global_data.cc

Go to the documentation of this file.

/*
Copyright (c) by respective owners including Yahoo!, Microsoft, and
individual contributors. All rights reserved.  Released under a BSD (revised)
license as described in the file LICENSE.
 */
#include <stdio.h>
#include <float.h>
#include <errno.h>
#include <iostream>
#include <sstream>
#include <math.h>
#include <assert.h>

#include "global_data.h"
#include "gd.h"
#include "vw_exception.h"

struct global_prediction
{
  float p;
  float weight;
};

size_t really_read(int sock, void* in, size_t count)
{
  char* buf = (char*)in;
  size_t done = 0;
  int r = 0;
  while (done < count)
  {
    if ((r =
#ifdef _WIN32
                recv(sock, buf, (unsigned int)(count - done), 0)
#else
                read(sock, buf, (unsigned int)(count - done))
#endif
                ) == 0)
      return 0;
    else if (r < 0)
    {
      THROWERRNO("read(" << sock << "," << count << "-" << done << ")");
    }
    else
    {
      done += r;
      buf += r;
    }
  }
  return done;
}

void get_prediction(int sock, float& res, float& weight)
{
  global_prediction p;
  really_read(sock, &p, sizeof(p));
  res = p.p;
  weight = p.weight;
}

void send_prediction(int sock, global_prediction p)
{
  if (
#ifdef _WIN32
      send(sock, reinterpret_cast<const char*>(&p), sizeof(p), 0)
#else
      write(sock, &p, sizeof(p))
#endif
      < (int)sizeof(p))
    THROWERRNO("send_prediction write(" << sock << ")");
}

void binary_print_result(int f, float res, float weight, v_array<char>)
{
  if (f >= 0)
  {
    global_prediction ps = {res, weight};
    send_prediction(f, ps);
  }
}

int print_tag(std::stringstream& ss, v_array<char> tag)
{
  if (tag.begin() != tag.end())
  {
    ss << ' ';
    ss.write(tag.begin(), sizeof(char) * tag.size());
  }
  return tag.begin() != tag.end();
}

void print_result(int f, float res, float, v_array<char> tag)
{
  if (f >= 0)
  {
    std::stringstream ss;
    auto saved_precision = ss.precision();
    if (floorf(res) == res)
      ss << std::setprecision(0);
    ss << std::fixed << res << std::setprecision(saved_precision);
    print_tag(ss, tag);
    ss << '\n';
    ssize_t len = ss.str().size();
    ssize_t t = io_buf::write_file_or_socket(f, ss.str().c_str(), (unsigned int)len);
    if (t != len)
    {
      std::cerr << "write error: " << strerror(errno) << std::endl;
    }
  }
}

void print_raw_text(int f, std::string s, v_array<char> tag)
{
  if (f < 0)
    return;

  std::stringstream ss;
  ss << s;
  print_tag(ss, tag);
  ss << '\n';
  ssize_t len = ss.str().size();
  ssize_t t = io_buf::write_file_or_socket(f, ss.str().c_str(), (unsigned int)len);
  if (t != len)
  {
    std::cerr << "write error: " << strerror(errno) << std::endl;
  }
}

void set_mm(shared_data* sd, float label)
{
  sd->min_label = std::min(sd->min_label, label);
  if (label != FLT_MAX)
    sd->max_label = std::max(sd->max_label, label);
}

void noop_mm(shared_data*, float) {}

void vw::learn(example& ec)
{
  if (l->is_multiline)
    THROW("This reduction does not support single-line examples.");

  if (ec.test_only || !training)
    LEARNER::as_singleline(l)->predict(ec);
  else
    LEARNER::as_singleline(l)->learn(ec);
}

void vw::learn(multi_ex& ec)
{
  if (!l->is_multiline)
    THROW("This reduction does not support multi-line example.");

  if (!training)
    LEARNER::as_multiline(l)->predict(ec);
  else
    LEARNER::as_multiline(l)->learn(ec);
}

void vw::predict(example& ec)
{
  if (l->is_multiline)
    THROW("This reduction does not support single-line examples.");

  LEARNER::as_singleline(l)->predict(ec);
}

void vw::predict(multi_ex& ec)
{
  if (!l->is_multiline)
    THROW("This reduction does not support multi-line example.");

  LEARNER::as_multiline(l)->predict(ec);
}

void vw::finish_example(example& ec)
{
  if (l->is_multiline)
    THROW("This reduction does not support single-line examples.");

  LEARNER::as_singleline(l)->finish_example(*this, ec);
}

void vw::finish_example(multi_ex& ec)
{
  if (!l->is_multiline)
    THROW("This reduction does not support multi-line example.");

