Merge ssh://ssh.dcc.fc.up.pt:31064//home/vsc/yap

C/errors.c

@@ -632,7 +632,7 @@ void Yap_ThrowError__(const char *file, const char *function, int lineno,
     // fprintf(stderr, "warning: ");
     Yap_Error__(true, file, function, lineno, type, where, tmpbuf);
   } else {
-    Yap_Error__(true, file, function, lineno, type, where);
+    Yap_Error__(true, file, function, lineno, type, where, NULL);
   }
   if (LOCAL_RestartEnv && !LOCAL_delay) {
     Yap_RestartYap(5);

library/gensym.yap

@@ -28,21 +28,20 @@
 :- dynamic gensym_key/2.
 
 init_gensym(Key) :-
-    assert(gensym_key(Atom,0) ).
+    retractall(gensym_key(Key,_)),
+    assert(gensym_key(Key,0) ).
 
-gensym(Atom, New) :-
+gensym(Key, New) :-
-	retract(gensym_key(Atom,Id)), !,
+	retract(gensym_key(Key,Id)), !,
-	atomic_concat(Atom,Id,New),
+	atomic_concat(Key,Id,New),
 	NId is Id+1,
-	assert(gensym_key(Atom,NId)).
+	assert(gensym_key(Key,NId)).
 gensym(Atom, New) :-
-	atomic_concat(Atom,1,New),
+	atomic_concat(Atom,0,New),
-	assert(gensym_key(Atom,2)).
+	assert(gensym_key(Atom,1)).
 
 reset_gensym(Atom) :-
 	retract(gensym_key(Atom,_)).
 
 reset_gensym :-
 	retractall(gensym_key(_,_)).
-
-

packages/ProbLog/problog_lbfgs.yap

@@ -317,7 +317,7 @@ check_examples :-
 	 (user:example(ID,_,P,_), (\+ number(P); P>1 ; P<0))
 	->
 	 (
-	  format(user_error,'The training example ~q does not have a valid probability value (~q).~n',[ID,P]),
+	  format(user_error,'The trianing example ~q does not have a valid probability value (~q).~n',[ID,P]),
 	  throw(error(examples))
 	 ); true
 	),
@@ -422,40 +422,26 @@ do_learning_intern(Iterations,Epsilon) :-
 	%	ground_truth_difference,
 	gradient_descent,
 
-	problog_flag(log_frequency,Log_Frequency),
+	once(save_model),
-
-	(
-	 ( Log_Frequency>0, 0 =:= CurrentIteration mod Log_Frequency)
-	->
-	 once(save_model);
-	 true
-	),
-
 	update_values,
-
+	mse_trainingset,
 	(
 	 last_mse(Last_MSE)
 	->
-	 (
+
 	  retractall(last_mse(_)),
 	  logger_get_variable(mse_trainingset,Current_MSE),
 	  assertz(last_mse(Current_MSE)),
 	  !,
 	  MSE_Diff is abs(Last_MSE-Current_MSE)
-	 );  (
+	 ;
 	      logger_get_variable(mse_trainingset,Current_MSE),
 	      assertz(last_mse(Current_MSE)),
 	      MSE_Diff is Epsilon+1
-	     )
-	),
 	 
-	(
-	  retractall(values_correct),
-	  retractall(query_is_similar(_,_)),
-	  retractall(query_md5(_,_,_)),
-	  empty_bdd_directory,
-	  init_queries
 	),
+	init_queries,
+
 
 
 	!,
@@ -466,7 +452,8 @@ do_learning_intern(Iterations,Epsilon) :-
 
 
 
-	RemainingIterations is Iterations-1,
+	current_iteration(ThisCurrentIteration),
+	RemainingIterations is Iterations-ThisCurrentIteration,
 
 	(
 	 MSE_Diff>Epsilon
@@ -499,6 +486,9 @@ init_learning :-
 	succeeds_n_times(user:example(_,_,_,_),TrainingExampleCount),
 	assertz(example_count(TrainingExampleCount)),
 	format_learning(3,'~q training examples~n',[TrainingExampleCount]),
+	forall(tunable_fact(FactID,GroundTruth),
+	       set_fact_probability(FactID,0.5)
+	      ),
 
 
 
@@ -709,8 +699,8 @@ mse_trainingset :-
 	format_learning(2,'MSE_Training ',[]),
 	update_values,
 	findall(t(LogCurrentProb,SquaredError),
-		(user:training_example(QueryID,Query,TrueQueryProb,_Type),
+		(user:example(QueryID,Query,TrueQueryProb,_Type),
-		 once(update_query(QueryID,'+',probability)),
+%		 once(update_query(QueryID,'+',probability)),
 		 query_probability(QueryID,CurrentProb),
  		 format(Handle,'ex(~q,training,~q,~q,~10f,~10f).~n',[Iteration,QueryID,Query,TrueQueryProb,CurrentProb]),
 
@@ -814,7 +804,7 @@ sigmoid(T,Slope,Sig) :-
     Sig <== OUT.
 
 inv_sigmoid(T,Slope,InvSig) :-
-	InvSig <== -log(1/T-1)/Slope.
+	InvSig is -log(1/T-1)/Slope.
 
