在MainActivity.java中使用非量化tflite文件所需的更改
这是为量化模型编写的,我尝试使用非量化模型 在进行上述更改之后,我的代码是在MainActivity.java中使用非量化tflite文件所需的更改,java,android,firebase,tensorflow,firebase-mlkit,Java,Android,Firebase,Tensorflow,Firebase Mlkit,这是为量化模型编写的,我尝试使用非量化模型 在进行上述更改之后,我的代码是 public class MainActivity extends AppCompatActivity implements AdapterView.OnItemSelectedListener { private static final String TAG = "MainActivity"; private Button mRun; private ImageView mImageView;
public class MainActivity extends AppCompatActivity implements AdapterView.OnItemSelectedListener {
private static final String TAG = "MainActivity";
private Button mRun;
private ImageView mImageView;
private Bitmap mSelectedImage;
private GraphicOverlay mGraphicOverlay;
// Max width (portrait mode)
private Integer mImageMaxWidth;
// Max height (portrait mode)
private Integer mImageMaxHeight;
private final String[] mFilePaths =
new String[]{"mountain.jpg", "tennis.jpg","96580.jpg"};
/**
* Name of the model file hosted with Firebase.
*/
private static final String HOSTED_MODEL_NAME = "mobilenet_v1_224_quant";
private static final String LOCAL_MODEL_ASSET = "retrained_graph_mobilenet_1_224.tflite";
/**
* Name of the label file stored in Assets.
*/
private static final String LABEL_PATH = "labels.txt";
/**
* Number of results to show in the UI.
*/
private static final int RESULTS_TO_SHOW = 3;
/**
* Dimensions of inputs.
*/
private static final int DIM_BATCH_SIZE = 1;
private static final int DIM_PIXEL_SIZE = 3;
private static final int DIM_IMG_SIZE_X = 224;
private static final int DIM_IMG_SIZE_Y = 224;
private static final int IMAGE_MEAN = 128;
private static final float IMAGE_STD = 128.0f;
/**
* Labels corresponding to the output of the vision model.
*/
private List<String> mLabelList;
private final PriorityQueue<Map.Entry<String, Float>> sortedLabels =
new PriorityQueue<>(
RESULTS_TO_SHOW,
new Comparator<Map.Entry<String, Float>>() {
@Override
public int compare(Map.Entry<String, Float> o1, Map.Entry<String, Float>
o2) {
return (o1.getValue()).compareTo(o2.getValue());
}
});
/* Preallocated buffers for storing image data. */
private final int[] intValues = new int[DIM_IMG_SIZE_X * DIM_IMG_SIZE_Y];
/**
* An instance of the driver class to run model inference with Firebase.
*/
private FirebaseModelInterpreter mInterpreter;
/**
* Data configuration of input & output data of model.
*/
private FirebaseModelInputOutputOptions mDataOptions;
@Override
protected void onCreate(Bundle savedInstanceState) {
super.onCreate(savedInstanceState);
setContentView(R.layout.activity_main);
mGraphicOverlay = findViewById(R.id.graphic_overlay);
mImageView = findViewById(R.id.image_view);
Spinner dropdown = findViewById(R.id.spinner);
List<String> items = new ArrayList<>();
for (int i = 0; i < mFilePaths.length; i++) {
items.add("Image " + (i + 1));
}
ArrayAdapter<String> adapter = new ArrayAdapter<>(this, android.R.layout
.simple_spinner_dropdown_item, items);
dropdown.setAdapter(adapter);
dropdown.setOnItemSelectedListener(this);
mLabelList = loadLabelList(this);
mRun = findViewById(R.id.button_run);
mRun.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
runModelInference();
}
});
int[] inputDims = {DIM_BATCH_SIZE, DIM_IMG_SIZE_X, DIM_IMG_SIZE_Y, DIM_PIXEL_SIZE};
int[] outputDims = {DIM_BATCH_SIZE, mLabelList.size()};
try {
mDataOptions =
new FirebaseModelInputOutputOptions.Builder()
