djl
djl copied to clipboard
deepjavalibrary
DOnt getting Any BoundingBox AND result same classes always?
hey i m following pikachu detection for creating my own dataset ...Dogs Breed Detection + claasification heres my Data prepearation code: **//OBJECT DETECTION DATASET
public class GestureDataset extends ObjectDetectionDataset {
private static final String VERSION = "1.0";
private static final String ARTIFACT_ID = "gesture";
private Usage usage;
private List<Path> imagePaths;
private PairList<Long, Rectangle> labels;
private Resource resource;
private boolean prepared;
protected GestureDataset(Builder builder) {
super(builder);
usage = builder.usage;
imagePaths = new ArrayList<Path>();
labels = new PairList<Long, Rectangle>();
// MRL mrl = MRL.dataset(CV.ANY, builder.groupId, builder.artifactId);
// resource = new Resource(builder.repository, mrl, VERSION);
}
/**
* Creates a new builder to build a {@link PikachuDetection}.
*
* @return a new builder
*/
public static Builder builder() {
return new Builder();
}
/** {@inheritDoc} */
@Override
public void prepare(Progress progress) throws IOException {
if (prepared) {
return;
}
String images="//content//drive//MyDrive//Colab Notebooks//DOGS";
Path root = Paths.get("//content//drive//MyDrive//Colab Notebooks//DOGS//");
File imgs=new File(images);
Path usagePath;
switch (usage) {
case TRAIN:
usagePath = Paths.get("//DOGS//");
break;
case TEST:
usagePath = Paths.get("//Validated//");
break;
case VALIDATION:
default:
throw new UnsupportedOperationException("Validation data not available.");
}
usagePath = root.resolve(usagePath);
String Annopath="//content//drive//MyDrive//Colab Notebooks//DOGS_ANNO";
File indexFiles = new File(Annopath);
List <File>folds=Arrays.asList(indexFiles.listFiles());
Collections.sort(folds);
long classs=1;
for(File dogs:folds){
try{
System.out.println(dogs.getName());
File dogname=new File(Annopath+"//"+dogs.getName());
for(File images_anno:dogname.listFiles()){
BufferedReader reader = Files.newBufferedReader(Paths.get(Annopath+"//"+dogs.getName()+"//"+images_anno.getName()) );
String lines=reader.lines().map(line -> line + "\n").collect(Collectors.joining());
Type mapType = new TypeToken<Map<String, List<Float>>>() {}.getType();
JSONObject json = XML.toJSONObject(lines);
System.out.println(json);
System.out.println("Wiooooooo:"+images_anno.getName());
GsonBuilder builder = new GsonBuilder();
builder.setPrettyPrinting();
Gson gson = builder.create();
Anno anno=gson.fromJson(json.toString(),Anno.class);
String imgName=anno.getAnnotation().getObject().getName();
System.out.println(imgName+"------"+root.resolve(dogs.getName()+"//"+images_anno.getName()));
imagePaths.add(root.resolve(dogs.getName()+"//"+images_anno.getName()+".jpg"));
long objectClass = classs;
System.out.println("looki");
System.out.println("Londiiii"+anno.getAnnotation().getObject().getBndbox().getXmin());
Rectangle objectLocation =
new Rectangle(
new Point(anno.getAnnotation().getObject().getBndbox().getXmin(), anno.getAnnotation().getObject().getBndbox().getYmin()), anno.getAnnotation().getObject().getBndbox().getXmax(),anno.getAnnotation().getObject().getBndbox().getYmax());
labels.add(objectClass, objectLocation);
System.out.println("AllSet");
}
}
catch(Exception ex){
System.out.println("Error:"+ex.getMessage());
}
classs+=1;
System.out.println("Classes:"+classs+"-Images-"+imagePaths.size()+"--Labels-"+labels.size());
}
System.out.println(labels);
prepared = true;
}
@Override
public PairList<Long, Rectangle> getObjects(long index) {
return labels;
}
/** {@inheritDoc} */
@Override
protected long availableSize() {
return imagePaths.size();
}
@Override
protected Image getImage(long index) throws IOException {
int idx = Math.toIntExact(index);
return ImageFactory.getInstance().fromFile(imagePaths.get(idx));
}
@Override
public Optional<Integer> getImageWidth() {
return Optional.empty();
}
@Override
public Optional<Integer> getImageHeight() {
return Optional.empty();
}
/** A builder for a {@link PikachuDetection}. */
public static final class Builder extends ImageDataset.BaseBuilder<Builder> {
Repository repository;
String groupId;
String artifactId;
Usage usage;
/** Constructs a new builder. */
Builder() {
repository = BasicDatasets.REPOSITORY;
groupId = BasicDatasets.GROUP_ID;
artifactId = ARTIFACT_ID;
usage = Usage.TRAIN;
}
/** {@inheritDoc} */
@Override
public Builder self() {
return this;
}
/**
* Sets the optional usage.
