Android 如何在tflite中使用posenet模型的输出
我正在使用来自的posenet的tflite模型。它接受输入1*353*257*3输入图像,并返回4个维度数组1*23*17*17、1*23*17*34、1*23*17*64和1*23*17*1。该模型的输出步幅为16。如何获得输入图像上所有17个姿势点的坐标?我已经试着从out1阵列的热图中打印信心分数,但每个像素的值接近0.00。代码如下:Android 如何在tflite中使用posenet模型的输出,android,tensorflow,computer-vision,tensorflow-lite,pose-estimation,Android,Tensorflow,Computer Vision,Tensorflow Lite,Pose Estimation,我正在使用来自的posenet的tflite模型。它接受输入1*353*257*3输入图像,并返回4个维度数组1*23*17*17、1*23*17*34、1*23*17*64和1*23*17*1。该模型的输出步幅为16。如何获得输入图像上所有17个姿势点的坐标?我已经试着从out1阵列的热图中打印信心分数,但每个像素的值接近0.00。代码如下: public class MainActivity extends AppCompatActivity { private static final i
public class MainActivity extends AppCompatActivity {
private static final int CAMERA_REQUEST = 1888;
private ImageView imageView;
private static final int MY_CAMERA_PERMISSION_CODE = 100;
Interpreter tflite = null;
private String TAG = "rohit";
//private Canvas canvas;
Map<Integer, Object> outputMap = new HashMap<>();
float[][][][] out1 = new float[1][23][17][17];
float[][][][] out2 = new float[1][23][17][34];
float[][][][] out3 = new float[1][23][17][64];
float[][][][] out4 = new float[1][23][17][1];
@Override
protected void onCreate(Bundle savedInstanceState) {
super.onCreate(savedInstanceState);
setContentView(R.layout.activity_main);
String modelFile="multi_person_mobilenet_v1_075_float.tflite";
try {
tflite=new Interpreter(loadModelFile(MainActivity.this,modelFile));
} catch (IOException e) {
e.printStackTrace();
}
final Tensor no = tflite.getInputTensor(0);
Log.d(TAG, "onCreate: Input shape"+ Arrays.toString(no.shape()));
int c = tflite.getOutputTensorCount();
Log.d(TAG, "onCreate: Output Count" +c );
for (int i = 0; i <4 ; i++) {
final Tensor output = tflite.getOutputTensor(i);
Log.d(TAG, "onCreate: Output shape" + Arrays.toString(output.shape()));
}
this.imageView = this.findViewById(R.id.imageView1);
Button photoButton = this.findViewById(R.id.button1);
photoButton.setOnClickListener(new View.OnClickListener() {
@Override
public void onClick(View v) {
if (checkSelfPermission(Manifest.permission.CAMERA)
!= PackageManager.PERMISSION_GRANTED) {
requestPermissions(new String[]{Manifest.permission.CAMERA},
MY_CAMERA_PERMISSION_CODE);
} else {
Intent cameraIntent = new Intent(android.provider.MediaStore.ACTION_IMAGE_CAPTURE);
startActivityForResult(cameraIntent, CAMERA_REQUEST);
}
}
});
}
public void onRequestPermissionsResult(int requestCode, @NonNull String[] permissions, @NonNull int[] grantResults) {
super.onRequestPermissionsResult(requestCode, permissions, grantResults);
if (requestCode == MY_CAMERA_PERMISSION_CODE) {
if (grantResults[0] == PackageManager.PERMISSION_GRANTED) {
Toast.makeText(this, "camera permission granted", Toast.LENGTH_LONG).show();
Intent cameraIntent = new