.setInputFormat(0, FirebaseModelDataType.FLOAT32, inputDims)
.setOutputFormat(0, FirebaseModelDataType.FLOAT32, outputDims)
.build();
FirebaseModelDownloadConditions conditions = new FirebaseModelDownloadConditions
.Builder()
.requireWifi()
.build();
FirebaseLocalModelSource localModelSource =
new FirebaseLocalModelSource.Builder("asset")
.setAssetFilePath(LOCAL_MODEL_ASSET).build();
FirebaseCloudModelSource cloudSource = new FirebaseCloudModelSource.Builder
(HOSTED_MODEL_NAME)
.enableModelUpdates(true)
.setInitialDownloadConditions(conditions)
.setUpdatesDownloadConditions(conditions) // You could also specify
// different conditions
// for updates
.build();
FirebaseModelManager manager = FirebaseModelManager.getInstance();
manager.registerLocalModelSource(localModelSource);
manager.registerCloudModelSource(cloudSource);
FirebaseModelOptions modelOptions =
new FirebaseModelOptions.Builder()
.setCloudModelName(HOSTED_MODEL_NAME)
.setLocalModelName("asset")
.build();
mInterpreter = FirebaseModelInterpreter.getInstance(modelOptions);
} catch (FirebaseMLException e) {
showToast("Error while setting up the model");
e.printStackTrace();
}
}
private void runModelInference() {
if (mInterpreter == null) {
Log.e(TAG, "Image classifier has not been initialized; Skipped.");
return;
}
// Create input data.
ByteBuffer imgData = convertBitmapToByteBuffer(mSelectedImage, mSelectedImage.getWidth(),
mSelectedImage.getHeight());
try {
FirebaseModelInputs inputs = new FirebaseModelInputs.Builder().add(imgData).build();
// Here's where the magic happens!!
mInterpreter
.run(inputs, mDataOptions)
.addOnFailureListener(new OnFailureListener() {
@Override
public void onFailure(@NonNull Exception e) {
e.printStackTrace();
showToast("Error running model inference");
}
})
.continueWith(
new Continuation<FirebaseModelOutputs, List<String>>() {
@Override
public List<String> then(Task<FirebaseModelOutputs> task) {
float[][] labelProbArray = task.getResult()
.<float[][]>getOutput(0);
List<String> topLabels = getTopLabels(labelProbArray);
mGraphicOverlay.clear();
GraphicOverlay.Graphic labelGraphic = new LabelGraphic
(mGraphicOverlay, topLabels);
mGraphicOverlay.add(labelGraphic);
return topLabels;
}
});
} catch (FirebaseMLException e) {
e.printStackTrace();
showToast("Error running model inference");
}
}
/**
* Gets the top labels in the results.
*/
private synchronized List<String> getTopLabels(float[][] labelProbArray) {
for (int i = 0; i < mLabelList.size(); ++i) {
sortedLabels.add(
new AbstractMap.SimpleEntry<>(mLabelList.get(i), (labelProbArray[0][i] )));
if (sortedLabels.size() > RESULTS_TO_SHOW) {
sortedLabels.poll();
}
}
List<String> result = new ArrayList<>();
final int size = sortedLabels.size();
for (int i = 0; i < size; ++i) {
Map.Entry<String, Float> label = sortedLabels.poll();
result.add(label.getKey() + ":" + label.getValue());
}
Log.d(TAG, "labels: " + result.toString());
return result;
}
/**
* Reads label list from Assets.
*/
private List<String> loadLabelList(Activity activity) {
List<String> labelList = new ArrayList<>();
try (BufferedReader reader =
new BufferedReader(new InputStreamReader(activity.getAssets().open
(LABEL_PATH)))) {
String line;
while ((line = reader.readLine()) != null) {
labelList.add(line);
}
} catch (IOException e) {
Log.e(TAG, "Failed to read label list.", e);
}
return labelList;
}
/**
* Writes Image data into a {@code ByteBuffer}.
