Very slow training (100h)

What is wrong?

Training takes super long - roughly 100h

Where does it happen?

In the a NeuralNetwork when trying to run a net in node.js on my mac

How do we replicate the issue?

  await getDataSet(type);

  const listings  = await getFilteredListings(type);
  const normalizer = new Normalizer(listings);

  normalizer.setOutputProperties(['runningTime']);
  normalizer.normalize();

  const inputs = normalizer.getBinaryInputDataset();
  const outputs = normalizer.getBinaryOutputDataset();
  const metadata = normalizer.getDatasetMetaData();

  fs.writeFileSync(METADATA_FILE[type],JSON.stringify(metadata));

  const trainingData=[];
  for(let i=0;i<inputs.length;i++){
    trainingData.push({input:inputs[i],output:outputs[i]});
  }
  const options={
    hiddenLayers:[180,60],
  };

  const net = new NeuralNetworkGPU(
    options
  );

  net.train(trainingData,{
    iterations: 60000, // the maximum times to iterate the training data --> number greater than 0
    errorThresh: 0.001, // the acceptable error percentage from training data --> number between 0 and 1
    log:true,
  });

Data has 13 Features and roughly 100MB data (60k items)

How important is this (1-5)?

4

Expected behavior (i.e. solution)

I would expect that this runs faster - even with CPU on Mac. Tried an amazon GPU instance - same. It is very very slow. Am I expecting too much of brain.js? What can this library handle?

[robertleeplummerjr]

Yea, that seems ok. Would there be a way to possibly get more detail about the background of the system? I’d really like to see the implementation and even run a profiler on it to get more insight as to why it is so slow. If you are interested, you can reach out to me via gmail of the same handle. We’ll make sure to post findings here as well.

[mubaidr]

Closing until more information is provided.