*
* @param usage the usage
* @return this builder
*/
public Builder optUsage(Usage usage) {
this.usage = usage;
return self();
}
/**
* Sets the optional repository.
*
* @param repository the repository
* @return this builder
*/
public Builder optRepository(Repository repository) {
this.repository = repository;
return self();
}
/**
* Sets optional groupId.
*
* @param groupId the groupId}
* @return this builder
*/
public Builder optGroupId(String groupId) {
this.groupId = groupId;
return this;
}
/**
* Sets the optional artifactId.
*
* @param artifactId the artifactId
* @return this builder
*/
public Builder optArtifactId(String artifactId) {
if (artifactId.contains(":")) {
String[] tokens = artifactId.split(":");
groupId = tokens[0];
this.artifactId = tokens[1];
} else {
this.artifactId = artifactId;
}
return this;
}
/**
* Builds the {@link PikachuDetection}.
*
* @return the {@link PikachuDetection}
*/
public GestureDataset build() {
if (pipeline == null) {
pipeline = new Pipeline(new ToTensor());
}
return new GestureDataset(this);
}
}
}**
All the model building is same as picachu detection:
`var usage=Dataset.Usage.TRAIN; var validate=Dataset.Usage.TEST; Pipeline pipeline = new Pipeline(new Resize(150,150)); pipeline.add(new ToTensor()); ObjectDetectionDataset dogsData = GestureDataset.builder() .optUsage(usage) .optLimit(1000)
.optPipeline(pipeline)
.setSampling(128,false)
.build();
dogsData.prepare(new ProgressBar());
dogsData.size();
`
at last my training code:
int[] numFilters = {16, 32, 64};
SequentialBlock baseBlock = new SequentialBlock();
for (int numFilter : numFilters) {
baseBlock.add(SingleShotDetection.getDownSamplingBlock(numFilter));
}
List<List<Float>> sizes = new ArrayList<>();
List<List<Float>> ratios = new ArrayList<>();
for (int i = 0; i < 5; i++) {
ratios.add(Arrays.asList(1f, 2f, 0.5f));
}
sizes.add(Arrays.asList(0.2f, 0.272f));
sizes.add(Arrays.asList(0.37f, 0.447f));
sizes.add(Arrays.asList(0.54f, 0.619f));
sizes.add(Arrays.asList(0.71f, 0.79f));
sizes.add(Arrays.asList(0.88f, 0.961f));
//TRAIN BLOCK
var ssd= SingleShotDetection.builder()
.setNumClasses(12)
.setNumFeatures(3)
.optGlobalPool(true)
.setRatios(ratios)
.setSizes(sizes)
.setBaseNetwork(baseBlock)
.build();
SequentialBlock ssdPredict = new SequentialBlock();
ssdPredict.add(ssd);
ssdPredict.add(new LambdaBlock(
output -> {
NDArray anchors = output.get(0);
System.out.println(anchors);
NDArray classPredictions = output.get(1).softmax(-1).transpose(0, 2, 1);
NDArray boundingBoxPredictions = output.get(2);
MultiBoxDetection multiBoxDetection =
MultiBoxDetection.builder().build();
NDList detections =
multiBoxDetection.detection(
new NDList(
classPredictions,
boundingBoxPredictions,
anchors));
return detections.singletonOrThrow().split(new long[] {1, 2}, 2);
})); //ADDING OUTPUT BLOCK
TRAINING START FROM HERE
` Model model = Model.newInstance("GestureDetector");
float detectionThreshold = 0.2f;
// load parameters back to original training block
model.setBlock(ssd);
// append prediction logic at end of training block with parameter loaded
Block ssdTrain = model.getBlock();
model.setBlock(ssdPredict);
model.setBlock(ssdTrain);
var epoch=50;
var train_config= new DefaultTrainingConfig(new SingleShotDetectionLoss())
.addEvaluator(new SingleShotDetectionAccuracy("classAccuracy"))
.addEvaluator(new BoundingBoxError("boundingBoxError"))
.optDevices(new Device[]{Device.gpu(0)})
.addTrainingListeners(TrainingListener.Defaults.logging());
Trainer trainer = model.newTrainer(train_config);