Intent(android.provider.MediaStore.ACTION_IMAGE_CAPTURE);
startActivityForResult(cameraIntent, CAMERA_REQUEST);
} else {
Toast.makeText(this, "camera permission denied", Toast.LENGTH_LONG).show();
}
}
}
protected void onActivityResult ( int requestCode, int resultCode, Intent data){
if (requestCode == CAMERA_REQUEST && resultCode == Activity.RESULT_OK) {
Bitmap photo = (Bitmap) data.getExtras().get("data");
Log.d(TAG,"bhai:"+photo.getWidth()+":"+photo.getHeight());
//imageView.setImageBitmap(photo);
photo = Bitmap.createScaledBitmap(photo, 353, 257, false);
photo = photo.copy(Bitmap.Config.ARGB_8888,true);
Log.d(TAG, "onActivityResult: Bitmap resized");
int width =photo.getWidth();
int height = photo.getHeight();
float[][][][] result = new float[1][width][height][3];
int[] pixels = new int[width*height];
photo.getPixels(pixels, 0, width, 0, 0, width, height);
int pixelsIndex = 0;
for (int i = 0; i < width; i++)
{
for (int j = 0; j < height; j++)
{
// result[i][j] = pixels[pixelsIndex];
int p = pixels[pixelsIndex];
result[0][i][j][0] = (p >> 16) & 0xff;
result[0][i][j][1] = (p >> 8) & 0xff;
result[0][i][j][2] = p & 0xff;
pixelsIndex++;
}
}
Object [] inputs = {result};
//inputs[0] = inp;
outputMap.put(0, out1);
outputMap.put(1, out2);
outputMap.put(2, out3);
outputMap.put(3, out4);
tflite.runForMultipleInputsOutputs(inputs,outputMap);
out1 = (float[][][][]) outputMap.get(0);
out2 = (float[][][][]) outputMap.get(1);
out3 = (float[][][][]) outputMap.get(2);
out4 = (float[][][][]) outputMap.get(3);
Canvas canvas = new Canvas(photo);
Paint p = new Paint();
p.setColor(Color.RED);
float[][][] scores = new float[out1[0].length][out1[0][0].length][17];
int[][] heatmap_pos = new int[17][2];
for(int i=0;i<17;i++)
{
float max = -1;
for(int j=0;j<out1[0].length;j++)
{
for(int k=0;k<out1[0][0].length;k++)
{
// Log.d("mylog", "onActivityResult: "+out1[0][j][k][i]);
scores[j][k][i] = sigmoid(out1[0][j][k][i]);
if(max<scores[j][k][i])
{
max = scores[j][k][i];
heatmap_pos[i][0] = j;
heatmap_pos[i][1] = k;
}
}
}
// Log.d(TAG, "onActivityResult: "+max+" "+heatmap_pos[i][0]+" "+heatmap_pos[i][1]);
}
for(int i=0;i<17;i++)
{
float max = -1;
for(int j=0;j<out1[0].length;j++)
{
for(int k=0;k<out1[0][0].length;k++)
{
Log.d("mylog", "onActivityResult: "+out1[0][j][k][i]);
scores[j][k][i] = sigmoid(out1[0][j][k][i]);
if(max<scores[j][k][i])
{
max = scores[j][k][i];
heatmap_pos[i][0] = j;
heatmap_pos[i][1] = k;
}
}
}
// Log.d(TAG, "onActivityResult: "+max+" "+heatmap_pos[i][0]+" "+heatmap_pos[i][1]);
}
for(int i=0;i<17;i++)
{
Log.d("heatlog", "onActivityResult: "+heatmap_pos[i][0]+" "+heatmap_pos[i][1]);
}
float[][] offset_vector = new float[17][2];
float[][] keypoint_pos = new float[17][2];
for(int i=0;i<17;i++)
{
