Edinburgh Speech Tools  2.4-release
 All Classes Functions Variables Typedefs Enumerations Enumerator Friends Pages
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 #define wagon_error(WMESS) (cerr << WMESS << endl,exit(-1))
51 
52 // I get floating point exceptions of Alphas when I do any comparisons
53 // with HUGE_VAL or FLT_MAX so I'll make my own
54 #define WGN_HUGE_VAL 1.0e20
55 
56 class WVector : public EST_FVector
57 {
58  public:
59  WVector(int n) : EST_FVector(n) {}
60  int get_int_val(int n) const { return (int)a_no_check(n); }
61  float get_flt_val(int n) const { return a_no_check(n); }
62  void set_int_val(int n,int i) { a_check(n) = (int)i; }
63  void set_flt_val(int n,float f) { a_check(n) = f; }
64 };
65 
66 typedef EST_TList<WVector *> WVectorList;
67 typedef EST_TVector<WVector *> WVectorVector;
68 
69 /* Different types of feature */
70 enum wn_dtype {/* for predictees and predictors */
71  wndt_binary, wndt_float, wndt_class,
72  /* for predictees only */
73  wndt_cluster, wndt_vector, wndt_matrix, wndt_trajectory,
74  wndt_ols,
75  /* for ignored features */
76  wndt_ignore};
77 
78 class WDataSet : public WVectorList {
79  private:
80  int dlength;
81  EST_IVector p_type;
82  EST_IVector p_ignore;
83  EST_StrVector p_name;
84  public:
85  void load_description(const EST_String& descfname,LISP ignores);
86  void ignore_non_numbers();
87 
88  int ftype(const int &i) const {return p_type(i);}
89  int ignore(int i) const {return p_ignore(i); }
90  void set_ignore(int i,int value) { p_ignore[i] = value; }
91  const EST_String &feat_name(const int &i) const {return p_name(i);}
92  int samples(void) const {return length();}
93  int width(void) const {return dlength;}
94 };
95 enum wn_oper {wnop_equal, wnop_binary, wnop_greaterthan,
96  wnop_lessthan, wnop_is, wnop_in, wnop_matches};
97 
98 class WQuestion {
99  private:
100  int feature_pos;
101  wn_oper op;
102  int yes;
103  int no;
104  EST_Val operand1;
105  EST_IList operandl;
106  float score;
107  public:
108  WQuestion() {;}
109  WQuestion(const WQuestion &s)
110  { feature_pos=s.feature_pos;
111  op=s.op; yes=s.yes; no=s.no; operand1=s.operand1;
112  operandl = s.operandl; score=s.score;}
113  ~WQuestion() {;}
114  WQuestion(int fp, wn_oper o,EST_Val a)
115  { feature_pos=fp; op=o; operand1=a; }
116  void set_fp(const int &fp) {feature_pos=fp;}
117  void set_oper(const wn_oper &o) {op=o;}
118  void set_operand1(const EST_Val &a) {operand1 = a;}
119  void set_yes(const int &y) {yes=y;}
120  void set_no(const int &n) {no=n;}
121  int get_yes(void) const {return yes;}
122  int get_no(void) const {return no;}
123  const int get_fp(void) const {return feature_pos;}
124  const wn_oper get_op(void) const {return op;}
125  const EST_Val get_operand1(void) const {return operand1;}
126  const EST_IList &get_operandl(void) const {return operandl;}
127  const float get_score(void) const {return score;}
128  void set_score(const float &f) {score=f;}
129  const int ask(const WVector &w) const;
130  friend ostream& operator<<(ostream& s, const WQuestion &q);
131 };
132 
133 enum wnim_type {wnim_unset, wnim_float, wnim_class,
134  wnim_cluster, wnim_vector, wnim_matrix, wnim_ols,
135  wnim_trajectory};
136 
137 // Impurity measure for cumulating impurities from set of data
138 class WImpurity {
139  private:
140  wnim_type t;
141  EST_SuffStats a;
142  EST_DiscreteProbDistribution p;
143 
144  float cluster_impurity();
145  float cluster_member_mean(int i);
146  float vector_impurity();
147  float trajectory_impurity();
148  float ols_impurity();
149  public:
150  EST_IList members; // Maybe there should be a cluster class
151  EST_FList member_counts; // AUP: Implement counts for vectors
152  EST_SuffStats **trajectory;
153  const WVectorVector *data; // Needed for ols
154  float score;
155  int l,width;
156 
157  WImpurity() { t=wnim_unset; a.reset(); trajectory=0; l=0; width=0; data=0;}
158  ~WImpurity();
159  WImpurity(const WVectorVector &ds);
160  void copy(const WImpurity &s)
161  {
162  int i,j;
163  t=s.t; a=s.a; p=s.p; members=s.members; member_counts = s.member_counts; l=s.l; width=s.width;
164  score = s.score;
165  data = s.data;
166  if (s.trajectory)
167  {
168  trajectory = new EST_SuffStats *[l];
169  for (i=0; i<l; i++)
170  {
171  trajectory[i] = new EST_SuffStats[width];
172  for (j=0; j<width; j++)
173  trajectory[i][j] = s.trajectory[i][j];
174  }
175  }
176  }
