Install caffe framework and run Faster R-CNN demo on Ubuntu 16.04. Test environment CPU: Intel(R) Core(TM) i3-4130 CPU @ 3.40GHz 4-Cores GPU: ASUSTeK GeForce GTX 1060 with 6GB Memory HD: WDC WD5000AAKX OS: Ubuntu 16.04 Test Flow Install software requirement Video pre-processing: get jpeg images from source video Image Labeling Use Faster R-CNN to genrate trained model Run Faster R-CNN demo Requirement Hardware: Good graphic card with large memory (6GB memory is okay, but it has problem in VGG traing.) ...

以前球隊在回防時, 教練總會說"快跑! 別回頭, 到定點再休息". 關於人生, 喘息點在哪兒? 我想就在滴水穿石之間! Deep Learning <–> keep learning Week31 (08/04) 網路文章 A Brief History of CNNs in Image Segmentation: From R-CNN to Mask R-CNN R-CNN: https://arxiv.org/abs/1311.2524 Visual Recognition就是從這裡爆發了… 透過Selective Search找到Region後丢給修改過的AlexNet學習, 最後再給Support Vector Machine (SVM) 這層CNN. 原本設計的R-CNN就只做這些: 用Selective Search找region 透過pre-trained AlexNet去算特徵後交給SVM看region裡的是什麼鬼東西 最後透過線性回歸(linear regression)畫出物件的座標 Fast R-CNN: https://arxiv.org/abs/1504.08083 Region proposals using Selective Search. RoI (Region of Interest) Pooling Combine All Models into One Network 將R-CNN的AlexNet, SVM, Regressor整合在一起變成單一個network Fast R-CNN instead used a single network to compute the extract image features (CNN), classify (SVM), and tighten bounding boxes (regressor). Faster R-CNN: https://arxiv.org/abs/1506.01497 用SS找region的方法太慢, 又重覆計算, 改由RPN後, 整個framework只要訓練一個CNN. Speeding Up Region Proposal: Selective Search is too slow. A single CNN is used to both carry out region proposals and classification. only one CNN needs to be trained Region Proposal Network(RPN) - How the Regions are Generated ...

以前球隊在回防時, 教練總會說"快跑! 別回頭, 到定點再休息". 關於人生, 喘息點在哪兒? 我想就在滴水穿石之間! Week27 (07/07) 網路文章 Elon Musk: The world’s population is accelerating toward collapse and nobody cares 網路資源 Barbara Oakley: “Learning How to Learn” | Talks at Google Week28 (07/14) 網路資源 Numpy Quick Start MathJax basic tutorial and quick reference MathJax語法速查及範例 Week29 (07/21) 網路文章 Path Guide: A New Approach to Indoor Navigation 讀書心得 The Numpy Stack in Python 用簡短的方式介紹numpy, pandas, matplotlib, scipy. numpy: 在matrix, vector處理會比ptython built-in還快很多且方便. 主要的object type是numpy.array, 可以跟list object互相轉換. pandas: 強大的DataFrames在處理csv files時可輕鬆取出及處理想要的內容. pandas.DataFrame與pandas.Series是很重要的object. matplotlib: 把資料圖形化 scipy: 介紹了常態分佈(normal distribution)又稱高斯分佈(Gaussian distribution). Week30 (07/28) 網路文章 A Neural Network in 11 lines of Python 詳細的說明ANN的實現方法, 但沒提到bias?? ...

以前球隊在回防時, 教練總會說"快跑! 別回頭, 到定點再休息". 關於人生, 喘息點在哪兒? 我想就在滴水穿石之間! Week22 (06/02) 讀書心得 OSDI: Memory Management Kernel space的memory management, 要考慮到的問題, Internal/External fregmentation, physical/logical contiguous, buddy system, slab allocator. User space的memory管理, page fault handler, Red-Black Tree / AVL Tree. 論文學習 FarmBeats: An IoT Platform for Data-Driven Agriculture PDF ,Slides FarmBeats這套系統提出了几個解決方法: 遠距傳輸, 以往的解法是用satellite或是手機訊號, 這裡是使用TVWS(TV White Space)以減少硬體成本, TVWS是利用數位電視未使用的頻寬做傳輸, 屬於低頻有傳輸距離長, 穿透力強的優點. 天氣感知系統, 透過OpenWeather API的資料來規劃電量的使用及電池的充電與否. 風力輔助系統, 收集農地的風向及風力數據, 透過飛行演算法(Flight Planning Algorithm)規劃出適當的路線, 減少逆風造成的額外耗電量. 本地影像及資料壓縮, 一般Drone的航空影像(Aerial Imagery)或sensors資料都是直接傳到雲端, 透過Farmbeats的gateway做中繼站並壓縮影像及sensors資料, 可支援1星期的在網路無法使用情況. Week23 (06/09) 網路文章 When should we use mutex and when should we use semaphore ...

