I’m using MacOS, Java13, DL4J M2.1 with the following dl4j configurations. I’m working on a project regarding value function approximation. In revenueMaximizationAdv, I have saved the feature vectors of samples in an INDArray [ '], with each sample’s feature vector being an INDArray of (1,n_features). The INDArray [ '] was iteratively assigned an INDArray of (1,n_features) from an attribute of the agent class. I did this because using Nd4j.zeros or Nd4j.vstack seems to be memory-demanding and time-consuming for my computer. In this same class, I tried to access a static method from class OSV that runs ParallelInference.output(INDArray… input). My Parallel Inference Configuration is
new ParallelInference.Builder(model).inferenceMode(InferenceMode.BATCHED).batchLimit(32).workers(4).build()
and hope that I’m able to get the output in form of INDArray back with each INDArray having shape (1,1). However, the following error arises. After searching for answers, I still was unable to sort out the reason.
Exception in thread "main" java.lang.NullPointerException at OSV.eval(OSV.java:122) at revenueMaximizationAdv.doMatching(revenueMaximizationAdv.java:143) at Simulator.sim(Simulator.java:414) at mainSim.main(mainSim.java:78)
I’ve solved my previous exception error. Turns out it’s because the ParallelInference is not initialized before calling OSV.eval(). However, a new error (or more like confusion regarding usage of method) arises. As I mentioned earlier, the INDArray[ '] s_a that I have stores 864 (1,n_features) INDArrays, which is also the input to the pi.ouput() method. From the javadoc, it is supposed to return a INDArray [ '], and as per my understanding, one that is of the same size and have corresponding order to the input array. However, by running INDArray[] VF_a = OSV.eval(s_a);(Please also see the static eval method attached below), the VF_a is only of length 1, and not 864 as expected. There has been limited documentation online regarding how to use ParallelInference.output(INDArray input), so I hope you can advise on how to address this issue.
public class OSV{
private static MultiLayerNetwork model;
private static ParallelInference pi;
public static INDArray[] eval(INDArray[] state){return pi.output(state);}
public void update(){
//sample and generate a DataSetIterator from experience replay
model.fit(DataSetIterator);
pi.updatedModel(model);
}
}
I’ve tried to comment out the last pi line and use model.output() in OSV.eval instead
Do we have any updates on my latest question entry?
@timmy1010697 apologies. There’s not much to document. It’s meant to mirror the networks (both MLN and ComputationGraph)
This should be a mismatch. I would recommend running without ParalellInference to see if anything happens. PI just delegates to a threadlocal version of a copy of your network you pass in. It doesn’t do anything special.