summaryrefslogtreecommitdiff
path: root/pro-train/mr_pro_reduce.cc
blob: aff410a0625cb6c7c6f7a77347d38e591ab415f5 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
#include <cstdlib>
#include <sstream>
#include <iostream>
#include <fstream>
#include <vector>

#include <boost/program_options.hpp>
#include <boost/program_options/variables_map.hpp>

#include "filelib.h"
#include "weights.h"
#include "sparse_vector.h"
#include "optimize.h"

using namespace std;
namespace po = boost::program_options;

// since this is a ranking model, there should be equal numbers of
// positive and negative examples, so the bias should be 0
static const double MAX_BIAS = 1e-10;

void InitCommandLine(int argc, char** argv, po::variables_map* conf) {
  po::options_description opts("Configuration options");
  opts.add_options()
        ("weights,w", po::value<string>(), "Weights from previous iteration (used as initialization and interpolation")
        ("interpolation,p",po::value<double>()->default_value(0.9), "Output weights are p*w + (1-p)*w_prev")
        ("memory_buffers,m",po::value<unsigned>()->default_value(200), "Number of memory buffers (LBFGS)")
        ("sigma_squared,s",po::value<double>()->default_value(0.1), "Sigma squared for Gaussian prior")
        ("min_reg,r",po::value<double>()->default_value(1e-8), "When tuning (-T) regularization strength, minimum regularization strenght")
        ("max_reg,R",po::value<double>()->default_value(10.0), "When tuning (-T) regularization strength, maximum regularization strenght")
        ("testset,t",po::value<string>(), "Optional held-out test set")
        ("tune_regularizer,T", "Use the held out test set (-t) to tune the regularization strength")
        ("help,h", "Help");
  po::options_description dcmdline_options;
  dcmdline_options.add(opts);
  po::store(parse_command_line(argc, argv, dcmdline_options), *conf);
  if (conf->count("help")) {
    cerr << dcmdline_options << endl;
    exit(1);
  }
}

void ParseSparseVector(string& line, size_t cur, SparseVector<weight_t>* out) {
  SparseVector<weight_t>& x = *out;
  size_t last_start = cur;
  size_t last_comma = string::npos;
  while(cur <= line.size()) {
    if (line[cur] == ' ' || cur == line.size()) {
      if (!(cur > last_start && last_comma != string::npos && cur > last_comma)) {
        cerr << "[ERROR] " << line << endl << "  position = " << cur << endl;
        exit(1);
      }
      const int fid = FD::Convert(line.substr(last_start, last_comma - last_start));
      if (cur < line.size()) line[cur] = 0;
      const weight_t val = strtod(&line[last_comma + 1], NULL);
      x.set_value(fid, val);

      last_comma = string::npos;
      last_start = cur+1;
    } else {
      if (line[cur] == '=')
        last_comma = cur;
    }
    ++cur;
  }
}

void ReadCorpus(istream* pin, vector<pair<bool, SparseVector<weight_t> > >* corpus) {
  istream& in = *pin;
  corpus->clear();
  bool flag = false;
  int lc = 0;
  string line;
  SparseVector<weight_t> x;
  while(getline(in, line)) {
    ++lc;
    if (lc % 1000 == 0) { cerr << '.'; flag = true; }
    if (lc % 40000 == 0) { cerr << " [" << lc << "]\n"; flag = false; }
    if (line.empty()) continue;
    const size_t ks = line.find("\t");
    assert(string::npos != ks);
    assert(ks == 1);
    const bool y = line[0] == '1';
    x.clear();
    ParseSparseVector(line, ks + 1, &x);
    corpus->push_back(make_pair(y, x));
  }
  if (flag) cerr << endl;
}

void GradAdd(const SparseVector<weight_t>& v, const double scale, vector<weight_t>* acc) {
  for (SparseVector<weight_t>::const_iterator it = v.begin();
       it != v.end(); ++it) {
    (*acc)[it->first] += it->second * scale;
  }
}

double TrainingInference(const vector<weight_t>& x,
                         const vector<pair<bool, SparseVector<weight_t> > >& corpus,
                         vector<weight_t>* g = NULL) {
  double cll = 0;
  for (int i = 0; i < corpus.size(); ++i) {
    const double dotprod = corpus[i].second.dot(x) + (x.size() ? x[0] : weight_t()); // x[0] is bias
    double lp_false = dotprod;
    double lp_true = -dotprod;
    if (0 < lp_true) {
      lp_true += log1p(exp(-lp_true));
      lp_false = log1p(exp(lp_false));
    } else {
      lp_true = log1p(exp(lp_true));
      lp_false += log1p(exp(-lp_false));
    }
    lp_true*=-1;
    lp_false*=-1;
    if (corpus[i].first) {  // true label
      cll -= lp_true;
      if (g) {
        // g -= corpus[i].second * exp(lp_false);
        GradAdd(corpus[i].second, -exp(lp_false), g);
        (*g)[0] -= exp(lp_false); // bias
      }
    } else {                  // false label
      cll -= lp_false;
      if (g) {
        // g += corpus[i].second * exp(lp_true);
        GradAdd(corpus[i].second, exp(lp_true), g);
        (*g)[0] += exp(lp_true); // bias
      }
    }
  }
  return cll;
}

