Wechat ID: NeuralTalk

Embedded-AI-Report

关注模型压缩、低比特量化、移动端推理加速优化、部署

embedded-ai.bi-weekly

Awesome-Emebedded-AI

Contents

• Device Benchmark
• Papers
  • Classic
  • Overview
  • Representation
  • Structure/Pattern
  • Binarization
  • Pruning
  • Quantization
  • Low Rank Approximation
  • Distillation
  • Execution/Frameworks
• App-Experience
• Demo-Codes
  • Android
  • iOS
  • Vulkan
• Frameworks
  • general
  • inference
  • benchmark
• Course/Guide/Tutorial
• Hardware
  • GPU
• Company
• News

Device Benchmark

• 高通处理器规格表 | mydriver

CPU Devive

• 高通骁龙处理器排行榜,强大性能一览无余 | Qualcomm
• 手机CPU性能天梯图 CPU performance of mobile comparison | mydriver

GPU Device

• Adreno – Wikipedia

• GPU GFLOPS | surge.sh

• Qualcomm Adreno GPU Performance as below:
  

• 手机处理器天梯图_最新CPU天梯图_手机CPU性能天梯图 | ZOL中关村在线

Papers

Classic

• [1512.03385] Deep Residual Learning for Image Recognition
• [1610.02357] Xception: Deep Learning with Depthwise Separable Convolutions
• [1611.05431] ResNeXt: Aggregated Residual Transformations for Deep Neural Networks

Overview

• [1703.09039] Efficient Processing of Deep Neural Networks: A Tutorial and Survey
• [1707.01209] Model compression as constrained optimization, with application to neural nets. Part I: general framework
• [1707.04319] Model compression as constrained optimization, with application to neural nets. Part II: quantization

Representation

• [1707.09926] A Framework for Super-Resolution of Scalable Video via Sparse Reconstruction of Residual Frames
• [1608.01409] Faster CNNs with Direct Sparse Convolutions and Guided Pruning

• [SenSys ’16] Sparsification and Separation of Deep Learning Layers for Constrained Resource Inference on Wearables
• [IoT-App ’15] An Early Resource Characterization of Deep Learning on Wearables, Smartphones and Internet-of-Things Devices

Structure

• [1704.06904] Residual Attention Network for Image Classification [code]
• BranchyNet: Fast Inference via Early Exiting from Deep Neural Networks

• [CVPR2017] Squeeze-and-Excitation networks (ILSVRC 2017 winner) at CVPR2017
• [1707.06342] ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression
• [1707.01083] ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
• [1704.04861] MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

• [1707.06990] Memory-Efficient Implementation of DenseNets

• [1706.03912] SEP-Nets: Small and Effective Pattern Networks

Binarization

• [CVPR2017] Local Binary Convolutional Neural Networks [code]

• [1707.04693] Binarized Convolutional Neural Networks with Separable Filters for Efficient Hardware Acceleration
• [1602.02830] Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1
• [1603.05279] XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
  • [1705.09864] BMXNet: An Open-Source Binary Neural Network Implementation Based on MXNet
• [1606.06160] DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients

Pruning

• Learning to Prune: Exploring the Frontier of Fast and Accurate Parsing [code]

• [CVPR'17] Designing Energy-Efficient Convolutional Neural Networks using Energy-Aware Pruning
• [ICLR'17] Pruning Filters for Efficient ConvNets
• [ICLR'17] Pruning Convolutional Neural Networks for Resource Efficient Inference
• [ICLR'17] Soft Weight-Sharing for Neural Network Compression
• [ICLR'16] Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding
• [NIPS'16] Dynamic Network Surgery for Efficient DNNs
• [NIPS'15] Learning both Weights and Connections for Efficient Neural Networks

