source: branches/3040_VectorBasedGP/HeuristicLab.Problems.DataAnalysis.Symbolic.Regression/3.4/SingleObjective/Evaluators/TensorFlowConstantOptimizationEvaluator.cs @ 18240

#region License Information
/* HeuristicLab
 * Copyright (C) Heuristic and Evolutionary Algorithms Laboratory (HEAL)
 *
 * This file is part of HeuristicLab.
 *
 * HeuristicLab is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * HeuristicLab is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with HeuristicLab. If not, see <http://www.gnu.org/licenses/>.
 */
#endregion

//#define EXPORT_GRAPH
//#define LOG_CONSOLE
//#define LOG_FILE

using System;
using System.Collections;
using System.Collections.Generic;
#if LOG_CONSOLE
using System.Diagnostics;
#endif
#if LOG_FILE
using System.Globalization;
using System.IO;
#endif
using System.Linq;
using System.Threading;
using HeuristicLab.Common;
using HeuristicLab.Core;
using HeuristicLab.Data;
using HeuristicLab.Encodings.SymbolicExpressionTreeEncoding;
using HeuristicLab.Parameters;
using HEAL.Attic;
using Tensorflow;
using Tensorflow.NumPy;
using static Tensorflow.Binding;
using static Tensorflow.KerasApi;
using DoubleVector = MathNet.Numerics.LinearAlgebra.Vector<double>;

namespace HeuristicLab.Problems.DataAnalysis.Symbolic.Regression {
  [StorableType("63944BF6-62E5-4BE4-974C-D30AD8770F99")]
  [Item("TensorFlowConstantOptimizationEvaluator", "")]
  public class TensorFlowConstantOptimizationEvaluator : SymbolicRegressionConstantOptimizationEvaluator {
    private const string MaximumIterationsName = "MaximumIterations";
    private const string LearningRateName = "LearningRate";

    //private static readonly TF_DataType DataType = tf.float64;
    //private static readonly TF_DataType DataType = tf.float32;

    #region Parameter Properties
    public IFixedValueParameter<IntValue> ConstantOptimizationIterationsParameter {
      get { return (IFixedValueParameter<IntValue>)Parameters[MaximumIterationsName]; }
    }
    public IFixedValueParameter<DoubleValue> LearningRateParameter {
      get { return (IFixedValueParameter<DoubleValue>)Parameters[LearningRateName]; }
    }
    #endregion

    #region Properties
    public int ConstantOptimizationIterations {
      get { return ConstantOptimizationIterationsParameter.Value.Value; }
    }
    public double LearningRate {
      get { return LearningRateParameter.Value.Value; }
    }
    #endregion

    public TensorFlowConstantOptimizationEvaluator()
      : base() {
      Parameters.Add(new FixedValueParameter<IntValue>(MaximumIterationsName, "Determines how many iterations should be calculated while optimizing the constant of a symbolic expression tree(0 indicates other or default stopping criterion).", new IntValue(10)));
      Parameters.Add(new FixedValueParameter<DoubleValue>(LearningRateName, "", new DoubleValue(0.001)));
    }

    protected TensorFlowConstantOptimizationEvaluator(TensorFlowConstantOptimizationEvaluator original, Cloner cloner)
      : base(original, cloner) { }

    public override IDeepCloneable Clone(Cloner cloner) {
      return new TensorFlowConstantOptimizationEvaluator(this, cloner);
    }

    [StorableConstructor]
    protected TensorFlowConstantOptimizationEvaluator(StorableConstructorFlag _) : base(_) { }

    protected override ISymbolicExpressionTree OptimizeConstants(
      ISymbolicExpressionTree tree, IRegressionProblemData problemData, IEnumerable<int> rows,
      CancellationToken cancellationToken = default(CancellationToken), EvaluationsCounter counter = null) {
      return OptimizeTree(tree,
        problemData, rows,
        ApplyLinearScalingParameter.ActualValue.Value, UpdateVariableWeights,
        ConstantOptimizationIterations, LearningRate,
        cancellationToken);
    }

    public static ISymbolicExpressionTree OptimizeTree(ISymbolicExpressionTree tree,
      IRegressionProblemData problemData, IEnumerable<int> rows,
      bool applyLinearScaling, bool updateVariableWeights, int maxIterations, double learningRate,
      CancellationToken cancellationToken = default(CancellationToken), IProgress<double> progress = null) {

      const bool eager = true;

