cyberbeat

The shape for LSTM seems to be wrong? Timestep should be at position 3 but is in position 2 from dagli?

I tried to patch NetworkBuilderLayerVisitor like this: ``` @Override public Void visit(NNLSTMLayer visited) { // the number of units is not always the same as the output width: long[] outputShape...

I fixed serialization of daglis `ReshapeMasklessVertex` by marking _newShape as JsonProperty ``` @JsonProperty("newShape") private final long[] _newShape; ``` May this be a bug to be fixed?

About serializing: I used beta8. I think the problem is, that private fields may not be serialized by default? About aggregation: thanks for your investigation. I would like to add...

These are the logged lines (2 iterations) ` [pool-1-thread-27] INFO org.deeplearning4j.optimize.listeners.TimeIterationListener - Remaining time : 2mn - End expected : Tue Feb 09 21:55:39 CET 2021 [pool-1-thread-27] INFO org.deeplearning4j.optimize.listeners.PerformanceListener -...

See here: https://github.com/linkedin/dagli/blob/13ebe37f13535e706f1f5ae128a4c93b8bbf7150/nn-dl4j/src/main/java/com/linkedin/dagli/dl4j/LossFunctionConverterVisitor.java#L18 where you use "LossMCXENT": https://github.com/eclipse/deeplearning4j/blob/fc735d30023981ebbb0fafa55ea9520ec44292e0/nd4j/nd4j-backends/nd4j-api-parent/nd4j-api/src/main/java/org/nd4j/linalg/lossfunctions/impl/LossMCXENT.java Perhaps you could also add an optional weight-array instead of a single double, you are converting it to an array anyway later...

I now have an exception, when using a weight: ``` Exception in thread "main" java.lang.RuntimeException: MultithreadedDAGExecutor terminated execution because it encountered an unexpected exception in a worker thread: java.lang.IllegalArgumentException: Weights...

Mhm, this seems a dl4j bug - the check for isRowVector should be replaced by isRowVectorOrScalar, right?

This is now fixed in dl4j: https://github.com/eclipse/deeplearning4j/issues/9582#issuecomment-1001112361 Also samediff loss is available: https://deeplearning4j.konduit.ai/samediff/reference/operation-namespaces/loss#weightedcrossentropywithlogits

I also found this: https://community.konduit.ai/t/per-sample-weights-or-label-fractions/706/5 https://github.com/eclipse/deeplearning4j-examples/blob/master/dl4j-examples/src/main/java/org/deeplearning4j/examples/advanced/features/customizingdl4j/lossfunctions/CustomLossUsageEx.java