Quantization

• [ICML'17] The ZipML Framework for Training Models with End-to-End Low Precision: The Cans, the Cannots, and a Little Bit of Deep Learning
• [1412.6115] Compressing Deep Convolutional Networks using Vector Quantization
• [CVPR '16] Quantized Convolutional Neural Networks for Mobile Devices
• [ICASSP'16] Fixed-Point Performance Analysis of Recurrent Neural Networks
• [arXiv'16] Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations
• [ICLR'17] Loss-aware Binarization of Deep Networks
• [ICLR'17] Towards the Limit of Network Quantization
• [CVPR'17] Deep Learning with Low Precision by Half-wave Gaussian Quantization
• [1706.02393] ShiftCNN: Generalized Low-Precision Architecture for Inference of Convolutional Neural Networks

LowRankApproximation

• [CVPR'15] Efficient and Accurate Approximations of Nonlinear Convolutional Networks
  • Accelerating Very Deep Convolutional Networks for Classification and Detection
• [1511.06067] Convolutional neural networks with low-rank regularization
• [NIPS'14] Exploiting Linear Structure Within Convolutional Networks for Efficient Evaluation
• [ICLR'16] Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applications

Distillation

• [1503.02531] Distilling the Knowledge in a Neural Network
• Face Model Compression by Distilling Knowledge from Neurons

Joint Compression

• [1707.09102] Fine-Pruning: Joint Fine-Tuning and Compression of a Convolutional Network with Bayesian Optimization

Kernel Selection

• [1703.09746] Coordinating Filters for Faster Deep Neural Networks
• [1606.05316] Learning Infinite-Layer Networks: Without the Kernel Trick

Computation Precison/Resolution

• [ICML2017] Analytical Guarantees on Numerical Precision of Deep Neural Networks
• [1707.09068] Tartan: Accelerating Fully-Connected and Convolutional Layers in Deep Learning Networks by Exploiting Numerical Precision Variability
• [1708.00999] Extreme Low Resolution Activity Recognition with Multi-Siamese Embedding Learning

Model Split

• [ASPLOS’17] Neurosurgeon: Collaborative intelligence between the cloud and mobile edge
• [1705.04630] Forecasting using incomplete models

Others

• [1606.05316] Learning Infinite-Layer Networks: Without the Kernel Trick

• [1608.02893] Syntactically Informed Text Compression with Recurrent Neural Networks

• [1608.05148] Full Resolution Image Compression with Recurrent Neural Networks

• [1707.09422] Hyperprofile-based Computation Offloading for Mobile Edge Networks

• [1707.09855] Convolution with Logarithmic Filter Groups for Efficient Shallow CNN

• [1707.09597] ScanNet: A Fast and Dense Scanning Framework for Metastatic Breast Cancer Detection from Whole-Slide Images

• [1604.08772] Towards Conceptual Compression

FrameworkPaper

• [1605.04614] DeepLearningKit - an GPU Optimized Deep Learning Framework for Apple's iOS, OS X and tvOS developed in Metal and Swift
• [MobiSys '17] DeepMon: Mobile GPU-based Deep Learning Framework for Continuous Vision Applications
• [MobiSys '17] DeepEye: Resource Efficient Local Execution of Multiple Deep Vision Models using Wearable Commodity Hardware
• [EMDL '17] MobiRNN: Efficient Recurrent Neural Network Execution on Mobile GPU
• [WearSys '16] DeepSense: A GPU-based deep convolutional neural network framework on commodity mobile devices
• [IPSN '16] DeepX: A Software Accelerator for Low-Power Deep Learning Inference on Mobile Devices
• [ISCA '16] EIE: Efficient Inference Engine on Compressed Deep Neural Network
• [MobiSys '16] MCDNN: An Approximation-Based Execution Framework for Deep Stream Processing Under Resource Constraints
• [MobiCASE '16] DXTK: Enabling Resource-efficient Deep Learning on Mobile and Embedded Devices with the DeepX Toolkit
• [MM '16] CNNdroid: GPU-Accelerated Execution of Trained Deep Convolutional Neural Networks on Android

