Edinburgh Speech Tools  2.4-release
EST_Wagon.h
1 /*************************************************************************/
2 /* */
3 /* Centre for Speech Technology Research */
4 /* University of Edinburgh, UK */
5 /* Copyright (c) 1996,1997 */
6 /* All Rights Reserved. */
7 /* */
8 /* Permission is hereby granted, free of charge, to use and distribute */
9 /* this software and its documentation without restriction, including */
10 /* without limitation the rights to use, copy, modify, merge, publish, */
11 /* distribute, sublicense, and/or sell copies of this work, and to */
12 /* permit persons to whom this work is furnished to do so, subject to */
13 /* the following conditions: */
14 /* 1. The code must retain the above copyright notice, this list of */
15 /* conditions and the following disclaimer. */
16 /* 2. Any modifications must be clearly marked as such. */
17 /* 3. Original authors' names are not deleted. */
18 /* 4. The authors' names are not used to endorse or promote products */
19 /* derived from this software without specific prior written */
20 /* permission. */
21 /* */
22 /* THE UNIVERSITY OF EDINBURGH AND THE CONTRIBUTORS TO THIS WORK */
23 /* DISCLAIM ALL WARRANTIES WITH REGARD TO THIS SOFTWARE, INCLUDING */
24 /* ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN NO EVENT */
25 /* SHALL THE UNIVERSITY OF EDINBURGH NOR THE CONTRIBUTORS BE LIABLE */
26 /* FOR ANY SPECIAL, INDIRECT OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES */
27 /* WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN */
28 /* AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, */
29 /* ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE OF */
30 /* THIS SOFTWARE. */
31 /* */
32 /*************************************************************************/
33 /* Author : Alan W Black */
34 /* Date : May 1996 */
35 /*-----------------------------------------------------------------------*/
36 /* */
37 /* Public declarations for Wagon (CART builder) */
38 /* */
39 /*=======================================================================*/
40 #ifndef __WAGON_H__
41 #define __WAGON_H__
42 
43 #include "EST_String.h"
44 #include "EST_Val.h"
45 #include "EST_TVector.h"
46 #include "EST_TList.h"
47 #include "EST_simplestats.h" /* For EST_SuffStats class */
48 #include "EST_Track.h"
49 #include "siod.h"
50 
51 // When set to one wagon supports using multiple threads if
52 // --omp_nthreads X is used (works for most gccs)
53 // #define OMP_WAGON 1
54 #ifdef OMP_WAGON
55 #include "omp.h"
56 #endif
57 
58 #define wagon_error(WMESS) (cerr << WMESS << endl,exit(-1))
59 
60 // I get floating point exceptions of Alphas when I do any comparisons
61 // with HUGE_VAL or FLT_MAX so I'll make my own
62 #define WGN_HUGE_VAL 1.0e20
63 
64 class WVector : public EST_FVector
65 {
66  public:
67  WVector(int n) : EST_FVector(n) {}
68  int get_int_val(int n) const { return (int)a_no_check(n); }
69  float get_flt_val(int n) const { return a_no_check(n); }
70  void set_int_val(int n,int i) { a_check(n) = (int)i; }
71  void set_flt_val(int n,float f) { a_check(n) = f; }
72 };
73 
76 
77 /* Different types of feature */
78 enum wn_dtype {/* for predictees and predictors */
79  wndt_binary, wndt_float, wndt_class,
80  /* for predictees only */
81  wndt_cluster, wndt_vector, wndt_matrix, wndt_trajectory,
82  wndt_ols,
83  /* for ignored features */
84  wndt_ignore};
85 
86 class WDataSet : public WVectorList {
87  private:
88  int dlength;
89  EST_IVector p_type;
90  EST_IVector p_ignore;
91  EST_StrVector p_name;
92  public:
93  void load_description(const EST_String& descfname,LISP ignores);
94  void ignore_non_numbers();
95 
96  int ftype(const int &i) const {return p_type(i);}
97  int ignore(int i) const {return p_ignore(i); }
98  void set_ignore(int i,int value) { p_ignore[i] = value; }
99  const EST_String &feat_name(const int &i) const {return p_name(i);}
100  int samples(void) const {return length();}
101  int width(void) const {return dlength;}
