICML2016) 10.6 38.4 BigDL 8.7 32.4

18 Deep Speech 2 on BigDL: Summary Feature transformers: § Flac/wav Reader, Windower, TimeSegmenter, TranscriptVectorizer, DFTSpecgram, MelFrequencyFilterBank Model training and inference § Big DL container, optimizer, Convolution, BatchNormalization, Bi-RNN CTC (Connectionist Temporal Classification) loss § Scala or JNI (warp-ctc) Decoder § ArgmaxDecoder, VocabDecoder Evaluation § wer, cer

20 SSD: Single Shot Multibox Detector Images from PASCAL(http://host.robots.ox.ac.uk/pascal/VOC/) Liu, Wei, et al. "SSD: Single shot multiboxdetector." European Conference on Computer Vision. Springer International Publishing, 2016. § State-of-the-art object detection pipeline § Single shot

The Single Shot Detector (SSD) base network fc6 and fc7 converted to convolutional Multi-scale feature maps for detection: observe how conv feature maps decrease in size and allow predictions at multiple scales collection of bounding boxes and scores for each category Convolutional predictors for detection https://arxiv.org/abs/1512.02325

22 SSD Pipeline Preprocess Resize Normalize ToBatch BigDL SSD Model Pre-trained Model Post- processor Boxes & scores Visualizer Raw Data Ground Truth MAP RDD[RoiByteImage] RDD[Tensor] Validation prediction

23 SSD + VGG 300x300 test over Pascal VOC 2007 • SSD + VGG 300x300 with pretrained model over voc07+12 • Mean Average Precision Caffe Model BigDL MAP 77.2 77.3

24 SSD + VGG 512x512 test over Pascal VOC 2007 • SSD + VGG 512x512 with pretrained model over voc07+12 • Mean Average Precision Caffe Model BigDL MAP 79.6 79.6

25 analytics-zoo https://github.com/intel-analytics/analytics-zoo Join Our Mail List bigdl-user-group+subscribe@googlegroups.com Report Bugs And Create Feature Request https://github.com/intel-analytics/BigDL/issues https://software.intel.com/ai BigDL Community

27 Legal Disclaimer INFORMATION IN THIS DOCUMENT IS PROVIDED IN CONNECTION WITH INTEL PRODUCTS. NO LICENSE, EXPRESS OR IMPLIED, BY ESTOPPEL OR OTHERWISE, TO ANY INTELLECTUAL PROPERTY RIGHTS IS GRANTED BY THIS DOCUMENT. EXCEPT AS PROVIDED IN INTEL'S TERMS AND CONDITIONS OF SALE FOR SUCH PRODUCTS, INTEL ASSUMES NO LIABILITY WHATSOEVER AND INTEL DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY, RELATING TO SALE AND/OR USE OFINTEL PRODUCTS INCLUDING LIABILITY OR WARRANTIES RELATING TO FITNESS FOR A PARTICULAR PURPOSE, MERCHANTABILITY, OR INFRINGEMENT OF ANY PATENT, COPYRIGHT OR OTHER INTELLECTUAL PROPERTY RIGHT. A "Mission Critical Application" is any application in which failure of the Intel Product could result, directly or indirectly, in personal injury or death. SHOULD YOU PURCHASE OR USE INTEL'S PRODUCTS FOR ANY SUCH MISSION CRITICALAPPLICATION, YOU SHALL INDEMNIFY AND HOLD INTEL AND ITS SUBSIDIARIES, SUBCONTRACTORS AND AFFILIATES, AND THE DIRECTORS, OFFICERS, AND EMPLOYEES OF EACH, HARMLESS AGAINST ALLCLAIMS COSTS, DAMAGES, AND EXPENSES AND REASONABLE ATTORNEYS' FEES ARISING OUT OF, DIRECTLY OR INDIRECTLY, ANY CLAIM OF PRODUCT LIABILITY, PERSONALINJURY,OR DEATH ARISING IN ANY WAY OUT OFSUCH MISSION CRITICALAPPLICATION, WHETHER OR NOT INTEL OR ITS SUBCONTRACTOR WAS NEGLIGENT IN THE DESIGN, MANUFACTURE, OR WARNING OFTHE INTEL PRODUCT OR ANY OF ITS PARTS. Intel may make changes to specifications and product descriptions at any time, without notice. Designers must not rely on the absence or characteristics of any features or instructions marked "reserved" or "undefined". Intel reserves thesefor future definition and shall haveno responsibility whatsoever for conflicts or incompatibilities arising from future changes to them. The information here is subject to change without notice. Do not finalize a design with this information. The products described in this document may contain design defects or errors known as erratawhich may cause the product to deviate from published specifications. Current characterized errata are available on request. Contact your local Intel sales office or your distributor to obtain the latest specifications and before placing your product order. Copies of documents which have an order number and are referenced in this document, or other Intel literature, may be obtained by calling 1-800-548-4725, or go to: http://www.intel.com/design/literature.htm Intel, Quark, VTune, Xeon, Cilk, Atom, LookInside and the Intel logo are trademarks of Intel Corporation in the United States and other countries. *Other names and brands may be claimed as the property of others. Copyright ©2015 Intel Corporation.

