之前一直是在使用的是tensorflow mobile，來作為模型的執行環境。但是，tensoflow mobile的libtensorflow_inference.so有19MB，load到記憶體裡以後，會佔用較多的記憶體。。。測試看下來，大概佔用20M，加上模型佔用記憶體，導致我們的模組，記憶體佔用變大了30MB。記憶體佔用偏高，因此想優化一下。

首先想到的手段是模型裁剪。

但是，只有裁剪模型的網路結構，才能降低記憶體佔用，可是這樣做，又會導致模型的準確率降低，因此放棄。

剩下一條路，就是抱著試一試的心態，嘗試換用較新的tensorflow lite了。

網上很多教程都是用gradle，或者brzel編譯使用tensorflow lite的，但是我想在Android原始碼裡編譯使用tensorflow lite。因此無法使用gradle或者brzel編譯。查了不少資料，折騰了兩天，終於搞定了。在這裡分享給大家。

編譯出適合手機的jar和so

Build TensorFlow Lite and the demo app from source

Clone the TensorFlow repo

git clone https://github.com/tensorflow/tensorflow

Install Bazel

If bazel is not installed on your system, see Installing Bazel.

Note: Bazel does not currently support Android builds on Windows. Windows users should download the 

prebuilt binary.

Install Android NDK and SDK

The Android NDK is required to build the native (C/C++) TensorFlow Lite code. The current recommended version is 14band can be found on the NDK Archives page.

The Android SDK and build tools can be downloaded separately or used as part of Android Studio

. To build the TensorFlow Lite Android demo, build tools require API >= 23 (but it will run on devices with API >= 21).

In the root of the TensorFlow repository, update the WORKSPACE file with the api_level and location of the SDK and NDK. If you installed it with Android Studio, the SDK path can be found in the SDK manager. The default NDK path is:{SDK path}/ndk-bundle. For example:

android_sdk_repository (
    name = "androidsdk",
    api_level = 23,
    build_tools_version = "23.0.2",
    path = "/home/xxxx/android-sdk-linux/",
)

android_ndk_repository(
    name = "androidndk",
    path = "/home/xxxx/android-ndk-r10e/",
    api_level = 19,
)

Some additional details are available on the TF Lite Android App page.

Build the source code

To build the demo app, run bazel:

bazel build --cxxopt=--std=c++11 //tensorflow/contrib/lite/java/demo/app/src/main:TfLiteCameraDemo

Caution: Because of an bazel bug, we only support building the Android demo app within a Python 2 environment.

就是這個"Some additional details are available on the TF Lite Android App page.",一定要點進去看。。。

Building from Source with Bazel

1. • You'll need at least SDK version 23.
   • Make sure to install the latest version of Bazel. Some distributions ship with Bazel 0.5.4, which is too old.
   • Bazel requires Android Build Tools 26.0.1 or higher.
   • Bazel is incompatible with NDK revisions 15 and above, with revision 16 being a compile-breaking change.Download an older version manually instead of using the SDK Manager.
   • You also need to install the Android Support Repository, available through Android Studio under Android SDK Manager -> SDK Tools -> Android Support Repository.

一定要按照第一步的提示“Follow the Bazel steps for the TF Demo App:”，使用NDK14的版本，才能編譯成功。:P

官網最開始寫19。。。坑啊。。。各種編譯不過。我嘗試了19,17,16。。。都失敗了。。直到最後仔細檢視每個連結。。。仔細閱讀每篇教程。。才終於找到The current recommended version is 14b, which may be found here.

Bazel

NOTE: Bazel does not currently support building for Android on Windows. Full support for gradle/cmake builds is coming soon, but in the meantime we suggest that Windows users download the prebuilt binaries instead.

Install Bazel and Android Prerequisites

Bazel is the primary build system for TensorFlow. To build with Bazel, it and the Android NDK and SDK must be installed on your system.

1. Install the latest version of Bazel as per the instructions on the Bazel website.

2. The Android NDK is required to build the native (C/C++) TensorFlow code. The current recommended version is 14b, which may be found here.

   • NDK 16, the revision released in November 2017, is incompatible with Bazel. See here.

