%%% -*- Mode: Prolog; -*- %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % % $Date: 2011-04-21 14:18:59 +0200 (Thu, 21 Apr 2011) $ % $Revision: 6364 $ % % This file is part of ProbLog % http://dtai.cs.kuleuven.be/problog % % ProbLog was developed at Katholieke Universiteit Leuven % % Copyright 2008, 2009, 2010 % Katholieke Universiteit Leuven % % Main authors of this file: % Bernd Gutmann, Vitor Santos Costa % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % % Artistic License 2.0 % % Copyright (c) 2000-2006, The Perl Foundation. % % Everyone is permitted to copy and distribute verbatim copies of this % license document, but changing it is not allowed. Preamble % % This license establishes the terms under which a given free software % Package may be copied, modified, distributed, and/or % redistributed. The intent is that the Copyright Holder maintains some % artistic control over the development of that Package while still % keeping the Package available as open source and free software. % % You are always permitted to make arrangements wholly outside of this % license directly with the Copyright Holder of a given Package. If the % terms of this license do not permit the full use that you propose to % make of the Package, you should contact the Copyright Holder and seek % a different licensing arrangement. Definitions % % "Copyright Holder" means the individual(s) or organization(s) named in % the copyright notice for the entire Package. % % "Contributor" means any party that has contributed code or other % material to the Package, in accordance with the Copyright Holder's % procedures. % % "You" and "your" means any person who would like to copy, distribute, % or modify the Package. % % "Package" means the collection of files distributed by the Copyright % Holder, and derivatives of that collection and/or of those files. A % given Package may consist of either the Standard Version, or a % Modified Version. % % "Distribute" means providing a copy of the Package or making it % accessible to anyone else, or in the case of a company or % organization, to others outside of your company or organization. % % "Distributor Fee" means any fee that you charge for Distributing this % Package or providing support for this Package to another party. It % does not mean licensing fees. % % "Standard Version" refers to the Package if it has not been modified, % or has been modified only in ways explicitly requested by the % Copyright Holder. % % "Modified Version" means the Package, if it has been changed, and such % changes were not explicitly requested by the Copyright Holder. % % "Original License" means this Artistic License as Distributed with the % Standard Version of the Package, in its current version or as it may % be modified by The Perl Foundation in the future. % % "Source" form means the source code, documentation source, and % configuration files for the Package. % % "Compiled" form means the compiled bytecode, object code, binary, or % any other form resulting from mechanical transformation or translation % of the Source form. % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % % Permission for Use and Modification Without Distribution % % (1) You are permitted to use the Standard Version and create and use % Modified Versions for any purpose without restriction, provided that % you do not Distribute the Modified Version. % % Permissions for Redistribution of the Standard Version % % (2) You may Distribute verbatim copies of the Source form of the % Standard Version of this Package in any medium without restriction, % either gratis or for a Distributor Fee, provided that you duplicate % all of the original copyright notices and associated disclaimers. At % your discretion, such verbatim copies may or may not include a % Compiled form of the Package. % % (3) You may apply any bug fixes, portability changes, and other % modifications made available from the Copyright Holder. The resulting % Package will still be considered the Standard Version, and as such % will be subject to the Original License. % % Distribution of Modified Versions of the Package as Source % % (4) You may Distribute your Modified Version as Source (either gratis % or for a Distributor Fee, and with or without a Compiled form of the % Modified Version) provided that you clearly document how it differs % from the Standard Version, including, but not limited to, documenting % any non-standard features, executables, or modules, and provided that % you do at least ONE of the following: % % (a) make the Modified Version available to the Copyright Holder of the % Standard Version, under the Original License, so that the Copyright % Holder may include your modifications in the Standard Version. (b) % ensure that installation of your Modified Version does not prevent the % user installing or running the Standard Version. In addition, the % modified Version must bear a name that is different from the name of % the Standard Version. (c) allow anyone who receives a copy of the % Modified Version to make the Source form of the Modified Version % available to others under (i) the Original License or (ii) a license % that permits the licensee to freely copy, modify and redistribute the % Modified Version using the same licensing terms that apply to the copy % that the licensee received, and requires that the Source form of the % Modified Version, and of any works derived from it, be made freely % available in that license fees are prohibited but Distributor Fees are % allowed. % % Distribution of Compiled Forms of the Standard Version or % Modified Versions without the Source % % (5) You may Distribute Compiled forms of the Standard Version without % the Source, provided that you include complete instructions on how to % get the Source of the Standard Version. Such instructions must be % valid at the time of your distribution. If these instructions, at any % time while you are carrying out such distribution, become invalid, you % must provide new instructions on demand or cease further % distribution. If you provide valid instructions or cease distribution % within thirty days after you become aware that the instructions are % invalid, then you do not forfeit any of your rights under this % license. % % (6) You may Distribute a Modified Version in Compiled form without the % Source, provided that you comply with Section 4 with respect to the % Source of the Modified Version. % % Aggregating or Linking the Package % % (7) You may aggregate the Package (either the Standard Version or % Modified Version) with other packages and Distribute the resulting % aggregation provided that you do not charge a licensing fee for the % Package. Distributor Fees are permitted, and licensing fees for other % components in the aggregation are permitted. The terms of this license % apply to the use and Distribution of the Standard or Modified Versions % as included in the aggregation. % % (8) You are permitted to link Modified and Standard Versions with % other works, to embed the Package in a larger work of your own, or to % build stand-alone binary or bytecode versions of applications that % include the Package, and Distribute the result without restriction, % provided the result does not expose a direct interface to the Package. % % Items That are Not Considered Part of a Modified Version % % (9) Works (including, but not limited to, modules and scripts) that % merely extend or make use of the Package, do not, by themselves, cause % the Package to be a Modified Version. In addition, such works are not % considered parts of the Package itself, and are not subject to the % terms of this license. % % General Provisions % % (10) Any use, modification, and distribution of the Standard or % Modified Versions is governed by this Artistic License. By using, % modifying or distributing the Package, you accept this license. Do not % use, modify, or distribute the Package, if you do not accept this % license. % % (11) If your Modified Version has been derived from a Modified Version % made by someone other than you, you are nevertheless required to % ensure that your Modified Version complies with the requirements of % this license. % % (12) This license does not grant you the right to use any trademark, % service mark, tradename, or logo of the Copyright Holder. % % (13) This license includes the non-exclusive, worldwide, % free-of-charge patent license to make, have made, use, offer to sell, % sell, import and otherwise transfer the Package with respect to any % patent claims licensable by the Copyright Holder that are necessarily % infringed by the Package. If you institute patent litigation % (including a cross-claim or counterclaim) against any party alleging % that the Package constitutes direct or contributory patent % infringement, then this Artistic License to you shall terminate on the % date that such litigation is filed. % % (14) Disclaimer of Warranty: THE PACKAGE IS PROVIDED BY THE COPYRIGHT % HOLDER AND CONTRIBUTORS "AS IS' AND WITHOUT ANY EXPRESS OR IMPLIED % WARRANTIES. THE IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A % PARTICULAR PURPOSE, OR NON-INFRINGEMENT ARE DISCLAIMED TO THE EXTENT % PERMITTED BY YOUR LOCAL LAW. UNLESS REQUIRED BY LAW, NO COPYRIGHT % HOLDER OR CONTRIBUTOR WILL BE LIABLE FOR ANY DIRECT, INDIRECT, % INCIDENTAL, OR CONSEQUENTIAL DAMAGES ARISING IN ANY WAY OUT OF THE USE % OF THE PACKAGE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% :- module(learning,[do_learning/1, do_learning/2, reset_learning/0, sigmoid/3, inv_sigmoid/3 ]). % switch on all the checks to reduce bug searching time :- style_check(all). :- yap_flag(unknown,error). % load modules from the YAP library :- use_module(library(lists), [member/2,max_list/2, min_list/2, sum_list/2, reverse/2]). :- use_module(library(system), [file_exists/1, shell/2]). :- use_module(library(rbtrees)). :- use_module(library(lbfgs)). :- reexport(library(matrix)). % load our own modules :- reexport(problog). :- use_module('problog/logger'). :- use_module('problog/flags'). :- use_module('problog/os'). :- use_module('problog/print_learning'). :- use_module('problog/utils_lbdd'). :- use_module('problog/utils'). :- use_module('problog/tabling'). % used to indicate the state of the system :- dynamic(values_correct/0). :- dynamic(learning_initialized/0). :- dynamic(current_iteration/1). :- dynamic(example_count/1). :- dynamic(query_probability_intern/2). %:- dynamic(query_gradient_intern/4). :- dynamic(last_mse/1). :- dynamic(query_is_similar/2). :- dynamic(query_md5/2). % used to identify queries which have identical proofs :- dynamic(query_is_similar/2). :- dynamic(query_md5/3). % used to identify queries which have identical proofs :- dynamic(query_is_similar/2). :- dynamic(query_md5/3). :- multifile(user:example/4). :- multifile(user:problog_discard_example/1). user:example(A,B,C,=) :- current_predicate(user:example/3), user:example(A,B,C), \+ user:problog_discard_example(B). :- multifile(user:test_example/4). user:test_example(A,B,C,=) :- current_predicate(user:test_example/3), user:test_example(A,B,C), \+ user:problog_discard_example(B). %======================================================================== %= store the facts with the learned probabilities to a file %======================================================================== save_model:- current_iteration(Iteration), create_factprobs_file_name(Iteration,Filename), export_facts(Filename). %======================================================================== %= find out whether some example IDs are used more than once %= if so, complain and stop %= %======================================================================== check_examples :- %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Check example IDs %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% ( (user:example(ID,_,_,_), \+ atomic(ID)) -> ( format(user_error,'The example id of training example ~q ',[ID]), format(user_error,'is not atomic (e.g foo42, 23, bar, ...).~n',[]), throw(error(examples)) ); true ), ( (user:test_example(ID,_,_,_), \+ atomic(ID)) -> ( format(user_error,'The example id of test example ~q ',[ID]), format(user_error,'is not atomic (e.g foo42, 23, bar, ...).~n',[]), throw(error(examples)) ); true ), %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Check example probabilities %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% ( (user:example(ID,_,P,_), (\+ number(P); P>1 ; P<0)) -> ( format(user_error,'The trianing example ~q does not have a valid probability value (~q).~n',[ID,P]), throw(error(examples)) ); true ), ( (user:test_example(ID,_,P,_), (\+ number(P); P>1 ; P<0)) -> ( format(user_error,'The test example ~q does not have a valid probability value (~q).~n',[ID,P]), throw(error(examples)) ); true ), %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Check that no example ID is repeated, % and if it is repeated make sure the query is the same %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% ( ( ( user:example(ID,QueryA,_,_), user:example(ID,QueryB,_,_), QueryA \= QueryB ) ; ( user:test_example(ID,QueryA,_,_), user:test_example(ID,QueryB,_,_), QueryA \= QueryB ); ( user:example(ID,QueryA,_,_), user:test_example(ID,QueryB,_,_), QueryA \= QueryB ) ) -> ( format(user_error,'The example id ~q is used several times.