#if LOG_FILE
      var directoryName = $"C:\\temp\\TFboard\\logdir\\TF_{DateTime.Now.ToString("yyyyMMddHHmmss")}_{maxIterations}_{learningRate.ToString(CultureInfo.InvariantCulture)}";
      Directory.CreateDirectory(directoryName);
      using var predictionTargetLossWriter = new StreamWriter(File.Create(Path.Combine(directoryName, "PredictionTargetLos.csv")));
      using var weightsWriter = new StreamWriter(File.Create(Path.Combine(directoryName, "Weights.csv")));
      using var treeGradsWriter = new StreamWriter(File.Create(Path.Combine(directoryName, "TreeGrads.csv")));
      using var lossGradsWriter = new StreamWriter(File.Create(Path.Combine(directoryName, "LossGrads.csv")));

      predictionTargetLossWriter.WriteLine(string.Join(";", "Prediction", "Target", "Loss"));
      weightsWriter.WriteLine(string.Join(";", Enumerable.Range(0, 4).Select(i => $"w_{i}")));
      treeGradsWriter.WriteLine(string.Join(";", Enumerable.Range(0, 4).Select(i => $"Tg_{i}")));
      lossGradsWriter.WriteLine(string.Join(";", Enumerable.Range(0, 4).Select(i => $"Lg_{i}")));
#endif

      //foreach (var row in rows) {

      bool prepared = TreeToTensorConverter.TryPrepareTree(
        tree,
        problemData, rows.ToList(),
        //problemData, new List<int>(){ row },
        updateVariableWeights, applyLinearScaling,
        eager,
        out Dictionary<string, Tensor> inputFeatures, out Tensor target,
        out Dictionary<ISymbolicExpressionTreeNode, ResourceVariable[]> variables);
      if (!prepared)
        return (ISymbolicExpressionTree)tree.Clone();

      var optimizer = keras.optimizers.Adam((float)learningRate);

      for (int i = 0; i < maxIterations; i++) {
        if (cancellationToken.IsCancellationRequested) break;

#if LOG_FILE || LOG_CONSOLE
        using var tape = tf.GradientTape(persistent: true);
#else
        using var tape = tf.GradientTape(persistent: false);
#endif

        bool success = TreeToTensorConverter.TryEvaluate(
          tree,
          inputFeatures, variables,
          updateVariableWeights, applyLinearScaling,
          eager,
          out Tensor prediction);
        if (!success)
          return (ISymbolicExpressionTree)tree.Clone();

        var loss = tf.reduce_mean(tf.square(target - prediction));

        progress?.Report(loss.ToArray<float>()[0]);
        
        var variablesList = variables.Values.SelectMany(x => x).ToList();
        var gradients = tape.gradient(loss, variablesList);

#if LOG_FILE
        predictionTargetLossWriter.WriteLine(string.Join(";", new[] { prediction.ToArray<float>()[0], target.ToArray<float>()[0], loss.ToArray<float>()[0] }));
        weightsWriter.WriteLine(string.Join(";", variablesList.Select(v => v.numpy().ToArray<float>()[0])));
        treeGradsWriter.WriteLine(string.Join(";", tape.gradient(prediction, variablesList).Select(t => t.ToArray<float>()[0])));
        lossGradsWriter.WriteLine(string.Join(";", tape.gradient(loss, variablesList).Select(t => t.ToArray<float>()[0])));
#endif


        //break;

        optimizer.apply_gradients(zip(gradients, variablesList));
      }
      //}

      var cloner = new Cloner();
      var newTree = cloner.Clone(tree);
      var newConstants = variables.ToDictionary(
        kvp => (ISymbolicExpressionTreeNode)cloner.GetClone(kvp.Key),
        kvp => kvp.Value.Select(x => (double)(x.numpy().ToArray<float>()[0])).ToArray()
      );
      UpdateConstants(newTree, newConstants);