Memory management主要的功能是page mapping, memory protection, and sharing, 但隨著時間不斷的演進已經算是kernel裡不小的subsystem. 這份研究分析2009~2015年之間4587筆有關memory management(mm)的patches, Linux版本由v2.6.32.1 ~ v4.0-rc4. Source: An Evolutionary Study of Linux Memory Management for Fun and Profit Memory Bugs 5種bugs存在mm: memory error, checking, concurrency, logic and programming. 透過heat map可清楚看到主要bug發生地方 主要的bug fix在: Memory Allocation, Garbage Collection (GC), Virtual Memory Management. 大部份memory leak並不是因為忘了free, 而是錯誤的page fault handling和 free page的計算. MM較大的問題是很難去track正確的狀態. Memory optimization 定義3種optimization patches data structure: 避免nested data structure. Scalability的實現, scalability問題是因為locking for atomic access to shared data structures. memory policy: 使用合理的policy design (latency/throughput, sync/async, lazy/non-lazy, local/global, fairness/performance). fast path: 加速頻繁使用的source code, reduction跟lockless optimization是廣泛被使用. Optimistic barrier是為了減少呼叫barrier/fence system call時的synchronous overheads. (Code reduction, lockless optimization, new function, state caching, inline, code shifting, group excution, optimistic barrier) MM常見的Data structure Radix tree: In adress_space, 主要特色是有效率的存放(sparse)資料. Red-black tree: In vm_area_struct, 可快速的search, insert, delete. 相較於AVL 雖然在search上較快但需要額外的空間, insert/delete比較慢, rotation比較困難. Bitmap: 通常用在RAM的page indexing. List: 廣泛使用的DS, 例如LRU(Last Recently Used) list, 用來追蹤active/inactive pages. ...

How to use weston with DRM backend in Raspbian, and gets benchmark score by GLMark2. With regards to hardware acceleration, We can enable VC4 driver through raspi-config. # update system and raspi-config script $ sudo apt-get update && sudo apt-get upgrade $ sudo apt-get install raspi-config # enable Broadcom VC4 driver $ sudo raspi-config Advanced Options -> GL Driver -> GL (Full KMS) Test environment: 2017-04-10-raspbian-jessie-lite HDMI 1080P 60Hz CPU/GPU use default settings(no overclock) Test packages: ...

以前球隊在回防時, 教練總會說"快跑! 別回頭, 到定點再休息". 關於人生, 喘息點在哪兒? 我想就在滴水穿石之間! 主要在了解linux graphic system架構, 並分析wayland weston compositor/glmark2的repaint flow. Week18 (05/05) 網路文章 和百年公斤原器說再見！七大國際單位制將改頭換面 原來用抽像來定義實體早己行之有年. 網路資源 C++ Language Interface Foundation (CLIF) Google 做了一個C++ wrapper generator, 目前看來只for Python, 目標是支援更多的語言. Week19 (05/12) 網路文章 Weston repaint scheduling Week20 (05/19) Wayland的weston compositor可透過wesgr來分析repaint過程. glmark2-es2-wayland 加入debug去看glmark repaint過程 [18:50:21.959] canvas_clear 1510 us [18:50:21.960] scenc_draw 389 us [18:50:21.960] scenc_update 159 us [18:50:21.960] weston weston_output_repaint[+] [18:50:21.960] weston surface_flush_damage [18:50:21.960] weston surface_flush_damage [18:50:21.961] weston surface_flush_damage [18:50:21.961] weston surface_flush_damage [18:50:21.969] weston weston_output_repaint[-] [18:50:21.970] weston weston_output_finish_frame [18:50:21.972] weston weston_surface_commit_state [18:50:21.972] weston core_repaint_req [18:50:21.972] canvas_update 12193 us [18:50:21.972] glmark2 draw 14443 us 網路文章 Boosting Site Speed Using Brotli Compression LinkedIn 使用Brotli 壓縮提高網頁的瀏覽速度, 並使用JavaScript檔案做了Brotli, Zopfli, and gzip演算法比較, 壓縮後的空間以gzip -6為基準Brotil -11結省30%, Brotli -5結省14%左右, Zopfli -1000結省約5%. 時間上gzip -6 壓縮169ms, 解壓35ms, Brotli -11 壓縮11913ms, 解壓22ms, Brotli壓縮時間花的較長, 但解壓及壓縮率完勝. 最終的A/B Test, 網頁在美國地區快了23.6%, 印度地區快了66.5%. LinkedIn 計劃採用Dynamic Brotli 處理動態資料(Base HTML document, JSON API responses, etc.), 也需要改善gzip以配合不支援Brotli的Browser. Browser support: ...