Brutex
/
cavis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

#8409 Fix compgraph backprop issue with dual embedding layers from single input (#52) 

Signed-off-by: AlexDBlack <blacka101@gmail.com>
master
Alex Black 2019-11-16 23:09:41 +11:00 committed by GitHub
parent 09a827fb6d
commit a856922fe9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 25 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
deeplearning4j/deeplearning4j-core/src/test/java/org/deeplearning4j/nn/graph/TestComputationGraphNetwork.java
View File

@ -2143,4 +2143,23 @@ public class TestComputationGraphNetwork extends BaseDL4JTest {
INDArray in = Nd4j.create(DataType.FLOAT, 1, 3, 16, 16, 16);
INDArray out = cg.outputSingle(in);
}
@Test
public void testDualEmbedding(){
ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder()
.graphBuilder()
.addInputs("in")
.addLayer("e1", new EmbeddingLayer.Builder().nIn(10).nOut(5).build(), "in")
.addLayer("e2", new EmbeddingLayer.Builder().nIn(10).nOut(5).build(), "in")
.addLayer("out", new OutputLayer.Builder().nIn(10).nOut(2).activation(Activation.SOFTMAX).lossFunction(LossFunctions.LossFunction.MCXENT).build(), "e1", "e2")
.setOutputs("out")
.build();
ComputationGraph cg = new ComputationGraph(conf);
cg.init();
INDArray in = Nd4j.createFromArray(3).reshape(1, 1);
INDArray label = Nd4j.createFromArray(1, 0).reshape(1, 2);
cg.fit(new DataSet(in, label));
}
}

5
deeplearning4j/deeplearning4j-nn/src/main/java/org/deeplearning4j/nn/graph/ComputationGraph.java
View File

@ -2734,7 +2734,12 @@ public class ComputationGraph implements Serializable, Model, NeuralNetwork {
if (setVertexEpsilon[gv.getVertexIndex()]) {
//This vertex: must output to multiple vertices... we want to add the epsilons here
INDArray currentEps = gv.getEpsilon();
if(currentEps == null){
//Edge case: this can be null for dual embedding layer case - in -> e1, in -> e2
gv.setEpsilon(currentEps);
} else {
gv.setEpsilon(currentEps.addi(epsilons[j++])); //TODO is this always safe?
}
} else {
gv.setEpsilon(epsilons[j++]);
}