Experience

• 【技术共享】怎么把人脸检测的速度做到极致
• 基于OpenGL ES 的深度学习框架编写 - jxt1234and2010的专栏 - CSDN博客

Codes

Model Compression

• yonghenglh6/DepthwiseConvolution: A personal mobile convolution implementation on caffe by liuhao.(only GPU)
• liuzhuang13/DenseNet: Densely Connected Convolutional Networks, In CVPR 2017 (Best Paper Award)
• kevinzakka/DenseNet: PyTorch Implementation of "Densely Connected Convolutional Networks"
• hollance/MobileNet-CoreML: The MobileNet neural network using Apple's new CoreML framework
• AngusG/tensorflow-xnor-bnn: BinaryNets in TensorFlow with XNOR GEMM op  
• jonathanmarek1/binarynet-tensorflow
• farmingyard/caffe-mobilenet: A caffe implementation of mobilenet's depthwise convolution layer  
• kedartatwawadi/NN_compression
• chuanqi305/MobileNet-SSD: Caffe implementation of Google MobileNet SSD detection network, with pretrained weights on VOC0712 and mAP=0.727.

Model Encryption

• OpenMined/Syft: Homomorphically Encrypted Deep Learning Library

Model Application

• msracver/FCIS: Fully Convolutional Instance-aware Semantic Segmentation
• bearpaw/PyraNet: Code for "Learning Feature Pyramids for Human Pose Estimation" (ICCV 2017)
• aquaviter/iot-demo-mxnet-greengrass
• bearpaw/PyraNet: Code for "Learning Feature Pyramids for Human Pose Estimation" (ICCV 2017)
• CongWeilin/mtcnn-caffe: Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks  
• foreverYoungGitHub/MTCNN: Repository for "Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks", implemented with Caffe, C++ interface.
• OAID/mtcnn: C++ project to implement MTCNN, a perfect face detect algorithm, on different DL frameworks. The most popular frameworks: caffe/mxnet/tensorflow, are all suppported now
• Seanlinx/mtcnn: this repository is the implementation of MTCNN in MXnet
• LaoDar/cnn_head_pose_estimator: a simple and fast mxnet version CNN based head pose estimator

AR

• ProjectDent/ARKit-CoreLocation: Combines the high accuracy of AR with the scale of GPS data
• bjarnel/arkit-tictactoe: Tic-Tac-Toe implemented using ARKit+Scenekit
• arirawr/ARKit-FloorIsLava: Basic ARKit example that detects planes and makes them lava.
• exyte/ARTetris: Augmented Reality Tetris made with ARKit and SceneKit
• bjarnel/arkit-portal: Simple portal demo implemented with ARKit+SceneKit, the trick is to change the rendering order and render invisible "masks" to hide what's inside.
• bjarnel/scenekit-tictactoe

Android

• harvardnlp/nmt-android: Neural Machine Translation on Android
• TensorFlow Android Camera Demo
• KleinYuan/Caffe2-iOS: Caffe2 on iOS Real-time Demo. Test with Your Own Model and Photos.
• MXNet Android Classification App - Image classification on Android with MXNet.
• bwasti/AICamera: Demonstration of using Caffe2 inside an Android application.
• mtmd/Mobile_ConvNet: RenderScript based implementation of Convolutional Neural Networks for Android phones

• madeye/yolo-android: Quantized Tiny Yolo Demo on Android

iOS

• MXNet iOS Classification App - Image classification on iOS with MXNet.
• Compile MXnet on Xcode (in Chinese) - a step-by-step tutorial of compiling MXnet on Xcode for iOS app
• KleinYuan/Caffe2-iOS: Caffe2 on iOS Real-time Demo. Test with Your Own Model and Photos.
• KimDarren/FaceCropper: Crop faces, inside of your image, with iOS 11 Vision api.
• hollance/TensorFlow-iOS-Example: Source code for my blog post "Getting started with TensorFlow on iOS"

• kingreza/SeeFood: Inspired by HBO's Silicon Valley: SeeFood is an iOS app that uses CoreML to detect various dishes
• hollance/TensorFlow-iOS-Example: Source code for my blog post "Getting started with TensorFlow on iOS"
• Naituw/CoreMLDemo: Demo for CoreML & Vision Framework

Vulkan

• SaschaWillems/Vulkan: Examples and demos for the new Vulkan API
• ARM-software/vulkan-sdk: ARM Vulkan SDK
• alexhultman/libvc: Vulkan Compute for C++ (experimentation project)

