无法根据 TFLite 官方示例加载 TFLite 模型

Question

我正在尝试使用 TFLite 创建一个智能回复应用程序，我正在关注来自 github 的 pre-built example。

当从 git 克隆引用的项目并编译时，它可以完美运行。

但是，当我将引用的项目代码（还有 gradle 依赖项、资产、库和其他东西）复制到我的项目时， 它无法加载 tflite 模型并引发运行时错误：

E/AndroidRuntime: FATAL EXCEPTION: main
Process: com.legendsayantan.replai, PID: 14279
java.lang.UnsatisfiedLinkError: No implementation found for long com.legendsayantan.replai.SmartReplyClient.loadJNI(java.nio.MappedByteBuffer, java.lang.String[]) (tried Java_com_legendsayantan_replai_SmartReplyClient_loadJNI and Java_com_legendsayantan_replai_SmartReplyClient_loadJNI__Ljava_nio_MappedByteBuffer_2_3Ljava_lang_String_2)
    at com.legendsayantan.replai.SmartReplyClient.loadJNI(Native Method)
    at com.legendsayantan.replai.SmartReplyClient.loadModel(SmartReplyClient.java:64)
    at com.legendsayantan.replai.MainActivity.lambda$onStart$0(MainActivity.java:90)
    at com.legendsayantan.replai.-$$Lambda$MainActivity$Xdq7R5vPx_buuatNOneWHck6N2o.run(Unknown Source:0)
    at android.os.Handler.handleCallback(Handler.java:888)
    at android.os.Handler.dispatchMessage(Handler.java:100)
    at android.os.Looper.loop(Looper.java:213)
    at android.app.ActivityThread.main(ActivityThread.java:8178)
    at java.lang.reflect.Method.invoke(Native Method)
    at com.android.internal.os.RuntimeInit$MethodAndArgsCaller.run(RuntimeInit.java:513)
    at com.android.internal.os.ZygoteInit.main(ZygoteInit.java:1101)

这是我的 MainActivity.java-

import android.content.Context;
import android.content.SharedPreferences;
import android.os.Bundle;
import android.os.Handler;
import android.util.Log;

import android.view.Menu;


import com.google.android.material.navigation.NavigationView;

import androidx.navigation.NavController;
import androidx.navigation.Navigation;
import androidx.navigation.ui.AppBarConfiguration;
import androidx.navigation.ui.NavigationUI;
import androidx.drawerlayout.widget.DrawerLayout;
import androidx.appcompat.app.AppCompatActivity;
import androidx.appcompat.widget.Toolbar;

import org.tensorflow.lite.Interpreter;

public class MainActivity extends AppCompatActivity {

private AppBarConfiguration mAppBarConfiguration;
public static SharedPreferences sharedPreferences;
public static Context context;

public static final String TAG = "SmartReply";
public static SmartReplyClient client;
public static Handler handler;
public static Interpreter model;

@Override
protected void onCreate(Bundle savedInstanceState) {
    super.onCreate(savedInstanceState);
    context=getApplicationContext();
    setContentView(R.layout.activity_main);

    client = new SmartReplyClient(getApplicationContext());
    handler = new Handler();

    sharedPreferences = getPreferences(Context.MODE_PRIVATE);
    Toolbar toolbar = findViewById(R.id.toolbar);
    setSupportActionBar(toolbar);
    DrawerLayout drawer = findViewById(R.id.drawer_layout);
    NavigationView navigationView = findViewById(R.id.nav_view);
    // Passing each menu ID as a set of Ids because each
    // menu should be considered as top level destinations.
    mAppBarConfiguration = new AppBarConfiguration.Builder(
            R.id.nav_home, R.id.nav_gallery, R.id.nav_slideshow)
            .setDrawerLayout(drawer)
            .build();
    NavController navController = Navigation.findNavController(this, R.id.nav_host_fragment);
    NavigationUI.setupActionBarWithNavController(this, navController, mAppBarConfiguration);
    NavigationUI.setupWithNavController(navigationView, navController);