*/
private synchronized ByteBuffer convertBitmapToByteBuffer(
Bitmap bitmap, int width, int height) {
ByteBuffer imgData =
ByteBuffer.allocateDirect(
4*DIM_BATCH_SIZE * DIM_IMG_SIZE_X * DIM_IMG_SIZE_Y * DIM_PIXEL_SIZE);
imgData.order(ByteOrder.nativeOrder());
Bitmap scaledBitmap = Bitmap.createScaledBitmap(bitmap, DIM_IMG_SIZE_X, DIM_IMG_SIZE_Y,
true);
imgData.rewind();
scaledBitmap.getPixels(intValues, 0, scaledBitmap.getWidth(), 0, 0,
scaledBitmap.getWidth(), scaledBitmap.getHeight());
// Convert the image to int points.
int pixel = 0;
for (int i = 0; i < DIM_IMG_SIZE_X; ++i) {
for (int j = 0; j < DIM_IMG_SIZE_Y; ++j) {
final int val = intValues[pixel++];
imgData.putFloat((((val >> 16) & 0xFF)-IMAGE_MEAN)/IMAGE_STD);
imgData.putFloat((((val >> 8) & 0xFF)-IMAGE_MEAN)/IMAGE_STD);
imgData.putFloat(((val & 0xFF)-IMAGE_MEAN)/IMAGE_STD);
}
}
return imgData;
}
private void showToast(String message) {
Toast.makeText(getApplicationContext(), message, Toast.LENGTH_SHORT).show();
}
public void onItemSelected(AdapterView<?> parent, View v, int position, long id) {
mGraphicOverlay.clear();
mSelectedImage = getBitmapFromAsset(this, mFilePaths[position]);
if (mSelectedImage != null) {
// Get the dimensions of the View
Pair<Integer, Integer> targetedSize = getTargetedWidthHeight();
int targetWidth = targetedSize.first;
int maxHeight = targetedSize.second;
// Determine how much to scale down the image
float scaleFactor =
Math.max(
(float) mSelectedImage.getWidth() / (float) targetWidth,
(float) mSelectedImage.getHeight() / (float) maxHeight);
Bitmap resizedBitmap =
Bitmap.createScaledBitmap(
mSelectedImage,
(int) (mSelectedImage.getWidth() / scaleFactor),
(int) (mSelectedImage.getHeight() / scaleFactor),
true);
mImageView.setImageBitmap(resizedBitmap);
mSelectedImage = resizedBitmap;
}
}
@Override
public void onNothingSelected(AdapterView<?> parent) {
// Do nothing
}
// Utility functions for loading and resizing images from app asset folder.
public static Bitmap getBitmapFromAsset(Context context, String filePath) {
AssetManager assetManager = context.getAssets();
InputStream is;
Bitmap bitmap = null;
try {
is = assetManager.open(filePath);
bitmap = BitmapFactory.decodeStream(is);
} catch (IOException e) {
e.printStackTrace();
}
return bitmap;
}
// Returns max image width, always for portrait mode. Caller needs to swap width / height for
// landscape mode.
private Integer getImageMaxWidth() {
if (mImageMaxWidth == null) {
// Calculate the max width in portrait mode. This is done lazily since we need to
// wait for a UI layout pass to get the right values. So delay it to first time image
// rendering time.
mImageMaxWidth = mImageView.getWidth();
}
return mImageMaxWidth;
}
// Returns max image height, always for portrait mode. Caller needs to swap width / height for
// landscape mode.
private Integer getImageMaxHeight() {
if (mImageMaxHeight == null) {
// Calculate the max width in portrait mode. This is done lazily since we need to
// wait for a UI layout pass to get the right values. So delay it to first time image
// rendering time.
mImageMaxHeight =
mImageView.getHeight();
}
return mImageMaxHeight;
}
// Gets the targeted width / height.