177  WImpurity &operator = (const WImpurity &a) { copy(a); return *this; }
178 
179  float measure(void);
180  double samples(void);
181  wnim_type type(void) const { return t;}
182  void cumulate(const float pv,double count=1.0);
183  EST_Val value(void);
184  EST_DiscreteProbDistribution &pd() { return p; }
185  float cluster_distance(int i); // distance i from centre in sds
186  int in_cluster(int i); // distance i from centre < most remote member
187  float cluster_ranking(int i); // position in closeness to centre
188  friend ostream& operator<<(ostream &s, WImpurity &imp);
189 };
190 
191 class WDlist {
192  private:
193  float p_score;
194  WQuestion p_question;
195  EST_String p_token;
196  int p_freq;
197  int p_samples;
198  WDlist *next;
199  public:
200  WDlist() { next=0; }
201  ~WDlist() { if (next != 0) delete next; }
202  void set_score(float s) { p_score = s; }
203  void set_question(const WQuestion &q) { p_question = q; }
204  void set_best(const EST_String &t,int freq, int samples)
205  { p_token = t; p_freq = freq; p_samples = samples;}
206  float score() const {return p_score;}
207  const EST_String &token(void) const {return p_token;}
208  const WQuestion &question() const {return p_question;}
209  EST_Val predict(const WVector &w);
210  friend WDlist *add_to_dlist(WDlist *l,WDlist *a);
211  friend ostream &operator<<(ostream &s, WDlist &d);
212 };
213 
214 class WNode {
215  private:
216  WVectorVector data;
217  WQuestion question;
218  WImpurity impurity;
219  WNode *left;
220  WNode *right;
221  void print_out(ostream &s, int margin);
222  int leaf(void) const { return ((left == 0) || (right == 0)); }
223  int pure(void);
224  public:
225  WNode() { left = right = 0; }
226  ~WNode() { if (left != 0) {delete left; left=0;}
227  if (right != 0) {delete right; right=0;} }
228  WVectorVector &get_data(void) { return data; }
229  void set_subnodes(WNode *l,WNode *r) { left=l; right=r; }
230  void set_impurity(const WImpurity &imp) {impurity=imp;}
231  void set_question(const WQuestion &q) {question=q;}
232  void prune(void);
233  void held_out_prune(void);
234  WImpurity &get_impurity(void) {return impurity;}
235  WQuestion &get_question(void) {return question;}
236  EST_Val predict(const WVector &w);
237  WNode *predict_node(const WVector &d);
238  int samples(void) const { return data.n(); }
239  friend ostream& operator<<(ostream &s, WNode &n);
240 };
241 
242 extern Discretes wgn_discretes;
243 extern WDataSet wgn_dataset;
244 extern WDataSet wgn_test_dataset;
245 extern EST_FMatrix wgn_DistMatrix;
246 extern EST_Track wgn_VertexTrack;
247 extern EST_Track wgn_UnitTrack;
248 extern EST_Track wgn_VertexFeats;
249 
250 void wgn_load_datadescription(EST_String fname,LISP ignores);
251 void wgn_load_dataset(WDataSet &ds,EST_String fname);
252 WNode *wgn_build_tree(float &score);
253 WNode *wgn_build_dlist(float &score,ostream *output);
254 WNode *wagon_stepwise(float limit);
255 float wgn_score_question(WQuestion &q, WVectorVector &ds);
256 void wgn_find_split(WQuestion &q,WVectorVector &ds,
257  WVectorVector &y,WVectorVector &n);
258 float summary_results(WNode &tree,ostream *output);
259 
260 extern int wgn_min_cluster_size;
261 extern int wgn_held_out;
262 extern int wgn_prune;
263 extern int wgn_quiet;
264 extern int wgn_verbose;
265 extern int wgn_predictee;
266 extern int wgn_count_field;
267 extern EST_String wgn_count_field_name;
268 extern EST_String wgn_predictee_name;
269 extern float wgn_float_range_split;
270 extern float wgn_balance;
271 extern EST_String wgn_opt_param;
272 extern EST_String wgn_vertex_output;
273 
274 #define wgn_ques_feature(X) (get_c_string(car(X)))
275 #define wgn_ques_oper_str(X) (get_c_string(car(cdr(X))))
276 #define wgn_ques_operand(X) (car(cdr(cdr(X))))
277 
278 int wagon_ask_question(LISP question, LISP value);
279 
280 int stepwise_ols(const EST_FMatrix &X,
281  const EST_FMatrix &Y,
282  const EST_StrList &feat_names,
283  float limit,
284  EST_FMatrix &coeffs,
285  const EST_FMatrix &Xtest,
286  const EST_FMatrix &Ytest,
287  EST_IVector &included,
288  float &best_score);
289 int robust_ols(const EST_FMatrix &X,
290  const EST_FMatrix &Y,
291  EST_IVector &included,
292  EST_FMatrix &coeffs);
293 int ols_apply(const EST_FMatrix &samples,
294  const EST_FMatrix &coeffs,
295  EST_FMatrix &res);
296 int ols_test(const EST_FMatrix &real,
297  const EST_FMatrix &predicted,
298  float &correlation,
299  float &rmse);
300 
301 #endif /* __WAGON_H__ */