~n',[ID]), throw(error(examples)) ); true ). %======================================================================== %= %======================================================================== reset_learning :- retractall(learning_initialized), retractall(values_correct), retractall(current_iteration(_)), retractall(example_count(_)), retractall(query_probability_intern(_,_)), % retractall(query_gradient_intern(_,_,_,_)), retractall(last_mse(_)), retractall(query_is_similar(_,_)), retractall(query_md5(_,_,_)), set_problog_flag(alpha,auto), set_problog_flag(learning_rate,examples), logger_reset_all_variables. %======================================================================== %= initialize everything and perform Iterations times gradient descent %= can be called several times %= if it is called with an epsilon parameter, it stops when the change %= in the MSE is smaller than epsilon %======================================================================== do_learning(Iterations) :- do_learning(Iterations,-1). do_learning(Iterations,Epsilon) :- current_predicate(user:example/4), !, integer(Iterations), number(Epsilon), Iterations>0, do_learning_intern(Iterations,Epsilon). do_learning(_,_) :- format(user_error,'~n~Error: No training examples specified.~n~n',[]). do_learning_intern(0,_) :- !. do_learning_intern(Iterations,Epsilon) :- Iterations>0, init_learning, current_iteration(CurrentIteration), retractall(current_iteration(_)), NextIteration is CurrentIteration+1, assertz(current_iteration(NextIteration)), EndIteration is CurrentIteration+Iterations-1, format_learning(1,'~nIteration ~d of ~d~n',[CurrentIteration,EndIteration]), logger_set_variable(iteration,CurrentIteration), logger_start_timer(duration), % mse_testset, % ground_truth_difference, %leash(0),trace, gradient_descent, once(save_model), 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 ), init_queries, !, logger_stop_timer(duration), logger_write_data, current_iteration(ThisCurrentIteration), RemainingIterations is Iterations-ThisCurrentIteration, ( MSE_Diff>Epsilon -> do_learning_intern(RemainingIterations,Epsilon); true ). %======================================================================== %= find proofs and build bdds for all training and test examples %= %= %======================================================================== init_learning :- learning_initialized, !. init_learning :- check_examples, retractall(current_iteration(_)), assert(current_iteration(0)), % empty_output_directory, logger_write_header, format_learning(1,'Initializing everything~n',[]), succeeds_n_times(user:test_example(_,_,_,_),TestExampleCount), format_learning(3,'~q test examples~n',[TestExampleCount]), succeeds_n_times(user:example(_,_,_,_),TrainingExampleCount), assertz(example_count(TrainingExampleCount)), format_learning(3,'~q training examples~n',[TrainingExampleCount]), %current_probs <== array[TrainingExampleCount ] of floats, %current_lls <== array[TrainingExampleCount ] of floats, forall(tunable_fact(FactID,_GroundTruth), set_fact_probability(FactID,0.5) ), %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % build BDD script for every example %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% once(init_queries), %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % done %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% assertz(current_iteration(-1)), assertz(learning_initialized), format_learning(1,'~n',[]). %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Check, if continuous facts are used. % if yes, switch to problog_exact % continuous facts are not supported yet. %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% set_default_gradient_method :- ( problog_flag(continuous_facts, true ) -> problog_flag(init_method,(_,_,_,_,OldCall)), ( ( continuous_fact(_), OldCall\=problog_exact_save(_,_,_,_,_) ) -> ( format_learning(2,'Theory uses continuous facts.~nWill use problog_exact/3 as initalization method.~2n',[]), set_problog_flag(init_method,(Query,Probability,BDDFile,ProbFile,problog_exact_save(Query,Probability,_Status,BDDFile,ProbFile))) ); true ) ; problog_tabled(_) -> ( format_learning(2,'Theory uses tabling.~nWill use problog_exact/3 as initalization method.