 
 %========================================================================
@@ -835,15 +825,30 @@ save_old_probabilities :-
 gradient_descent :-
     problog_flag(sigmoid_slope,Slope),
 %	current_iteration(Iteration),
-	findall(FactID,tunable_fact(FactID,GroundTruth),L), length(L,N),
+    findall(FactID,tunable_fact(FactID,_GroundTruth),L),
+    length(L,N),
 %	leash(0),trace,
     lbfgs_initialize(N,X,0,Solver),
-	forall(tunable_fact(FactID,GroundTruth),
+    forall(tunable_fact(FactID,_GroundTruth),
-	       (XZ is 0.0, X[FactID] <== XZ,sigmoid(XZ,Slope,Pr),set_fact_probability(FactID,Pr))),
+	   set_fact( FactID, Slope, X)
-	problog_flag(sigmoid_slope,Slope),
+	   ),
     lbfgs_run(Solver,_BestF),
     lbfgs_finalize(Solver).
 
+set_fact(FactID, Slope, X ) :-
+    get_fact_probability(FactID,Pr),
+    (Pr > 0.99
+	    ->
+		NPr = 0.99
+			;
+			Pr < 0.01
+			       ->
+				   NPr = 0.01 ;
+					   Pr = NPr ),
+    inv_sigmoid(NPr, Slope, XZ),
+    X[FactID] <== XZ.
+
+
 set_tunable(I,Slope,P) :-
     X <== P[I],
     sigmoid(X,Slope,Pr),
@@ -866,9 +871,7 @@ user:evaluate(LLH_Training_Queries, X,Grad,N,_,_) :-
 	full_example(QueryID,QueryProb,BDD),
 	   compute_grad(QueryID, BDD, QueryProb,Grad, Probs, Slope,LLs)
     ),
-    LLH_Training_QueriesSum <== sum(LLs),
+    LLH_Training_Queries <== sum(LLs).
-LLH_Training_Queries is LLH_Training_QueriesSum/TrainingExampleCount .
-%wrap(X, Grad, GradCount).
 
 full_example(QueryID,QueryProb,BDD) :-
     user:example(QueryID,_Query,QueryProb,_),
@@ -985,18 +988,21 @@ bind_maplist([Node-Pr|MapList], Slope, X) :-
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 % stop calculate gradient
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
-user:progress(FX,X,_G,X_Norm,G_Norm,Step,_N,Iteration,Ls,0) :-
+user:progress(FX,_X,_G,X_Norm,_G_Norm,_Step,_N,_Iteration,_Ls,-1) :-
+    FX < 0, !,
+    format('stopped on bad FX=~4f~n',[FX]).
+user:progress(FX,X,_G,X_Norm,G_Norm,Step,_N,_Iteration,Ls,0) :-
     problog_flag(sigmoid_slope,Slope),
     forall(tunable_fact(FactID,_GroundTruth), set_tunable(FactID,Slope,X)),
     current_iteration(CurrentIteration),
     retractall(current_iteration(_)),
     NextIteration is CurrentIteration+1,
     assertz(current_iteration(NextIteration)),
-    save_model,
     logger_set_variable(mse_trainingset, FX),
+    save_model,
     X0 <== X[0], sigmoid(X0,Slope,P0),
     X1 <== X[1], sigmoid(X1,Slope,P1),
-    format('~d. Iteration : (x0,x1)=(~4f,~4f)  f(X)=~4f  |X|=~4f  |X\'|=~4f  Step=~4f  Ls=~4f~n',[Iteration,P0                                               ,P1,FX,X_Norm,G_Norm,Step,Ls]).
+    format('~d. Iteration : (x0,x1)=(~4f,~4f)  f(X)=~4f  |X|=~4f  |X\'|=~4f  Step=~4f  Ls=~4f~n',[CurrentIteration,P0                                               ,P1,FX,X_Norm,G_Norm,Step,Ls]).
 
 
 %========================================================================