28 Risk Factors The above statements and any others in this document that refer to plans and expectations for the first quarter, the year and the future are forward-looking statements that involve a number of risks and uncertainties. Words suchas “anticipates,” “expects,” “intends,” “plans,” “believes,” “seeks,” “estimates,” “may,” “will,” “should” and their variations identify forward-lookingstatements. Statements that refer to or are based on projections, uncertain events or assumptionsalsoidentify forward-looking statements. Many factorscould affect Intel’s actual results, and variances fromIntel’s current expectations regardingsuch factorscould cause actual results to differ materially from those expressed in these forward-lookingstatements. Intel presently considers the followingto be the important factors that could cause actual results to differ materially from the company’sexpectations. Demand could be different from Intel's expectations due to factors includingchanges in business and economic conditions; customer acceptance of Intel’s and competitors’ products; supply constraints andother disruptions affecting customers; changes in customer order patterns includingorder cancellations; andchanges in the level of inventory at customers. Uncertainty in globaleconomic and financial conditions poses a risk that consumers and businesses may defer purchases in response to negative financial events, which could negatively affect product demand andother related matters. Intel operates in intensely competitive industries that are characterized by a high percentage of costs that are fixed or difficult to reduce in the short term and product demand that is highly variable and difficult to forecast. Revenue and the gross margin percentage are affected by the timing of Intel product introductions and the demand for and market acceptance of Intel's products; actions taken by Intel's competitors, including product offeringsand introductions, marketing programsand pricingpressures and Intel’s response to such actions; andIntel’s ability to respondquickly to technologicaldevelopments and to incorporate new features into its products. The gross margin percentage couldvary significantly fromexpectations based on capacity utilization; variations in inventory valuation, including variations related to the timing of qualifying products for sale; changesin revenue levels; segment product mix; the timing and execution of the manufacturing ramp and associated costs; start-up costs; excessor obsolete inventory; changes in unit costs; defects or disruptionsin the supply of materials or resources; product manufacturingquality/yields; and impairments of long-lived assets, including manufacturing, assembly/test and intangible assets. Intel's results couldbe affected by adverse economic, social, political and physical/infrastructure conditions in countries where Intel, its customers orits suppliers operate, including military conflict and other security risks, naturaldisasters, infrastructure disruptions, health concerns and fluctuations in currency exchange rates. Expenses, particularly certain marketing and compensation expenses, as well as restructuringand asset impairment charges, vary dependingon the level of demand for Intel's products and the level of revenue and profits. Intel’s resultscould be affected by the timing of closing of acquisitions and divestitures. Intel's results could be affected by adverse effects associated with product defects and errata (deviations from published specifications), and by litigation or regulatory matters involving intellectual property, stockholder, consumer, antitrust, disclosure and other issues, such as the litigation and regulatory matters described in Intel's SEC reports. An unfavorable ruling could include monetary damages oran injunction prohibiting Intel frommanufacturingor selling one or more products, precluding particular business practices, impacting Intel’s ability to design its products, or requiring other remedies suchas compulsory licensingof intellectual property. A detailed discussion of these and other factors that could affect Intel’s results is included in Intel’s SEC filings, includingthe company’s most recent reports on Form10-Q, Form 10-K and earnings release.

Deep Learning to Big Data Analytics on Apache Spark Using BigDL with Xianyan Jia and Yuhao Yang

More Related Content

What's hot

Similar to Deep Learning to Big Data Analytics on Apache Spark Using BigDL with Xianyan Jia and Yuhao Yang

More from Databricks

Recently uploaded

Deep Learning to Big Data Analytics on Apache Spark Using BigDL with Xianyan Jia and Yuhao Yang