davisking/dlib

Feature: Android NN API usage

Open

#1,332 建立於 2018年5月26日

在 GitHub 查看
 (1 留言) (0 反應) (0 負責人)C++ (12,570 star) (3,314 fork)batch import
help wanted

描述

Android now has a neural network API that delegates to NPU/DSP/GPU or falls back to CPU, on devices running Android 8.1 or higher. https://developer.android.com/ndk/guides/neuralnetworks/

Ideally there would be a backend in dlib to support this framework as it is likely going to be the most optimised pathway going forwards.

NNAPI is a bit limited in what it can do at this stage. It only does int8, int32 and float32 for types and a few dozen operations.

It was expanded in NNAPI 1.1, adding some more operations [Div/Mul/Pad/Mean/Sub] (which requires Android 9.0).

Example implementation can be found in the Tensorflow Lite library: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/nnapi/NeuralNetworksShim.h

貢獻者指南

技術棧
cppandroid
領域
machine learningmobilebackend
議題類型
feature
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
5
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
over 1 week
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
mostly clear
前置要求
C++Android NDKneural network concepts
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
15
研究方向
The issue requests implementing a backend in dlib for Android's NNAPI, which delegates to NPU/DSP/GPU or CPU on Android 8.1+. The reference implementation is TensorFlow Lite's NeuralNetworksShim.h. To proceed, study dlib's existing backend architecture (e.g., dlib/dnn/cpu dlib.h) and understand how operations like convolution and pooling map to NNAPI operations (int8, int32, float32). NNAPI 1.1 added Div, Mul, Pad, Mean, Sub. The implementation should handle Android version conditional features. No current assignee or PR exists, so first time contributors should discuss design in the issue before coding.