path: root/cg.cc

/*
Copyright (c) 2009 Yahoo! Inc.  All rights reserved.  The copyrights
embodied in the content of this file are licensed under the BSD
(revised) open source license

The algorithm here is generally based on Jonathan Shewchuck's tutorial.

 */
#include <fstream>
#include <float.h>
#include <netdb.h>
#include <string.h>
#include <stdio.h>
#include <assert.h>
#include "parse_example.h"
#include "constant.h"
#include "sparse_dense.h"
#include "cg.h"
#include "cache.h"
#include "multisource.h"
#include "simple_label.h"
#include "delay_ring.h"

void quad_grad_update(weight* weights, feature& page_feature, v_array<feature> &offer_features, size_t mask, float g)
{
  size_t halfhash = quadratic_constant * page_feature.weight_index;
  float update = g * page_feature.x;
  for (feature* ele = offer_features.begin; ele != offer_features.end; ele++)
    {
      weight* w=&weights[(halfhash + ele->weight_index) & mask];
      w[1] += update * ele->x;
    }
}

void quad_precond_update(weight* weights, feature& page_feature, v_array<feature> &offer_features, size_t mask, float g)
{
  size_t halfhash = quadratic_constant * page_feature.weight_index;
  float update = g * page_feature.x;
  for (feature* ele = offer_features.begin; ele != offer_features.end; ele++)
    {
      weight* w=&weights[(halfhash + ele->weight_index) & mask];
      w[4] += update * ele->x * ele->x;
    }
}

// w[0] = weight
// w[1] = accumulated first derivative
// w[2] = step direction
// w[3] = old first derivative
// w[4] = preconditioner

float predict_and_gradient(regressor& reg, example* &ec)
{
  float raw_prediction = inline_predict(reg,ec,0);
  float fp = finalize_prediction(raw_prediction);
  
  label_data* ld = (label_data*)ec->ld;

  float loss_grad = reg.loss->first_derivative(fp,ld->label)*ld->weight;
  
  size_t thread_mask = global.thread_mask;
  weight* weights = reg.weight_vectors[0];
  for (size_t* i = ec->indices.begin; i != ec->indices.end; i++) 
    {
      feature *f = ec->subsets[*i][0];
      for (; f != ec->subsets[*i][1]; f++)
	{
	  weight* w = &weights[f->weight_index & thread_mask];
	  w[1] += loss_grad * f->x;
	}
    }
  for (vector<string>::iterator i = global.pairs.begin(); i != global.pairs.end();i++) 
    {
      if (ec->subsets[(int)(*i)[0]].index() > 0)
	{
	  v_array<feature> temp = ec->atomics[(int)(*i)[0]];
	  temp.begin = ec->subsets[(int)(*i)[0]][0];
	  temp.end = ec->subsets[(int)(*i)[0]][1];
	  for (; temp.begin != temp.end; temp.begin++)
	    quad_grad_update(weights, *temp.begin, ec->atomics[(int)(*i)[1]], thread_mask, loss_grad);
	} 
    }
  return fp;
}

void update_preconditioner(regressor& reg, example* &ec)
{
  label_data* ld = (label_data*)ec->ld;
  float curvature = reg.loss->second_derivative(ec->final_prediction,ld->label) * ld->weight;
  
  size_t thread_mask = global.thread_mask;
  weight* weights = reg.weight_vectors[0];
  for (size_t* i = ec->indices.begin; i != ec->indices.end; i++)
    {
      feature *f = ec->subsets[*i][0];
      for (; f != ec->subsets[*i][1]; f++)
        {
          weight* w = &weights[f->weight_index & thread_mask];
          w[4] += f->x * f->x * curvature;
        }
    }
  for (vector<string>::iterator i = global.pairs.begin(); i != global.pairs.end();i++)
    {
      if (ec->subsets[(int)(*i)[0]].index() > 0)
        {
          v_array<feature> temp = ec->atomics[(int)(*i)[0]];
          temp.begin = ec->subsets[(int)(*i)[0]][0];
          temp.end = ec->subsets[(int)(*i)[0]][1];
          for (; temp.begin != temp.end; temp.begin++)
            quad_precond_update(weights, *temp.begin, ec->atomics[(int)(*i)[1]], thread_mask, curvature);
        }
    }
}  

