#5056 implement Yolo2OutputLayer.computeScoreForExamples

deeplearning4j · Jul 25, 2018 · a767452 · a767452
1 parent 90fba44
commit a767452
Show file tree

Hide file tree

Showing 2 changed files with 38 additions and 5 deletions.
diff --git a/...ing4j-core/src/test/java/org/deeplearning4j/nn/layers/objdetect/TestYolo2OutputLayer.java b/...ing4j-core/src/test/java/org/deeplearning4j/nn/layers/objdetect/TestYolo2OutputLayer.java
@@ -89,8 +89,9 @@ public void testYoloActivateScoreBasic() {
 
 
         MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
+                .l2(0.01)
                 .list()
-                .layer(new ConvolutionLayer.Builder().nIn(1).nOut(1).kernelSize(1,1).build())
+                .layer(new ConvolutionLayer.Builder().nIn(depth).nOut(depth).kernelSize(1,1).build())
                 .layer(new Yolo2OutputLayer.Builder()
                         .boundingBoxPriors(bbPrior)
                         .build())
@@ -150,6 +151,16 @@ public void testYoloActivateScoreBasic() {
         double score2 = y2impl.computeScore(0, 0, true, LayerWorkspaceMgr.noWorkspaces());
 
         assertEquals(score, score2, 1e-8);
+
+        //Test computeScoreForExamples:
+        INDArray scoreArr1 = net.scoreExamples(new DataSet(input, labels), false);
+        INDArray scoreArr2 = net.scoreExamples(new DataSet(input, labels), true);
+        assertFalse(scoreArr1.isAttached());
+        assertFalse(scoreArr2.isAttached());
+
+        assertArrayEquals(new long[]{mb,1}, scoreArr1.shape());
+        assertArrayEquals(new long[]{mb,1}, scoreArr2.shape());
+        assertNotEquals(scoreArr1, scoreArr2);
     }
 
 

diff --git a/...plearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/objdetect/Yolo2OutputLayer.java b/...plearning4j-nn/src/main/java/org/deeplearning4j/nn/layers/objdetect/Yolo2OutputLayer.java
@@ -98,12 +98,12 @@ public Yolo2OutputLayer(NeuralNetConfiguration conf) {
 
     @Override
     public Pair<Gradient, INDArray> backpropGradient(INDArray epsilon, LayerWorkspaceMgr workspaceMgr) {
-        INDArray epsOut = computeBackpropGradientAndScore(workspaceMgr, false);
+        INDArray epsOut = computeBackpropGradientAndScore(workspaceMgr, false, false);
 
         return new Pair<>(EMPTY_GRADIENT, epsOut);
     }
 
-    private INDArray computeBackpropGradientAndScore(LayerWorkspaceMgr workspaceMgr, boolean scoreOnly){
+    private INDArray computeBackpropGradientAndScore(LayerWorkspaceMgr workspaceMgr, boolean scoreOnly, boolean computeScoreForExamples){
         assertInputSet(true);
         Preconditions.checkState(labels != null, "Cannot calculate gradients/score: labels are null");
         Preconditions.checkState(labels.rank() == 4, "Expected rank 4 labels array with shape [minibatch, 4+numClasses, h, w]" +
@@ -235,6 +235,26 @@ private INDArray computeBackpropGradientAndScore(LayerWorkspaceMgr workspaceMgr,
         //Calculate the loss:
         ILossFunction lossConfidence = new LossL2();
         IActivation identity = new ActivationIdentity();
+
+
+        if(computeScoreForExamples){
+            INDArray positionLoss = layerConf().getLossPositionScale().computeScoreArray(labelXYCenter2d, predictedXYCenter2d, identity, mask1_ij_obj_2d );
+            INDArray sizeScaleLoss = layerConf().getLossPositionScale().computeScoreArray(labelWHSqrt2d, predictedWHSqrt2d, identity, mask1_ij_obj_2d);
+            INDArray confidenceLossPt1 = lossConfidence.computeScoreArray(labelConfidence2d, predictedConfidence2d, identity, mask1_ij_obj_2d);
+            INDArray confidenceLossPt2 = lossConfidence.computeScoreArray(labelConfidence2d, predictedConfidence2d, identity, mask1_ij_noobj_2d).muli(lambdaNoObj);
+            INDArray classPredictionLoss = layerConf().getLossClassPredictions().computeScoreArray(classLabels2d, classPredictionsPreSoftmax2d, new ActivationSoftmax(), mask1_ij_obj_2d);
+
+            INDArray scoreForExamples = positionLoss.addi(sizeScaleLoss).muli(lambdaCoord)
+                    .addi(confidenceLossPt1).addi(confidenceLossPt2.muli(lambdaNoObj))
+                    .addi(classPredictionLoss)
+                    .dup('c');
+
+            scoreForExamples = scoreForExamples.reshape('c', mb, b*h*w).sum(1).addi(fullNetworkL1 + fullNetworkL2);
+
+            return workspaceMgr.leverageTo(ArrayType.ACTIVATIONS, scoreForExamples);
+        }
+
+
         double positionLoss = layerConf().getLossPositionScale().computeScore(labelXYCenter2d, predictedXYCenter2d, identity, mask1_ij_obj_2d, false );
         double sizeScaleLoss = layerConf().getLossPositionScale().computeScore(labelWHSqrt2d, predictedWHSqrt2d, identity, mask1_ij_obj_2d, false);
         double confidenceLoss = lossConfidence.computeScore(labelConfidence2d, predictedConfidence2d, identity, mask1_ij_obj_2d, false)
@@ -366,7 +386,7 @@ public double computeScore(double fullNetworkL1, double fullNetworkL2, boolean t
         this.fullNetworkL1 = fullNetworkL1;
         this.fullNetworkL2 = fullNetworkL2;
 
-        computeBackpropGradientAndScore(workspaceMgr, true);
+        computeBackpropGradientAndScore(workspaceMgr, true, false);
         return score();
     }
 
@@ -530,7 +550,9 @@ public Pair<Gradient, Double> gradientAndScore() {
 
     @Override
     public INDArray computeScoreForExamples(double fullNetworkL1, double fullNetworkL2, LayerWorkspaceMgr workspaceMgr) {
-        throw new UnsupportedOperationException();
+        this.fullNetworkL1 = fullNetworkL1;
+        this.fullNetworkL2 = fullNetworkL2;
+        return computeBackpropGradientAndScore(workspaceMgr, false, true);
     }
 
     @Override