}

@Override
public boolean onCreateOptionsMenu(Menu menu) {
    // Inflate the menu; this adds items to the action bar if it is present.
    getMenuInflater().inflate(R.menu.main, menu);
    return true;
}

@Override
public boolean onSupportNavigateUp() {
    NavController navController = Navigation.findNavController(this, R.id.nav_host_fragment);
    return NavigationUI.navigateUp(navController, mAppBarConfiguration)
            || super.onSupportNavigateUp();
}

@Override
protected void onStart() {
    super.onStart();
    Log.v(TAG, "onStart");
    handler.post(
            () -> {
               client.loadModel();
            });
}

@Override
protected void onStop() {
    super.onStop();
    Log.v(TAG, "onStop");
    handler.post(
            () -> {
                client.unloadModel();
            });
}

private static void send(final String message) {
    handler.post(
            () -> {
                StringBuilder textToShow = new StringBuilder();
                textToShow.append("Input: ").append(message).append("\n\n");

                // Get suggested replies from the model.
                SmartReply[] ans = client.predict(new String[] {message});
                for (SmartReply reply : ans) {
                    textToShow.append("Reply: ").append(reply.getText()).append("\n");
                }
                textToShow.append("------").append("\n");


            });
      }
}

这里是 SmartReplyClient.java（与参考 github 项目完全相同的文件）：

import android.content.Context;
import android.content.res.AssetFileDescriptor;

import androidx.annotation.Keep;
import androidx.annotation.WorkerThread;

import java.io.BufferedReader;
import java.io.FileInputStream;
import java.io.IOException;
import java.io.InputStreamReader;
import java.nio.MappedByteBuffer;
import java.nio.channels.FileChannel;
import java.util.ArrayList;
import java.util.List;


public class SmartReplyClient implements AutoCloseable {
  private static final String TAG = "SmartReplyDemo";
  private static final String MODEL_PATH = "smartreply.tflite";
  private static final String BACKOFF_PATH = "backoff_response.txt";
  private static final String JNI_LIB = "smartreply_jni";

  private final Context context;
  private long storage;
  private MappedByteBuffer model;

  private volatile boolean isLibraryLoaded;

  public SmartReplyClient(Context context) {
    this.context = context;
  }

  public boolean isLoaded() {
    return storage != 0;
  }

  @WorkerThread
  public synchronized void loadModel() {
    if (!isLibraryLoaded) {
      System.loadLibrary(JNI_LIB);
      isLibraryLoaded = true;
    }

    try {
      model = loadModelFile();
      String[] backoff = loadBackoffList();
      storage = loadJNI(model, backoff); //This line is throwing the error
      // But this same java file works nice in the reference project
    } catch (Exception e) {
      System.out.println(e.getMessage());
      return;
    }
  }

  @WorkerThread
  public synchronized SmartReply[] predict(String[] input) {
    if (storage != 0) {
      return predictJNI(storage, input);
    } else {
      return new SmartReply[] {};
    }
  }

  @WorkerThread
  public synchronized void unloadModel() {
     close();
  }

  @Override
  public synchronized void close() {
    if (storage != 0) {
      unloadJNI(storage);
      storage = 0;
    }
  }

  public MappedByteBuffer loadModelFile() throws IOException {
     try (AssetFileDescriptor fileDescriptor =
         AssetsUtil.getAssetFileDescriptorOrCached(context, MODEL_PATH);
         FileInputStream inputStream = new FileInputStream(fileDescriptor.getFileDescriptor())) {
         FileChannel fileChannel = inputStream.getChannel();
         long startOffset = fileDescriptor.getStartOffset();
         long declaredLength = fileDescriptor.getDeclaredLength();
         return fileChannel.map(FileChannel.MapMode.READ_ONLY, startOffset, declaredLength);
     }
  }

   private String[] loadBackoffList() throws IOException {
    List<String> labelList = new ArrayList<String>();
    try (BufferedReader reader =
        new BufferedReader(new InputStreamReader(context.getAssets().open(BACKOFF_PATH)))) {
      String line;
      while ((line = reader.readLine()) != null) {
        if (!line.isEmpty()) {
          labelList.add(line);
        }
      }
    }
    String[] ans = new String[labelList.size()];
    labelList.toArray(ans);
    return ans;
  }