      //var numRows = rows.Count();
      
      //var variablesFeed = new Hashtable();
      //foreach (var kvp in inputFeatures) {
      //  var variableName = kvp.Key;
      //  var variablePlaceholder = kvp.Value;
      //  if (problemData.Dataset.VariableHasType<double>(variableName)) {
      //    var data = problemData.Dataset.GetDoubleValues(variableName, rows).Select(x => (float)x).ToArray();
      //    variablesFeed.Add(variablePlaceholder, np.array(data).reshape(new Shape(numRows, 1)));
      //  } else if (problemData.Dataset.VariableHasType<DoubleVector>(variableName)) {
      //    var data = problemData.Dataset.GetDoubleVectorValues(variableName, rows).SelectMany(x => x.Select(y => (float)y)).ToArray();
      //    variablesFeed.Add(variablePlaceholder, np.array(data).reshape(new Shape(numRows, -1)));
      //  } else
      //    throw new NotSupportedException($"Type of the variable is not supported: {variableName}");
      //}
      //var targetData = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, rows).Select(x => (float)x).ToArray();
      //variablesFeed.Add(target, np.array(targetData));

      //using var session = tf.Session();

      //var loss2 = tf.constant(1.23f, TF_DataType.TF_FLOAT);

      //var graphOptimizer = tf.train.AdamOptimizer((float)learningRate);
      //var minimizationOperations = graphOptimizer.minimize(loss2);

      //var init = tf.global_variables_initializer();
      //session.run(init);

      //session.run((minimizationOperations, loss2), variablesFeed);

      
      

      return newTree;

      
//#if EXPORT_GRAPH
//      //https://github.com/SciSharp/TensorFlow.NET/wiki/Debugging
//      tf.train.export_meta_graph(@"C:\temp\TFboard\graph.meta", as_text: false,
//        clear_devices: true, clear_extraneous_savers: false, strip_default_attrs: true);
//#endif

      

//      //// features as feed items
//      //var variablesFeed = new Hashtable();
//      //foreach (var kvp in parameters) {
//      //  var variable = kvp.Key;
//      //  var variableName = kvp.Value;
//      //  if (problemData.Dataset.VariableHasType<double>(variableName)) {
//      //    var data = problemData.Dataset.GetDoubleValues(variableName, rows).Select(x => (float)x).ToArray();
//      //    variablesFeed.Add(variable, np.array(data).reshape(new Shape(numRows, 1)));
//      //  } else if (problemData.Dataset.VariableHasType<DoubleVector>(variableName)) {
//      //    var data = problemData.Dataset.GetDoubleVectorValues(variableName, rows).SelectMany(x => x.Select(y => (float)y)).ToArray();
//      //    variablesFeed.Add(variable, np.array(data).reshape(new Shape(numRows, -1)));
//      //  } else
//      //    throw new NotSupportedException($"Type of the variable is not supported: {variableName}");
//      //}
//      //var targetData = problemData.Dataset.GetDoubleValues(problemData.TargetVariable, rows).Select(x => (float)x).ToArray();
//      //variablesFeed.Add(target, np.array(targetData));

    }

    private static void UpdateConstants(ISymbolicExpressionTree tree, Dictionary<ISymbolicExpressionTreeNode, double[]> constants) {
      foreach (var kvp in constants) {
        var node = kvp.Key;
        var value = kvp.Value;

        switch (node) {
          case ConstantTreeNode constantTreeNode:
            constantTreeNode.Value = value[0];
            break;
          case VariableTreeNodeBase variableTreeNodeBase:
            variableTreeNodeBase.Weight = value[0];
            break;
          case FactorVariableTreeNode factorVarTreeNode: {
            for (int i = 0; i < factorVarTreeNode.Weights.Length; i++) {
              factorVarTreeNode.Weights[i] = value[i];
            }
            break;
          }
        }
      }
    }

    public static bool CanOptimizeConstants(ISymbolicExpressionTree tree) {
      return TreeToTensorConverter.IsCompatible(tree);
    }
  }
}