Frameworks & Acceleration Library

• Deep Learning in a Single File for Smart Devices — mxnet
• ARM-software/ComputeLibrary: The ARM Computer Vision and Machine Learning library is a set of functions optimised for both ARM CPUs and GPUs using SIMD technologies Intro
• Apple CoreML
• Microsoft Embedded Learning Library
• mil-tokyo/webdnn: Fastest DNN Execution Framework on Web Browser
• jiaxiang-wu/quantized-cnn: An efficient framework for convolutional neural networks
• Tencent/ncnn: ncnn is a high-performance neural network inference framework optimized for the mobile platform
• JDAI-CV/dabnn: dabnn is an accelerated binary neural networks inference framework for mobile platform

• Darknet with NNPACK: NNPACK was used to optimize Darknet without using a GPU. It is useful for embedded devices using ARM CPUs  
• naibaf7/libdnn: Greentea LibDNN - a universal convolution implementation supporting CUDA and OpenCL
• blei-lab/edward: A library for probabilistic modeling, inference, and criticism. Deep generative models, variational inference. Runs on TensorFlow
• dmlc/nnvm-fusion: Kernel Fusion and Runtime Compilation Based on NNVM

Benchmark

• baidu-research/DeepBench: Benchmarking Deep Learning operations on different hardware

• hollance/BNNS-vs-MPSCNN: Compares the speed of Apple's two deep learning frameworks: BNNS and Metal Performance Shaders
• DeepMark/deepmark: THE Deep Learning Benchmarks

Convertor

Model convertor. More convertors please refer deep-learning-model-convertor

• NervanaSystems/caffe2neon: Tools to convert Caffe models to neon's serialization format

Mobile Video Process Library/Player

• MTG/essentia: C++ library for audio and music analysis, description and synthesis, including Python bindings
• Pili-完美直播体验（Pili Streaming Cloud）  
• pili-engineering/PLDroidMediaStreaming: PLDroidMediaStreaming 是 Pili 直播 SDK 的 Android 推流端，支持 RTMP 推流，h.264 和 AAC 编码，硬编、软编支持。具有丰富的数据和状态回调，方便用户根据自己的业务定制化开发。具有直播场景下的重要功能，如：美颜、背景音乐、水印等功能。PLDroidMediaStreaming 是现在目前重点维护的版本，自带采集模块也支持用户自己做采集端。
• pili-engineering/PLDroidShortVideo: PLDroidShortVideo 是七牛推出的一款适用于 Android 平台的短视频 SDK，提供了包括美颜、滤镜、水印、断点录制、分段回删、视频编辑、混音特效、本地/云端存储在内的多种功能，支持高度定制以及二次开发。
• pili-engineering/PLDroidPlayer: PLDroidPlayer 是 Pili 直播 SDK 的安卓播放器。支持所有直播常用的格式，如：RTMP、HLS、FLV。拥有优秀的功能和特性，如：首屏秒开、追帧优化、丰富的数据和状态回调、硬解软解支持。而且可以根据自己的业务进行高度定制化开发。
• pili-engineering/PLMediaStreamingKit: PLMediaStreamingKit 是 Pili 直播 SDK 的 iOS 推流端，支持 RTMP 推流，h.264 和 AAC 编码，硬编、软编支持。具有丰富的数据和状态回调，方便用户根据自己的业务定制化开发。具有直播场景下的重要功能，如：美颜、背景音乐、水印等功能。
• pili-engineering/PLShortVideoKit: PLShortVideoKit 是七牛推出的一款适用于 iOS 平台的短视频 SDK，提供了包括美颜、滤镜、水印、断点录制、分段回删、视频编辑、混音特效、本地/云端存储在内的多种功能，支持高度定制以及二次开发。
• pili-engineering/PLPlayerKit: PLPlayerKit 是 Pili 直播 SDK 的 iOS 播放器。支持所有直播常用的格式，如：RTMP、HLS、FLV。拥有优秀的功能和特性，如：首屏秒开、追帧优化、丰富的数据和状态回调、硬解软解支持。而且可以根据自己的业务进行高度定制化开发。
• pili-engineering/PLPlayerKit: PLPlayerKit 是 Pili 直播 SDK 的 iOS 播放器。支持所有直播常用的格式，如：RTMP、HLS、FLV。拥有优秀的功能和特性，如：首屏秒开、追帧优化、丰富的数据和状态回调、硬解软解支持。而且可以根据自己的业务进行高度定制化开发。