  @Keep
  private native long loadJNI(MappedByteBuffer buffer, String[] backoff);

  @Keep
  private native SmartReply[] predictJNI(long storage, String[] text);

  @Keep
  private native void unloadJNI(long storage);
}

我也在build.gradle中实现了相同的tensorflow版本作为参考示例：

implementation 'org.tensorflow:tensorflow-lite:0.0.0-nightly-SNAPSHOT'

在这里，它会抛出上面提到的错误。

为了避免错误，我还尝试使用Interpreter 加载 TFLite 模型 在活动onCreate:

Interpreter interpreter;
        try {
            interpreter=new Interpreter(loadmodelfile());
        } catch (IOException e) {
            e.printStackTrace();
        }

方法loadmodelfile()：

public MappedByteBuffer loadmodelfile() throws IOException {
        AssetFileDescriptor assetFileDescriptor = this.getAssets().openFd("smartreply.tflite");
        FileInputStream fileInputStream = new FileInputStream(assetFileDescriptor.getFileDescriptor());
        FileChannel fileChannel = fileInputStream.getChannel();
        long startoff = assetFileDescriptor.getStartOffset();
        long length = assetFileDescriptor.getDeclaredLength();
        return fileChannel.map(FileChannel.MapMode.READ_ONLY,startoff,length);
    }

仍然存在，我得到了这个错误：

E/AndroidRuntime: FATAL EXCEPTION: main
    Process: com.legendsayantan.tflitesmartreplyremake, PID: 10879
    java.lang.RuntimeException: Unable to start activity ComponentInfo{com.legendsayantan.tflitesmartreplyremake/com.legendsayantan.tflitesmartreplyremake.MainActivity}: java.lang.IllegalStateException: Internal error: Unexpected failure when preparing tensor allocations: Encountered unresolved custom op: Normalize.
    Node number 0 (Normalize) failed to prepare.
    
        at android.app.ActivityThread.performLaunchActivity(ActivityThread.java:3782)
        at android.app.ActivityThread.handleLaunchActivity(ActivityThread.java:3961)
        at android.app.servertransaction.LaunchActivityItem.execute(LaunchActivityItem.java:91)
        at android.app.servertransaction.TransactionExecutor.executeCallbacks(TransactionExecutor.java:149)
        at android.app.servertransaction.TransactionExecutor.execute(TransactionExecutor.java:103)
        at android.app.ActivityThread$H.handleMessage(ActivityThread.java:2386)
        at android.os.Handler.dispatchMessage(Handler.java:107)
        at android.os.Looper.loop(Looper.java:213)
        at android.app.ActivityThread.main(ActivityThread.java:8178)
        at java.lang.reflect.Method.invoke(Native Method)
        at com.android.internal.os.RuntimeInit$MethodAndArgsCaller.run(RuntimeInit.java:513)
        at com.android.internal.os.ZygoteInit.main(ZygoteInit.java:1101)
     Caused by: java.lang.IllegalStateException: Internal error: Unexpected failure when preparing tensor allocations: Encountered unresolved custom op: Normalize.
    Node number 0 (Normalize) failed to prepare.
    
        

请让我知道我遗漏的一些东西或我在这些尝试中犯的任何错误。

我是 Tensorflow 的新手，但不是创建 Android 应用程序的新手，我只是想不通我在这里还有什么错误。

任何帮助或建议将不胜感激！

Answer 1

遇到未解决的自定义操作：规范化。

看起来您的模型有一个自定义操作 Normalize，这意味着您需要实现自己的 TFLite 自定义操作并将其注册到 TF​​Lite 解释器。如果您没有实现自定义操作的计划，请考虑使用 Select TF op 选项，该选项可以在移动设备上利用 TensorFlow 操作。

https://www.tensorflow.org/lite/guide/ops_select