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

AI 人工智慧學校分享

blue chen
April 20, 2018
360

AI 人工智慧學校分享

blue chen

April 20, 2018
Tweet

Transcript

  1. NN Layer Output Input 算法上的幾個關鍵 轉換 轉換 1. FFT (audio)

    2. MFCC (audio) 3. RGB resize (vision) 4. Object detection (vision) 5. Face detection (vision) 1. IFFT (audio) 2. IMFCC (audio) 1. Model Size 2. Memory size 3. int / float (HW 加速)
  2. NN Layer Output Input 軟體 Support Effort 很龐⼤大 (先不論 AI

    model 設計) 轉換 轉換 1.OS driver porting 2. 轉換器算法 3. openCV/ open VX 4. Dlib 5. FFmpeg (streaming) Coretex A - Android NN Coretex M - CMSIS NN 1. IFFT, IMFCC… 2. Trigger Application 極度不成熟 缺少比較新的 layer support Porting 底層⼈人才稀少 後續應⽤用層 ⼜又是另外⼀一層 Know how
  3. Training Deployment RelaJet Zoo (各種模型) RelaJet 團隊開發架構 (視覺為例例) Caffe OpenCV

    Linux OS Win/Mac Android OS IOS Tensorflow Theano Keras Caffe2 Caffe Pytorch … etc Base on Cortex A / X86 powerful platform OpenVX Base on CortexM 單⼼心片架構系統 語⾔言層 C / C++ C / C++ python/Js/ .. etc C / C++
  4. We provide 1. Vision tool 2. Tensorflow / Caffe /

    Caffe2 / Keras model transfer tool 3. Capture Image and video tool 4. Training model tool RelaJet Omnism 全知 platform USD 79