102 };
103 enum wn_oper {wnop_equal, wnop_binary, wnop_greaterthan,
104  wnop_lessthan, wnop_is, wnop_in, wnop_matches};
105 
106 class WQuestion {
107  private:
108  int feature_pos;
109  wn_oper op;
110  int yes;
111  int no;
112  EST_Val operand1;
113  EST_IList operandl;
114  float score;
115  public:
116  WQuestion() {;}
117  WQuestion(const WQuestion &s)
118  { feature_pos=s.feature_pos;
119  op=s.op; yes=s.yes; no=s.no; operand1=s.operand1;
120  operandl = s.operandl; score=s.score;}
121  ~WQuestion() {;}
122  WQuestion(int fp, wn_oper o,EST_Val a)
123  { feature_pos=fp; op=o; operand1=a; }
124  void set_fp(const int &fp) {feature_pos=fp;}
125  void set_oper(const wn_oper &o) {op=o;}
126  void set_operand1(const EST_Val &a) {operand1 = a;}
127  void set_yes(const int &y) {yes=y;}
128  void set_no(const int &n) {no=n;}
129  int get_yes(void) const {return yes;}
130  int get_no(void) const {return no;}
131  const int get_fp(void) const {return feature_pos;}
132  const wn_oper get_op(void) const {return op;}
133  const EST_Val get_operand1(void) const {return operand1;}
134  const EST_IList &get_operandl(void) const {return operandl;}
135  const float get_score(void) const {return score;}
136  void set_score(const float &f) {score=f;}
137  const int ask(const WVector &w) const;
138  friend ostream& operator<<(ostream& s, const WQuestion &q);
139 };
140 
141 enum wnim_type {wnim_unset, wnim_float, wnim_class,
142  wnim_cluster, wnim_vector, wnim_matrix, wnim_ols,
143  wnim_trajectory};
144 
145 // Impurity measure for cumulating impurities from set of data
146 class WImpurity {
147  private:
148  wnim_type t;
149  EST_SuffStats a;
151 
152  float cluster_impurity();
153  float cluster_member_mean(int i);
154  float vector_impurity();
155  float trajectory_impurity();
156  float ols_impurity();
157  public:
158  EST_IList members; // Maybe there should be a cluster class
159  EST_FList member_counts; // AUP: Implement counts for vectors
160  EST_SuffStats **trajectory;
161  const WVectorVector *data; // Needed for ols
162  float score;
163  int l,width;
164 
165  WImpurity() { t=wnim_unset; a.reset(); trajectory=0; l=0; width=0; data=0;}
166  ~WImpurity();
167  WImpurity(const WVectorVector &ds);
168  void copy(const WImpurity &s)
169  {
170  int i,j;
171  t=s.t; a=s.a; p=s.p; members=s.members; member_counts = s.member_counts; l=s.l; width=s.width;
172  score = s.score;
173  data = s.data;
174  if (s.trajectory)
175  {
176  trajectory = new EST_SuffStats *[l];
177  for (i=0; i<l; i++)
178  {
179  trajectory[i] = new EST_SuffStats[width];
180  for (j=0; j<width; j++)
181  trajectory[i][j] = s.trajectory[i][j];
182  }
183  }
184  }
185  WImpurity &operator = (const WImpurity &a) { copy(a); return *this; }
186 
187  float measure(void);
188  double samples(void);
189  wnim_type type(void) const { return t;}
190  void cumulate(const float pv,double count=1.0);
191  EST_Val value(void);
192  EST_DiscreteProbDistribution &pd() { return p; }
193  float cluster_distance(int i); // distance i from centre in sds
194  int in_cluster(int i); // distance i from centre < most remote member
195  float cluster_ranking(int i); // position in closeness to centre
196  friend ostream& operator<<(ostream &s, WImpurity &imp);
197 };
198 
199 class WDlist {
200  private:
201  float p_score;
202  WQuestion p_question;
203  EST_String p_token;
204  int p_freq;
205  int p_samples;
206  WDlist *next;
207  public:
208  WDlist() { next=0; }
209  ~WDlist() { if (next != 0) delete next; }
210  void set_score(float s) { p_score = s; }
211  void set_question(const WQuestion &q) { p_question = q; }
212  void set_best(const EST_String &t,int freq, int samples)
213  { p_token = t; p_freq = freq; p_samples = samples;}
214  float score() const {return p_score;}
215  const EST_String &token(void) const {return p_token;}
216  const WQuestion &question() const {return p_question;}
217  EST_Val predict(const WVector &w);
218  friend WDlist *add_to_dlist(WDlist *l,WDlist *a);