SingleShotDetectionTranslator translator =SingleShotDetectionTranslator.builder()
.addTransform(new Resize(150,150))
.addTransform(new ToTensor())
.optThreshold(detectionThreshold)
.optSynset(Arrays.asList("Japanese","Shih-tzu","bluetick","English_fox","Irish_wolf","Saluki","Staffordshire","American_terrier","Border_terrier","Brittany_spaniel","Rottweiler","German_shephard"))
.build();
trainer.setMetrics(new Metrics());
System.out.println("whaaat");
trainer.initialize(new Shape(128,3,150, 150));
for(int epochs=0;epochs<epoch;epochs++){
for(Batch b:trainer.iterateDataset(dogsData)){
System.out.println("loo");
EasyTrain.trainBatch(trainer,b);
trainer.step();
}
}
// reset training and validation evaluators at end of epoch
trainer.notifyListeners(listener -> listener.onEpoch(trainer));`
HERES my result :
Model mn=trainer.getModel();
mn.setBlock(ssdPredict);
Predictor<Image, DetectedObjects> predictor = mn.newPredictor(translator);
Image image = ImageFactory.getInstance().fromFile(Paths.get("//content//drive//MyDrive//Colab Notebooks//DOGS//n02088632-bluetick//n02088632_1047.jpg"));
DetectedObjects detectedObjects = predictor.predict(image);
image.drawBoundingBoxes(detectedObjects);
// return number of pikachu detected
detectedObjects.getNumberOfObjects();
Map<String, List<Rectangle>> ret = new ConcurrentHashMap<>();
int imageWidth = image.getWidth();
int imageHeight = image.getHeight();
Map<String,Float> prlist=new ConcurrentHashMap<>();
List<DetectedObjects.DetectedObject> detections = detectedObjects.items();
int n=1;
for (DetectedObjects.DetectedObject detection : detections) {
String className = detection.getClassName();
float probability = (float) detection.getProbability();
Rectangle rect = detection.getBoundingBox().getBounds();
double x = rect.getX() * imageWidth;
double y = rect.getY() * imageHeight;
double w = rect.getWidth() * imageWidth;
double h = rect.getHeight() * imageHeight;
System.out.println(x+"---"+y+"----"+w+"-----"+h);
if(probability>0.99){
prlist.put(className+"_"+n,probability);
n++;
}
ret.compute(
className,
(k, list) -> {
if (list == null) {
list = new ArrayList<>();
}
//list.add(new BoundingBox( x, y, w, h));
return list;
});
}
System.out.println(prlist);
image.getWrappedImage()
no mater which dog image i use results are all same it gives me only this four result array with NO boundingboxes and sorry for too much code....i dont know what to do please suggest me
I see one problem with your getObjects function. The function return value is a list of objects because it is possible for there to be multiple objects in a single image. If you only have one object to detect, you should return a singleton list. However, you are returning the same value regardless of which image.
Consider changing the type of your labels instance variable to List<PairList<Long, Rectangle>>.
i m not getting what you are saying...means you are saying that i should change getObjects return type from pairList to List<PairList> ....and i just want my model to detect one dog only ..single object...can you share a small sample it would make me clear...
Update: when i change labels instance from Pair to List its gives me cant @Override error which is obivous that i m changing return type.....i dont understand...can you please share me A CodePiece for this...??
@Override
public PairList<Long, Rectangle> getObjects(long index) {
return new PairList(Collections.singletonList(labels.get(index));
}
i m getting Null pointer exception after doing this now what should i do?