deil87 / automl-genetic Goto Github PK

We need to make sure that our One-Hot-Encoding works properly and either do something like this

.withColumnReplace("label", $"label" - 1.0 )

for the input data or make it more automatic inside algorithm

For now every time we add model to the stage... we calculate predictions immediately. It maybe not the very optimal way.

https://stackoverflow.com/questions/43330902/spark-off-heap-memory-config-and-tungsten

Check what we should leave in second parameter of FitnessResult(..., here)

As a first version we can use api of some models for that.

How can we improve situation here?

We should probably sample without replacement or tune the probabilities of being sampled.

added ability to log per individual into separate log files instead of writing everything into console.

Store template inside IndividualAlgorithm class with mutable fitnessError property. Print it while traversing.

start from scratch?

This will help us to choose from at least something when we abrupt our evolutions due to timeboxes. Or maybe define strategy: we can store every individual at the beginning and then slowly reduce number.

Changes in AutoMLMainSuite needed

Spark as far as I see has perceptron with only sigmoid activation functions…. 

From the source code:
*Each layer has sigmoid activation function, output layer has softmax.

• Number of inputs has to be equal to the size of feature vectors.
• Number of outputs has to be equal to the total number of labels.

In paper you tried “Training regression models (as components of probabilistic classifiers) is fast and straightforward. We use several activation functions in simple perceptrons, namely Sigmoid, SigmoidNorm, Sine, Polynomial, Gaussian, Exponential and Linear. “ - so I believe we need to use Deeplearning4J implementations for these base classifiers

Or we can fork spark

Possible implementation could be MultiLayerNetwork class from deeplearning4j.com framework

libraryDependencies ++= Seq(
"org.slf4j" % "slf4j-api" % "1.7.7",
"org.slf4j" % "jcl-over-slf4j" % "1.7.7"
).map(_.force())

or update to latest versions.

At least for duplicates of individuals within one population we can cache and reuse the results.

Still need to decide how to make final predictions based on what we have from each Perceptron separately. How to measure confidence?

We should somehow share(not maybe physically) hyperparameters instances for all classifiers within one ensemble individual/generation/evolution. How to find similarities between classifiers in terms of optimal hyperparameters settings?

Think of default/initialization values for them

We can choose test examples that were difficult in terms of classification.

Techniques that have been proposed to ameliorate this difficulty include shared sampling, in
which test cases are chosen so as to be unsolvable by as many of the strategies in the population
as possible

libraryDependencies ++= Seq(
).map(_.force())

or update to the latest versions.

For now we use only Bagging but we can subsample not only training examples but also features space.
Random Forest is not general enough and works only with trees, but we need to apply it's core idea to any ensemble of classifiers.

For now we are getting issue with serialisation of Function2 because there is no Encoder for that type:

No Encoder found for (breeze.linalg.DenseVector[Double], breeze.linalg.DenseVector[Double]) => breeze.linalg.DenseVector[Double]

• field (class: "scala.Function2", name: "_1")
• root class: "scala.Tuple2"
  java.lang.UnsupportedOperationException: No Encoder found for (breeze.linalg.DenseVector[Double], breeze.linalg.DenseVector[Double]) => breeze.linalg.DenseVector[Double]
• field (class: "scala.Function2", name: "_1")
• root class: "scala.Tuple2"

We can define Set[EvolutionDimension] and create pipeline where we configure all those dimensions with strategies and weights.

Later we should search for hyperparameters with evolutions as well.

We can use Kamon or something similar

Maybe we need log4j's Logging for spark. Logback version seems to be nonserializable. Investigation needed(LinearPerceptronClassifier). Maybe upgrade to newer version?

acc.withColumnReplace("maxPrediction", selectMaxColumn($"maxPrediction", $"prediction_$focusClass"))