Upgrade to Pro — share decks privately, control downloads, hide ads and more …

To aim for democratization, Machine learning that anyone can use

SISO
August 04, 2017
320

To aim for democratization, Machine learning that anyone can use

LT session 2 of the "TENSORFLOW aims, machine learning that anyone can use" (https://elv.connpass.com/event/61466/).

SISO

August 04, 2017
Tweet

Transcript

  1. 自己紹介 • 現在、♩ n ♫ n 株式会社にてDeapLearning 開発業務 • 過去に携わった業務

    • メインフレーム、スーパーコンピュータ(HPC)やダイナミックリコンフィギャラブル プロセッサなどの開発 • リアルタイムアクセラレータ開発、デモアプリ:Pedestrian Dead-Reckoning(PDR) • モータコントロールマイコン、グラフィックスマイコンの製品企画 • 国内3Dグラフィックススタートアップ企業支援 • シリコンバレーでUSコンピュータベンダとの協業
  2. 誰もが使える機会学習・・・これから(CONSUMER 向け) • High Level DNN Libraryがあったらいいなぁ!! ➡ Keras with

    TensorFlow • モバイルで動くDNNがあったら簡単に試せるかも?! ➡TensorFlow Lite • 過去、Direct3Dが出たころ思い出します・・・ ➡ Direct DNN プリミティブモデルのプログラミング そして、Unity, Unreal engine..統合環境へ
  3. DEEP LEARNING PORTABILITY ➡ON THE MOBILE≒EDGE COMPUTING • Movidius™ Neural

    Compute Stick • Framework: Caffe • Host computer: • Ubuntu Linux* 16.04 LTS x86-64-bit with a USB 2.0 port and an active internet connection • Toolkit features: • Compilation: Translation of network weights and structures from Caffe deep learning frameworks into a Movidius™ Neural Compute Stick compatible format. • Profiling: The generation of detailed, per-layer performance statistics of how your network is running on the Movidius™ Neural Compute Stick. • Checking: Verification of classification accuracy by running inferences with the Movidius Neural Compute Stick. https://developer.movidius.com/ • Headline Features • Convolutions, NxN Convolution with Stride S. The following cases have been extensively tested:1x1s1,3x3s1,5x5s1,7x7s1, 7x7s2, 7x7s4, Group convolution, Depth Convolution kernel sizes 3x3, 5x5, 7x7, 9x9, strides 1, 2, 3, 4, 8 • Max Pooling Radix NxM with Stride S, Average Pooling Radix NxM with Stride S, Local Response Normalization, Relu, Relu-X, Prelu, Softmax, Sigmoid,Tanh, Deconvolution, Slice, Scale, ElmWise unit, Fully Connected Layers • Element wise operations : sum, prod, max • Elu, Reshape, Flatten, Power, Crop
  4. DEEP LEARNING PORTABILITY (CONT’D) ➡ON THE MOBILE≒EDGE COMPUTING • Snapdragon

    Neural Processing Engine • What's in the SDK? • Android and Linux runtimes for neural network model execution • Acceleration support for Qualcomm Hexagon™ DSPs, Qualcomm Adreno™ GPUs and Qualcomm Kryo™, CPUs2 • Support for models in Caffe, Caffe2 and TensorFlow formats3 • APIs for controlling loading, execution and scheduling on the runtimes 1 Indicative average numbers from internal testing, measured by running an inference use case with an Inception-v3 model on a Qualcomm® Snapdragon™ 835 device. https://developer.qualcomm.com/software/snap dragon-neural-processing-engine
  5. TENSORFLOW ANDROID CAMERA DEMO • Description • The demos in

    this folder are designed to give straightforward samples of using TensorFlow in mobile applications. • Inference is done using the TensorFlow Android Inference Interface, which may be built separately if you want a standalone library to drop into your existing application. Object tracking and efficient YUV -> RGB conversion are handled by libtensorflow_demo.so. • A device running Android 5.0 (API 21) or higher is required to run the demo due to the use of the camera2 API, although the native libraries themselves can run on API >= 14 devices. • Current samples: • TF Classify: Uses the Google Inception model to classify camera frames in real-time, displaying the top results in an overlay on the camera image. • TF Stylize: Uses a model based on A Learned Representation For Artistic Style to restyle the camera preview image to that of a number of different artists. • TF Detect: Demonstrates a model based on Scalable Object Detection using Deep Neural Networks to localize and track people in the camera preview in real-time. https://github.com/androidthings/sample-tensorflow-imageclassifier https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/android