3. The Android SDK and build tools may be obtained here, or alternatively as part of Android Studio. Build tools API >= 23 is required to build the TF Android demo (though it will run on API >= 21 devices).

   • The Android Studio SDK Manager's NDK installer will install the latest revision of the NDK, which is incompatiblewith Bazel. You'll need to download an older version manually, as (2) suggests.

Edit WORKSPACE

NOTE: As long as you have the SDK and NDK installed, the ./configure script will create these rules for you. Answer "Yes" when the script asks to automatically configure the ./WORKSPACE.

The Android entries in <workspace_root>/WORKSPACE must be uncommented with the paths filled in appropriately depending on where you installed the NDK and SDK. Otherwise an error such as: "The external label '//external:android/sdk' is not bound to anything" will be reported.

Also edit the API levels for the SDK in WORKSPACE to the highest level you have installed in your SDK. This must be >= 23 (this is completely independent of the API level of the demo, which is defined in AndroidManifest.xml). The NDK API level may remain at 14.

bazel編譯環境配置好後，就可以編譯了。

官網推薦的這個

bazel build --cxxopt=--std=c++11 //tensorflow/contrib/lite/java/demo/app/src/main:TfLiteCameraDemo

只能編譯出TfLiteCameraDemo的apk，

但是我想編譯出libtensorflowlite.jar和libtensorflowlite_jni.so，方便在Android原始碼環境裡編譯整合tensorflow lite，所以，在編譯tensorflowlite的原始碼時，需要調整一下bazel的編譯命令：

bazel build --cxxopt='--std=c++11' //tensorflow/contrib/lite/java:tensorflowlite --crosstool_top=//external:android/crosstool [email protected]_tools//tools/cpp:toolchain --cpu=arm64-v8a

然後，就能成功編譯出給arm64-v8a使用的libtensorflowlite.jar和libtensorflowlite_jni.so了。

想編譯支援其他型別的機器，修改--cpu引數即可。

Android原始碼裡編譯使用tensorflowlite

下面，就是在原始碼環境裡測試一下。

首先，把TfLiteCameraDemo的原始碼拷貝到Android原始碼目錄裡，然後新增Android.mk,這兩個配置是關鍵：

LOCAL_PREBUILT_STATIC_JAVA_LIBRARIES := tensorflowlite:libs/libtensorflowlite.jar

LOCAL_PREBUILT_LIBS := libtensorflowlite_jni:libs/libtensorflowlite_jni.so

以前使用tensorflow mobile的時候，libtensorflow_inference.so是需要自己手動System.loadLibrary的。

但是我看了下TfLiteCameraDemo，沒有發現System.loadLibrary的程式碼。

於是，就去檢視TensorflowLite的原始碼，發現原來libtensorflowlite_jni.so已經在TensorflowLite里加載了。（所以，我們在Android.mk裡配置的so的名字，也必須是“libtensorflowlite_jni”，否則執行時會報錯，載入so失敗）

/** Static utility methods loading the TensorFlowLite runtime. */
public final class TensorFlowLite {

  private static final String LIBNAME = "tensorflowlite_jni";

  private TensorFlowLite() {}

  /** Returns the version of the underlying TensorFlowLite runtime. */
  public static native String version();

  /**
   * Load the TensorFlowLite runtime C library.
   */
  static boolean init() {
    try {
      System.loadLibrary(LIBNAME);
      return true;
    } catch (UnsatisfiedLinkError e) {
      System.err.println("TensorFlowLite: failed to load native library: " + e.getMessage());
      return false;
    }
  }

  static {
    init();
  }
}

然後在原始碼裡make一下，就成功了。

push進手機後，TfLiteCameraDemo成功識別出了鍵盤，記憶體佔用大概20M,這是UI activity + bitmap + tensorflowlite + mobileNet一共佔用的，而以前tensorflow mobile自己就能佔用20M了。

總結

原來tensoflow mobile的libtensorflow_inference.so有19M，而libtensorflowlite_jni.so只有1.5M，因此至少在載入tensorflow需要的native so這一塊，lite比mobile就節省了不少記憶體。

特別感謝這兩篇文章，為我提供了不少思路：