  LEARNER::as_multiline(l)->finish_example(*this, ec);
}

void compile_gram(
    std::vector<std::string> grams, std::array<uint32_t, NUM_NAMESPACES>& dest, char* descriptor, bool quiet)
{
  for (size_t i = 0; i < grams.size(); i++)
  {
    std::string ngram = grams[i];
    if (isdigit(ngram[0]))
    {
      int n = atoi(ngram.c_str());
      if (!quiet)
        std::cerr << "Generating " << n << "-" << descriptor << " for all namespaces." << std::endl;
      for (size_t j = 0; j < 256; j++) dest[j] = n;
    }
    else if (ngram.size() == 1)
      std::cout << "You must specify the namespace index before the n" << std::endl;
    else
    {
      int n = atoi(ngram.c_str() + 1);
      dest[(uint32_t)(unsigned char)*ngram.c_str()] = n;
      if (!quiet)
        std::cerr << "Generating " << n << "-" << descriptor << " for " << ngram[0] << " namespaces." << std::endl;
    }
  }
}

void compile_limits(std::vector<std::string> limits, std::array<uint32_t, NUM_NAMESPACES>& dest, bool quiet)
{
  for (size_t i = 0; i < limits.size(); i++)
  {
    std::string limit = limits[i];
    if (isdigit(limit[0]))
    {
      int n = atoi(limit.c_str());
      if (!quiet)
        std::cerr << "limiting to " << n << "features for each namespace." << std::endl;
      for (size_t j = 0; j < 256; j++) dest[j] = n;
    }
    else if (limit.size() == 1)
      std::cout << "You must specify the namespace index before the n" << std::endl;
    else
    {
      int n = atoi(limit.c_str() + 1);
      dest[(uint32_t)limit[0]] = n;
      if (!quiet)
        std::cerr << "limiting to " << n << " for namespaces " << limit[0] << std::endl;
    }
  }
}

void trace_listener_cerr(void*, const std::string& message)
{
  std::cerr << message;
  std::cerr.flush();
}

int vw_ostream::vw_streambuf::sync()
{
  int ret = std::stringbuf::sync();
  if (ret)
    return ret;

  parent.trace_listener(parent.trace_context, str());
  str("");
  return 0;  // success
}

vw_ostream::vw_ostream() : std::ostream(&buf), buf(*this), trace_context(nullptr)
{
  trace_listener = trace_listener_cerr;
}

vw::vw()
{
  sd = &calloc_or_throw<shared_data>();
  sd->dump_interval = 1.;  // next update progress dump
  sd->contraction = 1.;
  sd->first_observed_label = FLT_MAX;
  sd->is_more_than_two_labels_observed = false;
  sd->max_label = 0;
  sd->min_label = 0;

  label_type = label_type::simple;

  l = nullptr;
  scorer = nullptr;
  cost_sensitive = nullptr;
  loss = nullptr;
  p = nullptr;

  reg_mode = 0;
  current_pass = 0;

  data_filename = "";
  delete_prediction = nullptr;

  bfgs = false;
  no_bias = false;
  hessian_on = false;
  active = false;
  num_bits = 18;
  default_bits = true;
  daemon = false;
  num_children = 10;
  save_resume = false;
  preserve_performance_counters = false;

  random_positive_weights = false;

  weights.sparse = false;

  set_minmax = set_mm;

  power_t = 0.5;
  eta = 0.5;  // default learning rate for normalized adaptive updates, this is switched to 10 by default for the other
              // updates (see parse_args.cc)
  numpasses = 1;

  final_prediction_sink.begin() = final_prediction_sink.end() = final_prediction_sink.end_array = nullptr;
  raw_prediction = -1;
  print = print_result;
  print_text = print_raw_text;
  lda = 0;
  random_seed = 0;
  random_weights = false;
  normal_weights = false;
  tnormal_weights = false;
  per_feature_regularizer_input = "";
  per_feature_regularizer_output = "";
  per_feature_regularizer_text = "";

#ifdef _WIN32
  stdout_fileno = _fileno(stdout);
#else
  stdout_fileno = fileno(stdout);
#endif

  searchstr = nullptr;

  nonormalize = false;
  l1_lambda = 0.0;
  l2_lambda = 0.0;

  eta_decay_rate = 1.0;
  initial_weight = 0.0;
  initial_constant = 0.0;

  all_reduce = nullptr;

  for (size_t i = 0; i < 256; i++)
  {
    ngram[i] = 0;
    skips[i] = 0;
    limit[i] = INT_MAX;
    affix_features[i] = 0;
    spelling_features[i] = 0;
  }

  invariant_updates = true;
  normalized_idx = 2;

  add_constant = true;
  audit = false;

  pass_length = std::numeric_limits<size_t>::max();
  passes_complete = 0;

  save_per_pass = false;

  stdin_off = false;
  do_reset_source = false;
  holdout_set_off = true;
  holdout_after = 0;
  check_holdout_every_n_passes = 1;
  early_terminate = false;

  max_examples = std::numeric_limits<size_t>::max();

  hash_inv = false;
  print_invert = false;

  // Set by the '--progress <arg>' option and affect sd->dump_interval
  progress_add = false;  // default is multiplicative progress dumps
  progress_arg = 2.0;    // next update progress dump multiplier

  sd->is_more_than_two_labels_observed = false;
  sd->first_observed_label = FLT_MAX;
  sd->second_observed_label = FLT_MAX;

  sd->report_multiclass_log_loss = false;
  sd->multiclass_log_loss = 0;
  sd->holdout_multiclass_log_loss = 0;
}