219  friend ostream &operator<<(ostream &s, WDlist &d);
220 };
221 
222 class WNode {
223  private:
224  WVectorVector data;
225  WQuestion question;
226  WImpurity impurity;
227  WNode *left;
228  WNode *right;
229  void print_out(ostream &s, int margin);
230  int leaf(void) const { return ((left == 0) || (right == 0)); }
231  int pure(void);
232  public:
233  WNode() { left = right = 0; }
234  ~WNode() { if (left != 0) {delete left; left=0;}
235  if (right != 0) {delete right; right=0;} }
236  WVectorVector &get_data(void) { return data; }
237  void set_subnodes(WNode *l,WNode *r) { left=l; right=r; }
238  void set_impurity(const WImpurity &imp) {impurity=imp;}
239  void set_question(const WQuestion &q) {question=q;}
240  void prune(void);
241  void held_out_prune(void);
242  WImpurity &get_impurity(void) {return impurity;}
243  WQuestion &get_question(void) {return question;}
244  EST_Val predict(const WVector &w);
245  WNode *predict_node(const WVector &d);
246  int samples(void) const { return data.n(); }
247  friend ostream& operator<<(ostream &s, WNode &n);
248 };
249 
250 extern Discretes wgn_discretes;
251 extern WDataSet wgn_dataset;
252 extern WDataSet wgn_test_dataset;
253 extern EST_FMatrix wgn_DistMatrix;
254 extern EST_Track wgn_VertexTrack;
255 extern EST_Track wgn_UnitTrack;
256 extern EST_Track wgn_VertexFeats;
257 
258 void wgn_load_datadescription(EST_String fname,LISP ignores);
259 void wgn_load_dataset(WDataSet &ds,EST_String fname);
260 WNode *wgn_build_tree(float &score);
261 WNode *wgn_build_dlist(float &score,ostream *output);
262 WNode *wagon_stepwise(float limit);
263 float wgn_score_question(WQuestion &q, WVectorVector &ds);
264 void wgn_find_split(WQuestion &q,WVectorVector &ds,
266 float summary_results(WNode &tree,ostream *output);
267 
268 extern int wgn_min_cluster_size;
269 extern int wgn_max_questions;
270 extern int wgn_held_out;
271 extern float wgn_dropout_feats;
272 extern float wgn_dropout_samples;
273 extern int wgn_cos;
274 extern int wgn_prune;
275 extern int wgn_quiet;
276 extern int wgn_verbose;
277 extern int wgn_predictee;
278 extern int wgn_count_field;
279 extern EST_String wgn_count_field_name;
280 extern EST_String wgn_predictee_name;
281 extern float wgn_float_range_split;
282 extern float wgn_balance;
283 extern EST_String wgn_opt_param;
284 extern EST_String wgn_vertex_output;
285 
286 #define wgn_ques_feature(X) (get_c_string(car(X)))
287 #define wgn_ques_oper_str(X) (get_c_string(car(cdr(X))))
288 #define wgn_ques_operand(X) (car(cdr(cdr(X))))
289 
290 int wagon_ask_question(LISP question, LISP value);
291 
292 int stepwise_ols(const EST_FMatrix &X,
293  const EST_FMatrix &Y,
294  const EST_StrList &feat_names,
295  float limit,
296  EST_FMatrix &coeffs,
297  const EST_FMatrix &Xtest,
298  const EST_FMatrix &Ytest,
299  EST_IVector &included,
300  float &best_score);
301 int robust_ols(const EST_FMatrix &X,
302  const EST_FMatrix &Y,
303  EST_IVector &included,
304  EST_FMatrix &coeffs);
305 int ols_apply(const EST_FMatrix &samples,
306  const EST_FMatrix &coeffs,
307  EST_FMatrix &res);
308 int ols_test(const EST_FMatrix &real,
309  const EST_FMatrix &predicted,
310  float &correlation,
311  float &rmse);
312 
313 #endif /* __WAGON_H__ */
INLINE const float & a_no_check(int n) const
read-only const access operator: without bounds checking
Definition: EST_TVector.h:257
EST_FVector()
Default constructor.
Definition: EST_FMatrix.h:121
const float & a_check(int n) const
read-only const access operator: with bounds checking
Definition: EST_TVector.cc:249
EST_TSimpleVector & operator=(const EST_TSimpleVector< float > &s)
assignment operator
void reset(void)
reset internal values
INLINE int length() const
number of items in vector.
Definition: EST_TVector.h:252
INLINE int n() const
number of items in vector.
Definition: EST_TVector.h:254
friend ostream & operator<<(ostream &st, const EST_TVector< float > &m)
print out vector.
Definition: EST_TVector.h:313