private Pair<Integer, Integer> getTargetedWidthHeight() {
int targetWidth;
int targetHeight;
int maxWidthForPortraitMode = getImageMaxWidth();
int maxHeightForPortraitMode = getImageMaxHeight();
targetWidth = maxWidthForPortraitMode;
targetHeight = maxHeightForPortraitMode;
return new Pair<>(targetWidth, targetHeight);
}
}
这是在代码实验室中分配缓冲区的方式:
ByteBuffer imgData = ByteBuffer.allocateDirect(
DIM_BATCH_SIZE * DIM_IMG_SIZE_X * DIM_IMG_SIZE_Y * DIM_PIXEL_SIZE);
DIM_BATCH_SIZE-典型用法是支持批处理(如果模型支持)。在我们的示例和您的测试中,您一次输入一个图像,并将其保持为1。
DIM_PIXEL_SIZE-我们在代码实验室中设置了3,对应于r/g/b每个1字节
但是,看起来您使用的是浮动模型。然后,使用浮点(4个字节)来表示r/g/b,而不是r/g/b的每个字节(您自己已经计算出了这一部分)。然后,使用上述代码分配的缓冲区不再足够
对于浮动模型,您可以遵循以下示例:
准确地说,imgData人口的分配公式如下:
ByteBuffer imgData = ByteBuffer.allocateDirect(
DIM_BATCH_SIZE * getImageSizeX() * getImageSizeY() * DIM_PIXEL_SIZE
* getNumBytesPerChannel());
在您的情况下,getNumBytesPerChannel()应该是4
[针对以下错误更新新问题]: 无法获取输入维度。第0个输入应有268203个字节,但找到1072812个字节 这是检查模型期望的字节数==传入的字节数。268203 = 299 * 299 * 3 & 1072812 = 4 * 299 * 299 * 3. 看起来您使用的是量化模型,但为其提供了浮点模型的数据。你能再检查一下你用的型号吗?为了简单起见,不要指定云模型源,而只使用来自资产的本地模型
[更新0628,开发者说他们训练了一个浮动模型]: 可能是你的模型错了;也可能是您下载了一个云模型,它覆盖了您的本地模型。但是错误消息告诉我们正在加载的模型不是浮点模型 为了隔离问题,我建议进行以下几项测试: 1) 从快速入门应用程序中删除setCloudModelName/registerCloudModelSource 2) 使用时,您必须下载注释中提到的模型,并将Camera2BasicFragment更改为使用该ImageClassifierFloatInception(而不是ImageClassifierQuantizedMobileNet)
3) 仍然使用相同的TFLite示例应用程序,切换到您自己的训练有素的模型。确保根据您的值调整图像大小。imgData是如何定义的?这可能就是问题所在。还有一个旁白:在你的问题中,你能不能只写两个单独的代码片段来说明这些变化,而不是把它们混在一起?@PannagSanketi要了解
imgData
是如何定义的,请参见MainActivity.java
。我把这个问题说得更清楚了。添加两个单独的代码会增加问题的大小。谢谢。@duplex143就像上面PannagSanketi建议的那样,如果你把问题中的所有代码都粘贴到这里,我们就更容易评论了。而且,在将来,即使您链接到的代码发生了更改,其他开发人员仍然可以知道确切的问题并可能找到答案。谢谢@duplex143,我给你的imgData人口逻辑贴了一个解释。但如果仍然遇到问题,请粘贴代码片段e2e,包括如何配置FirebaseModelInputOutputOptions。@isabella chen还有,作为另一个问题,我已将资产中的.tflite
和.txt
文件替换为在另一个数据集上训练过的文件。我还将自定义模型上传到了云端。当我试图从StillImageActivity
预测图像时,为什么我仍然得到ImageNet
标签?它不使用云计算的模型吗?即使在从资产中删除ImageNet
tflite文件之后,为什么还要进行ImageNet
标签预测<不过,LivePreviewActivity
中的code>Classification提供了我的自定义标签。我仍然收到一些错误。我已经更新了问题,我很困惑。我已经根据中的脚本在.pb
中生成了模型文件。我已使用中的脚本将其转换为.lite
格式。我试图在这里使用一个非量化模型(inception/nasnet)。请参见toco命令中的此标志--input_data_type=FLOAT
,感谢您的响应。我会让你知道我的测试结果。我已经尝试使用模型文件,并从我的云项目中删除了模型。即使这样,我也得到了V3的268203
,1072812
错误。你看过我对MainActivity.java所做的更改了吗?我认为还有一些需要更改的地方。如果可能的话,下载android和模型文件,对我制作的MainActivity.java进行更改,不要将任何模型上传到Firebase中的project,看看它是否工作。
ByteBuffer imgData = ByteBuffer.allocateDirect(
DIM_BATCH_SIZE * getImageSizeX() * getImageSizeY() * DIM_PIXEL_SIZE
* getNumBytesPerChannel());