float dot_with_direction(regressor& reg, example* &ec)
{
  float ret = 0;
  weight* weights = reg.weight_vectors[0];
  size_t thread_mask = global.thread_mask;
  weights +=2;//direction vector stored two advanced
  for (size_t* i = ec->indices.begin; i != ec->indices.end; i++) 
    {
      feature *f = ec->subsets[*i][0];
      for (; f != ec->subsets[*i][1]; f++)
	ret += weights[f->weight_index & thread_mask] * f->x;
    }
  for (vector<string>::iterator i = global.pairs.begin(); i != global.pairs.end();i++) 
    {
      if (ec->subsets[(int)(*i)[0]].index() > 0)
	{
	  v_array<feature> temp = ec->atomics[(int)(*i)[0]];
	  temp.begin = ec->subsets[(int)(*i)[0]][0];
	  temp.end = ec->subsets[(int)(*i)[0]][1];
	  for (; temp.begin != temp.end; temp.begin++)
	    ret += one_pf_quad_predict(weights, *temp.begin, ec->atomics[(int)(*i)[1]], thread_mask);
	} 
    }
  return ret;
}

double derivative_magnitude(regressor& reg)
{//compute derivative magnitude & shift new derivative to old
  double ret = 0.;
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  weight* weights = reg.weight_vectors[0];//shift by one for ease of indexing.
  for(uint32_t i = 0; i < length; i++)
    {
      ret += weights[stride*i+1]*weights[stride*i+1]*weights[stride*i+4];
      weights[stride*i+3] = weights[stride*i+1];
      weights[stride*i+1] = 0;
    }
  return ret;
}

void zero_derivative(regressor& reg)
{//compute derivative magnitude & shift new derivative to old
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  weight* weights = reg.weight_vectors[0];//shift by one for ease of indexing.
  for(uint32_t i = 0; i < length; i++)
    weights[stride*i+1] = 0;
}

double direction_magnitude(regressor& reg)
{//compute direction magnitude
  double ret = 0.;
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  weight* weights = reg.weight_vectors[0];
  for(uint32_t i = 0; i < length; i++)
    ret += weights[stride*i+2]*weights[stride*i+2];
  
  return ret;
}

double derivative_diff_mag(regressor& reg)
{//compute the derivative difference
  double ret = 0.;
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  weight* weights = reg.weight_vectors[0];
  for(uint32_t i = 0; i < length; i++)
    {
      ret += weights[stride*i+1]*weights[stride*i+4]*
	(weights[stride*i+1] - weights[stride*i+3]);
    }
  return ret;
}

double add_regularization(regressor& reg,float regularization)
{//compute the derivative difference
  double ret = 0.;
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  weight* weights = reg.weight_vectors[0];
  for(uint32_t i = 0; i < length; i++)
    weights[stride*i+1] += regularization*weights[stride*i];

  return ret;
}

void finalize_preconditioner(regressor& reg,float regularization)
{
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  weight* weights = reg.weight_vectors[0];
  for(uint32_t i = 0; i < length; i++) {
    weights[stride*i+4] += regularization;
    if (weights[stride*i+4] > 0)
      weights[stride*i+4] = 1. / weights[stride*i+4];
  }
}

double derivative_in_direction(regressor& reg)
{
  double ret = 0.;
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  for(uint32_t i = 0; i < length; i++)
    ret += reg.weight_vectors[0][stride*i+3]*reg.weight_vectors[0][stride*i+2];
  return ret;
}

void update_direction(regressor& reg, float old_portion)
{
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  weight* weights = reg.weight_vectors[0];
  for(uint32_t i = 0; i < length; i++)
    {
      weights[stride*i+2] = weights[stride*i+3]*weights[stride*i+4] + old_portion * weights[stride*i+2];
    }
}

void update_weight(regressor& reg, float step_size)
{
  uint32_t length = 1 << global.num_bits;
  size_t stride = global.stride;
  for(uint32_t i = 0; i < length; i++)
    {
      reg.weight_vectors[0][stride*i] += step_size * reg.weight_vectors[0][stride*i+2];
    }
}