// return held-out log likelihood
double LearnParameters(const vector<pair<bool, SparseVector<weight_t> > >& training,
                       const vector<pair<bool, SparseVector<weight_t> > >& testing,
                       const double sigsq,
                       const unsigned memory_buffers,
                       vector<weight_t>* px) {
  vector<weight_t>& x = *px;
  vector<weight_t> vg(FD::NumFeats(), 0.0);
  bool converged = false;
  LBFGSOptimizer opt(FD::NumFeats(), memory_buffers);
  double tppl = 0.0;
  while(!converged) {
    fill(vg.begin(), vg.end(), 0.0);
    double cll = TrainingInference(x, training, &vg);
    double ppl = cll / log(2);
    ppl /= training.size();
    ppl = pow(2.0, ppl);

    // evaluate optional held-out test set
    if (testing.size()) {
      tppl = TrainingInference(x, testing) / log(2);
      tppl /= testing.size();
      tppl = pow(2.0, tppl);
    }

    // handle regularizer
#if 1
    double norm = 0;
    for (int i = 1; i < x.size(); ++i) {
      const double mean_i = 0.0;
      const double param = (x[i] - mean_i);
      norm += param * param;
      vg[i] += param / sigsq;
    } 
    const double reg = norm / (2.0 * sigsq);
#else
    double reg = 0;
#endif
    cll += reg;
    cerr << cll << " (REG=" << reg << ")\tPPL=" << ppl << "\t TEST_PPL=" << tppl << "\t";
    try {
      vector<weight_t> old_x = x;
      do {
        opt.Optimize(cll, vg, &x);
        converged = opt.HasConverged();
      } while (!converged && x == old_x);
    } catch (...) {
      cerr << "Exception caught, assuming convergence is close enough...\n";
      converged = true;
    }
    if (fabs(x[0]) > MAX_BIAS) {
      cerr << "Biased model learned. Are your training instances wrong?\n";
      cerr << "  BIAS: " << x[0] << endl;
    }
  }
  return tppl;
}

int main(int argc, char** argv) {
  po::variables_map conf;
  InitCommandLine(argc, argv, &conf);
  string line;
  vector<pair<bool, SparseVector<weight_t> > > training, testing;
  const bool tune_regularizer = conf.count("tune_regularizer");
  if (tune_regularizer && !conf.count("testset")) {
    cerr << "--tune_regularizer requires --testset to be set\n";
    return 1;
  }
  const double min_reg = conf["min_reg"].as<double>();
  const double max_reg = conf["max_reg"].as<double>();
  double sigsq = conf["sigma_squared"].as<double>(); // will be overridden if parameter is tuned
  assert(sigsq > 0.0);
  assert(min_reg > 0.0);
  assert(max_reg > 0.0);
  assert(max_reg > min_reg);
  const double psi = conf["interpolation"].as<double>();
  if (psi < 0.0 || psi > 1.0) { cerr << "Invalid interpolation weight: " << psi << endl; }
  ReadCorpus(&cin, &training);
  if (conf.count("testset")) {
    ReadFile rf(conf["testset"].as<string>());
    ReadCorpus(rf.stream(), &testing);
  }
  cerr << "Number of features: " << FD::NumFeats() << endl;

  vector<weight_t> x, prev_x;  // x[0] is bias
  if (conf.count("weights")) {
    Weights::InitFromFile(conf["weights"].as<string>(), &x);
    x.resize(FD::NumFeats());
    prev_x = x;
  } else {
    x.resize(FD::NumFeats());
    prev_x = x;
  }
  cerr << "         Number of features: " << x.size() << endl;
  cerr << "Number of training examples: " << training.size() << endl;
  cerr << "Number of  testing examples: " << testing.size() << endl;
  double tppl = 0.0;
  vector<pair<double,double> > sp;
  vector<double> smoothed;
  if (tune_regularizer) {
    sigsq = min_reg;
    const double steps = 18;
    double sweep_factor = exp((log(max_reg) - log(min_reg)) / steps);
    cerr << "SWEEP FACTOR: " << sweep_factor << endl;
    while(sigsq < max_reg) {
      tppl = LearnParameters(training, testing, sigsq, conf["memory_buffers"].as<unsigned>(), &x);
      sp.push_back(make_pair(sigsq, tppl));
      sigsq *= sweep_factor;
    }
    smoothed.resize(sp.size(), 0);
    smoothed[0] = sp[0].second;
    smoothed.back() = sp.back().second; 
    for (int i = 1; i < sp.size()-1; ++i) {
      double prev = sp[i-1].second;
      double next = sp[i+1].second;
      double cur = sp[i].second;
      smoothed[i] = (prev*0.2) + cur * 0.6 + (0.2*next);
    }
    double best_ppl = 9999999;
    unsigned best_i = 0;
    for (unsigned i = 0; i < sp.size(); ++i) {
      if (smoothed[i] < best_ppl) {
        best_ppl = smoothed[i];
        best_i = i;
      }
    }
    sigsq = sp[best_i].first;
  }  // tune regularizer
  tppl = LearnParameters(training, testing, sigsq, conf["memory_buffers"].as<unsigned>(), &x);
  if (conf.count("weights")) {
    for (int i = 1; i < x.size(); ++i) {
      x[i] = (x[i] * psi) + prev_x[i] * (1.0 - psi);
    }
  }
  cout.precision(15);
  cout << "# sigma^2=" << sigsq << "\theld out perplexity=";
  if (tppl) { cout << tppl << endl; } else { cout << "N/A\n"; }
  if (sp.size()) {
    cout << "# Parameter sweep:\n";
    for (int i = 0; i < sp.size(); ++i) {
      cout << "# " << sp[i].first << "\t" << sp[i].second << "\t" << smoothed[i] << endl;
    }
  }
  Weights::WriteToFile("-", x);
  return 0;
}