Other Toolkit

• facebook/fb-caffe-exts: Some handy utility libraries and tools for the Caffe deep learning framework.
• Samsung/iotjs: Platform for Internet of Things with JavaScript code
• hollance/Forge: A neural network toolkit for Metal
• christopher5106/FastAnnotationTool: A tool using OpenCV to annotate images for image classification, optical character reading, etc.  
• raphui/rnk: rnk is a RTOS targeting ARM architecture.

Data Set

• HandNet - A dataset of depth images of hands

Course

This part contains related course, guides and tutorials.

• Deep learning systems: UW course schedule(focused on systems design, not learning)
• Squeezing Deep Learning Into Mobile Phones
• Deep Learning – Tutorial and Recent Trends
• Efficient Convolutional Neural Network Inference on Mobile GPUs
• ARM® Mali™ GPU OpenCL Developer Guide html pdf
• Optimal Compute on ARM MaliTM GPUs
• GPU Compute for Mobile Devices
• Compute for Mobile Devices Performance focused
• Hands On OpenCL
• Adreno OpenCL Programming Guide
• Better OpenCL Performance on Qualcomm Adreno GPU

• Tutorial on Hardware Architectures for Deep Neural Networks | MIT MICRO-50
• 基于mtcnn和facenet的实时人脸检测与识别系统开发 | 知乎专栏
• Creating insanely fast image classifiers with MobileNet in TensorFlow | HACKERNOON
• How to squeeze the most from your training data | KDNUGGETS
• Ubuntu16.04腾讯NCNN框架入门到应用 | CSDN
• Building Cross-Platform CUDA Applications with CMake | NVIDIA
• Caffe2 Bay Area Meetup (5/31/2017) | YouTube

Hardware

GPU

• Bifrost GPU architecture and ARM Mali-G71 GPU
• Midgard GPU Architecture
• ARM Mali-T880 GPU
• Mobile GPU market share

Company

• Lift: A novel approach to achieving performance portability on parallel accelerators. | Where High-Level Programming Meets Performance Portability
• mlmodelzoo.com – deep learning models on mobile

News

2017-08-07

• OpenCV 3.3版本发布
• 鱼和熊掌兼得，DNN加入 OpenCV 全家桶 | 知乎专栏
• Qualcomm Snapdragon Neural Processing Engine (NPE) | Qualcomm Developer Network
• AI让芯片业洗牌: 苹果、微软和谷歌挤入赛道，英特尔、英伟达、高通、AMD几家欢乐几家愁 | 新智元
• 解密图森：英伟达为何投资这家无人车公司；估值18亿背后有位长者 | 量子位  
• 被英伟达相中，给Tier1供货，天瞳威视仅靠AI就搞定ADAS | 车东西
• ARM的最新NB-IoT报告 | 5G
• ARM发飙！几个月后手机处理器将因它们而变天！ | 智趣狗
• 人工智能和云计算让芯片业洗牌，英特尔成了最大输家 | 量子位
• The Rise of AI Is Forcing Google and Microsoft to Become Chipmakers | WIRED
• 如何评价腾讯刚出的ncnn库？ | 知乎
• 沈向洋宣布微软开发 AI 芯片HPU，剑指英伟达等芯片巨头软肋 | 新智元  
• 超越GPU，FPGA、ASIC和更智能的手机 ｜ 新智元  
• "TensorFire - runs neural networks in the browser using WebGL" [Demo: style-transfer]
• Getting Started with Neural Compute Stick and Rasbperry Pi 3 | YouTube

2017-07-24

• We ported CAFFE to HIP - and here’s what happened… - GPUOpen
• Clarifai launches SDK for training AI on your iPhone | VentureBeat | AI | by Khari Johnson