Logistic Regression

logistic-regression.js

var LogisticRegression = function() {
  this.N        = 0;  // Number of Samples
  this.D        = 0;  // Number of Dimensions|Features
  this.K        = 0;  // Number of Classes
  this.classes  = []; // Classes to classify
  this.weights  = []; // Weights
  this.input    = []; // [{0 1 0 1 0 ... 512} , {}]
  this.output   = []; // [{zero:0,one:1.0...}, ...]
}

// NOTE:
// Increase in input samples, decrease learning curve, decrease iterations ??


// Steps
// Every i/p sample gets a Weight Matrix (class x features)
// We try to modify the weight matrix such that dW is min (close to 0 or threshold)

// 1. Assign Random Weights for each feature for each class W[i][j] = rand(-0.01,0.01) [10x512]
// 2. Initialize dW -> 0 (error)
// 3. For every input sample (n)
//      For every class (i)
//        calculated_output[class] = 0 // for every i/p u calculate output of all classes
//        For every feature|dimension in i/p (j)
            // Calculate output , weight*feature of i/p
//          calculated_output[n][i] = calculated_output[n][i] + W[i][j]*X[n][j];
          // now translate calculated_output in 0-1
//        Y = exp(calculated_output[n][i])/sum(exp(calculated_output[n][i])) 
//        For every feature|dimension in i/p (j)
//          dW[i][j] = dW[i][j] + (actual_output[n][i]-y[n][i]) * X[n][j]
// 4. For every class (i)
//      For every dimension|feature (j)        
//        W[i][j] = W[i][j] + n * dW[i][j]
// 5. Repeat 2 till dW[i][j] close to zero

LogisticRegression.prototype = {
    run: function(input) {
      var self = this;
      var Y = [];
      input = "1 " + input;
      input = input.split(" ");
      for(var i=0; i<self.K; i++) {
        Y.push(0);
        for(var j=0; j < self.D; j++) {
          Y[i] = Y[i] + input[j] * self.weights[i][j];
        }
      }
      return Y;
    }
    
  , train:  function(data, iterations) {
      var self = this;
      // Lets Init some Data
      self.init(data);
      
      for(var z=0; z < 1000; z+=1) {
        var coutput = []
          , dW = []
          , flag = 0;
        // console.log("reset coutput", coutput);
        // Init dW
        for(var i=0; i < self.K; i++) {
          dW.push([]);
          // Calculate calculated output
          for(var j=0; j < self.D; j++) {
            dW[i].push(0);
          }
        }
        
        // Start Logistic Regression
        for(var n=0; n < self.N; n++) {
          coutput.push([]);
          for(var i=0; i < self.K; i++) {
            coutput[n].push(0);
            // Calculate calculated output
            for(var j=0; j < self.D; j++) {
              coutput[n][i] = coutput[n][i] + self.weights[i][j]*self.input[n][j];
            }
          }
          // Calculate Sum of calculated output
          var sum_coutput = 0;
          for(var i=0; i < self.K; i++) {
            sum_coutput = sum_coutput + Math.exp(coutput[n][i]);
          }
          // Translate calculated output using X function to put in 0-1
          for(var i=0; i < self.K; i++) {
            coutput[n][i] = Math.exp(coutput[n][i])/sum_coutput;
            // console.log("for iteration z =>" + z + ", for class " + i, coutput[n][i]);
            // Calculate dW
            for(var j=0; j < self.D; j++) {
              dW[i][j] = dW[i][j] + (self.output[n][i]-coutput[n][i]) * self.input[n][j];
              self.weights[i][j] = self.weights[i][j] + 0.00001 * dW[i][j]
            }
          }
        }
        
        for(var i=0; i < self.K; i++) {
          for(var j=0; j < self.D; j++) {
            if(Math.abs(dW[i][j]) <= 0.001) {
              flag = 1;
            } else {
              flag = 0;
              break;
            }
          }
          if(flag == 0) break;
        }
        console.log(z);
        if (flag == 1 || z == 50) {
          console.log("logistic regression, ready afer =>", z);
          break;
        } else {
          continue;
        }
      }
    }
    
  , init: function(data) {
      var self = this;
      self.N = n = data.length; // Length
      // Lets create ip X[i]
      for(var i=0; i < n; i++) {
        data[i].input = "1 " + data[i].input;
        self.input.push(data[i].input.split(" ")); // ip length x 512
        self.output.push(data[i].output); // {zero:0, one:0 ... }
      }
      // Number of Dimensions
      self.D = self.input[0].length;
      // generate output
      self.generateKOutput();
      // generate random weights for W
      self.generateWeightsMatrix();
      
      console.log(self.D, self.N, self.K);
    }
    
  , generateKOutput: function() {
      // init vars
      var self = this
        , output = [];
      input = self.output;
      self.classes = classes = [];
      n = self.N;
      // get classes
      for(var i=0; i < n; i++) {
        var key = Object.keys(input[i])[0];
        // if class not found, then insert into classes
        if(classes.indexOf(key) === -1) classes.push(key);
      }
      // Number of Classes
      self.K = classes.length;
      // now restructure output in 1..K encoding
      for(var i=0; i < n; i++) {
        output[i] = [];
        for(var j=0; j < self.K; j++) {
          if(Object.keys(input[i])[0] === classes[j]) {
            output[i].push(1);
          } else {
            output[i].push(0);
          }
        }
      }
      // Re write output
      self.output = output;
    }
    
  , generateWeightsMatrix: function() {
      var self = this;
      for(var i=0; i < self.K; i++) {
        self.weights[i] = [];
        for(var j=0; j< self.D; j++) {
          self.weights[i][j] = Math.random() * 0.4  - 0.2;
        }
      }
    }
    
  , toJSON : function() {
    }

  , fromJSON : function(json) {
    }
}

module.exports = LogisticRegression;