offset_vector[i][0] = out2[0][heatmap_pos[i][0]][heatmap_pos[i][1]][i];
offset_vector[i][1] = out2[0][heatmap_pos[i][0]][heatmap_pos[i][1]][i+17];
Log.d("myoff",offset_vector[i][0]+":"+offset_vector[i][1]);
keypoint_pos[i][0] = heatmap_pos[i][0]*16+offset_vector[i][0];
keypoint_pos[i][1] = heatmap_pos[i][1]*16+offset_vector[i][1];
Log.d(TAG, "onActivityResult: "+keypoint_pos[i][0]+" "+keypoint_pos[i][1]);
canvas.drawCircle(keypoint_pos[i][0]+353/2,keypoint_pos[i][1]-257/2,5,p);
}
imageView.setImageBitmap(photo);
}
}
private MappedByteBuffer loadModelFile(Activity activity, String MODEL_FILE) throws IOException {
AssetFileDescriptor fileDescriptor = activity.getAssets().openFd(MODEL_FILE);
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);
}
public float sigmoid(float value) {
float p = (float)(1.0 / (1 + Math.exp(-value)));
return p;
}
}
public类MainActivity扩展了AppCompatActivity{
专用静态最终int摄像机_请求=1888;
私人影像视图;
私有静态最终int MY_CAMERA_PERMISSION_CODE=100;
解释器tflite=null;
私有字符串TAG=“rohit”;
//私人帆布;
Map outputMap=newhashmap();
浮动汇率[][]out1=新浮动汇率[1][23][17][17];
float[][]out2=新的float[1][23][17][34];
float[]out3=新的float[1][23][17][64];
浮动[]输出4=新浮动[1][23][17][1];
@凌驾
创建时受保护的void(Bundle savedInstanceState){
super.onCreate(savedInstanceState);
setContentView(R.layout.activity_main);
String modelFile=“multi\u person\u mobilenet\u v1\u 075\u float.tflite”;
试一试{
tflite=新的解释器(loadModelFile(MainActivity.this,modelFile));
}捕获(IOE异常){
e、 printStackTrace();
}
最终张量no=tflite.GetInputSensor(0);
Log.d(标记“onCreate:InputShape”+Arrays.toString(no.shape()));
int c=tflite.getOutputTensorCount();
Log.d(标记“onCreate:Output Count”+c);
对于(int i=0;i>16)&0xff;
结果[0][i][j][1]=(p>>8)&0xff;
结果[0][i][j][2]=p&0xff;
pixelsIndex++;
}
}
对象[]输入={result};
//输入[0]=inp;
outputMap.put(0,out1);
outputMap.put(1,out2);
outputMap.put(2,out3);
outputMap.put(3,out4);
runForMultipleInputsOutputs(输入,输出映射);
out1=(float[])outputMap.get(0);
out2=(float[])outputMap.get(1);
out3=(float[])outputMap.get(2);
out4=(float[])outputMap.get(3);
画布=新画布(照片);
油漆p=新油漆();
p、 setColor(Color.RED);
浮动[][]分数=新浮动[out1[0]。长度][out1[0][0]。长度][17];
int[][]热图位置=新int[17][2];
对于(inti=0;i我认为这个tflite模型文件有问题。
因此,我尝试使用模型中的权重创建posenet-tflite模型。
可以从tfjs模型下载模型中的所有权重:
然后,您可以生成模型,并按照以下回购协议进行所有前期和后期处理:
生成posenet模型后,可以导出到.pb文件或.tflite文件。
我已经成功地尝试了这个过程,并且posenet模型可以在我的Android应用程序中使用GPU成功运行。我认为这个tflite模型文件有问题。
因此,我尝试使用模型中的权重创建posenet-tflite模型。
可以从tfjs模型下载模型中的所有权重:
然后,您可以生成模型,并按照以下回购协议进行所有前期和后期处理:
生成posenet模型后,可以导出到.pb文件或.tflite文件。
我已经成功地尝试了这个过程,并且posenet模型可以在我的Android应用程序中使用GPU成功运行。谢谢你的努力。现在我不在城里,但我会尽快尝试。下载url不起作用。以下是我尝试的内容:你可以使用中的权重。你能为我们提供你的tflite文件Ying Li吗?@Ramandepsing,tflite file在这里:但是输入和输出大小不同。谢谢你的努力。现在我不在城里,但我会尽快尝试。下载url不起作用。我尝试的是:你可以在中使用权重。你能提供你的tflite文件吗?Ying Li?@Ramandepsing,tflite文件在这里:但是输入和输出大小不同。