void setup_cg(gd_thread_params t)
{
  regressor reg = t.reg;
  size_t thread_num = 0;
  example* ec = NULL;

  v_array<float> predictions;
  size_t example_number=0;
  double curvature=0.;

  bool gradient_pass=true;
  double loss_sum = 0;
  float step_size = 0.;
  double importance_weight_sum = 0.;
 
  double previous_d_mag=0;
  size_t current_pass = 0;
  double previous_loss_sum = 0;
  while ( true )
    {
      if ((ec = get_example(thread_num)) != NULL)//semiblocking operation.
	{
	  assert(ec->in_use);
	  if (ec->pass != current_pass)//we need to do work on all features.
	    {
	      if (current_pass == 0)
		finalize_preconditioner(reg,global.regularization*importance_weight_sum);
	      if (gradient_pass) // We just finished computing all gradients
		if (current_pass > 0 && loss_sum > previous_loss_sum)
		  {// we stepped to far last time, step back
		    step_size *= 0.5;
		    cout << "backstepping, new step_size = " << step_size << endl;
		    update_weight(reg,- step_size);
		    zero_derivative(reg);
		    loss_sum = 0.;
		  }
		else
		  {
		    previous_loss_sum = loss_sum;
		    loss_sum = 0.;
		    if (global.regularization > 0.)
		      add_regularization(reg,global.regularization*importance_weight_sum);
		    example_number = 0;
		    curvature = 0;
		    float mix_frac = 0;
		    if (current_pass != 0)
		      mix_frac = derivative_diff_mag(reg) / previous_d_mag;
		    if (mix_frac < 0 || isnan(mix_frac))
		      mix_frac = 0;
		    float new_d_mag = derivative_magnitude(reg);
		    previous_d_mag = new_d_mag;
		    
		    update_direction(reg, mix_frac);
		    gradient_pass = false;//now start computing curvature
		  }
	      else // just finished all second gradients
		{
		  if (global.regularization > 0.)
		    curvature += global.regularization*direction_magnitude(reg)*importance_weight_sum;
		  step_size = - derivative_in_direction(reg)/(max(curvature,1.));
		  predictions.erase();
		  update_weight(reg,step_size);
		  gradient_pass = true;
		}//now start computing derivatives.
	      current_pass++;
	    }
	  if (gradient_pass)
	    {
	      ec->final_prediction = predict_and_gradient(reg,ec);
	      if (current_pass == 0)
		{
		  label_data* ld = (label_data*)ec->ld;
		  importance_weight_sum += ld->weight;
		  update_preconditioner(reg,ec);
		}
	      label_data* ld = (label_data*)ec->ld;
	      ec->loss = reg.loss->getLoss(ec->final_prediction, ld->label) * ld->weight;
	      loss_sum += ec->loss;
	      push(predictions,ec->final_prediction);
	    }
	  else //computing curvature
	    {
	      float d_dot_x = dot_with_direction(reg,ec);
	      label_data* ld = (label_data*)ec->ld;
	      ec->final_prediction = predictions[example_number];
	      ec->loss = reg.loss->getLoss(ec->final_prediction, ld->label) * ld->weight;	      
	      float sd = reg.loss->second_derivative(predictions[example_number++],ld->label);
	      curvature += d_dot_x*d_dot_x*sd*ld->weight;
	    }
	  finish_example(ec);
	}
      else if (thread_done(thread_num))
	{
	  if (example_number == predictions.index())//do one last update
	    {
	      if (global.regularization > 0.)
		curvature += global.regularization*direction_magnitude(reg)*importance_weight_sum;
	      float step_size = - derivative_in_direction(reg)/(max(curvature,1.));
	      update_weight(reg,step_size);
	    }
	  if (global.local_prediction > 0)
	    shutdown(global.local_prediction, SHUT_WR);
	  free(predictions.begin);
	  return;
	}
      else 
	;//busywait when we have predicted on all examples but not yet trained on all.
    }

  free(predictions.begin);
  return;
}

void destroy_cg()
{
}