~2n',[]), set_problog_flag(init_method,(Query,Probability,BDDFile,ProbFile,problog_exact_save(Query,Probability,_Status,BDDFile,ProbFile))) ); true ). empty_bdd_directory :- current_key(_,I), integer(I), recorded(I,bdd(_,_,_),R), erase(R), fail. empty_bdd_directory. %======================================================================== %= This predicate goes over all training and test examples, %= calls the inference method of ProbLog and stores the resulting %= BDDs %======================================================================== init_queries :- empty_bdd_directory, format_learning(2,'Build BDDs for examples~n',[]), forall(user:test_example(ID,Query,_Prob,_),init_one_query(ID,Query,test)), forall(user:example(ID,Query,_Prob,_),init_one_query(ID,Query,training)). bdd_input_file(Filename) :- problog_flag(output_directory,Dir), concat_path_with_filename(Dir,'input.txt',Filename). init_one_query(QueryID,Query,_Type) :- %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % if BDD file does not exist, call ProbLog %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% fail, problog_flag(init_method,(Query,N,Bdd,user:graph2bdd(Query,N,Bdd))), !, b_setval(problog_required_keep_ground_ids,false), (QueryID mod 100 =:= 0 ->writeln(QueryID) ; true), Bdd = bdd(Dir, Tree0,MapList), user:graph2bdd(Query,N,Bdd), reverse(Tree0,Tree), %rb_new(H0), %maplist_to_hash(MapList, H0, Hash), %tree_to_grad(Tree, Hash, [], Grad), % ; % Bdd = bdd(-1,[],[]), % Grad=[] write('.'), recordz(QueryID,bdd(Dir, Tree, MapList),_). init_one_query(QueryID,Query,_Type) :- % format_learning(3,' ~q example ~q: ~q~n',[Type,QueryID,Query]), %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % if BDD file does not exist, call ProbLog %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% b_setval(problog_required_keep_ground_ids,false), problog_flag(init_method,(Query,_K,Bdd,Call)), !, Bdd = bdd(Dir, Tree0, MapList), % trace, once(Call), reverse(Tree0,Tree), %rb_new(H0), %maplist_to_hash(MapList, H0, Hash), %Tree \= [], % writeln(Dir:Tree:MapList), %tree_to_grad(Tree, Hash, [], Grads), recordz(QueryID,bdd(Dir, Tree, MapList),_). %======================================================================== %= %= %= %======================================================================== query_probability(QueryID,Prob) :- query_probability_intern(QueryID,Prob). %======================================================================== %= %= %= %======================================================================== % FIXME ground_truth_difference :- findall(Diff,(tunable_fact(FactID,GroundTruth), \+continuous_fact(FactID), \+ var(GroundTruth), %% get_fact_probability(FactID,Prob), Prob <== p[FactID], Diff is abs(GroundTruth-Prob)),AllDiffs), ( AllDiffs=[] -> ( MinDiff=0.0, MaxDiff=0.0, DiffMean=0.0 ) ; ( length(AllDiffs,Len), sum_list(AllDiffs,AllDiffsSum), min_list(AllDiffs,MinDiff), max_list(AllDiffs,MaxDiff), DiffMean is AllDiffsSum/Len ) ), logger_set_variable(ground_truth_diff,DiffMean), logger_set_variable(ground_truth_mindiff,MinDiff), logger_set_variable(ground_truth_maxdiff,MaxDiff). %======================================================================== %= Calculates the mse of training and test data %= %= -Float %======================================================================== mse_trainingset :- current_iteration(Iteration), create_training_predictions_file_name(Iteration,File_Name), open(File_Name, write,Handle), format_learning(2,'MSE_Training ',[]), update_values, findall(t(LogCurrentProb,SquaredError), (user:example(QueryID,Query,TrueQueryProb,_Type), query_probability(QueryID,CurrentProb), format(Handle,'ex(~q,training,~q,~q,~10f,~10f).~n',[Iteration,QueryID,Query,TrueQueryProb,CurrentProb]), once(update_query_cleanup(QueryID)), SquaredError is (CurrentProb-TrueQueryProb)**2, LogCurrentProb is log(CurrentProb) ), All), maplist(tuple, All, AllLogs, AllSquaredErrors), sum_list( AllLogs, LLH_Training_Queries), close(Handle), length(AllSquaredErrors,Length), ( Length>0 -> ( sum_list(AllSquaredErrors,SumAllSquaredErrors), min_list(AllSquaredErrors,MinError), max_list(AllSquaredErrors,MaxError), MSE is SumAllSquaredErrors/Length );( MSE=0.0, MinError=0.0, MaxError=0.0 ) ), logger_set_variable(mse_trainingset,MSE), logger_set_variable(mse_min_trainingset,MinError), logger_set_variable(mse_max_trainingset,MaxError), logger_set_variable(llh_training_queries,LLH_Training_Queries), %%%%% format(' (~8f)~n',[MSE]). format_learning(2,' (~8f)~n',[MSE]). tuple(t(X,Y),X,Y). mse_testset :- current_iteration(Iteration), create_test_predictions_file_name(Iteration,File_Name), open(File_Name, write,Handle), format_learning(2,'MSE_Test ',[]), update_values, bb_put(llh_test_queries,0.0), findall(SquaredError, (user:test_example(QueryID,Query,TrueQueryProb,Type), once(update_query(QueryID,'+',probability)), query_probability(QueryID,CurrentProb), format(Handle,'ex(~q,test,~q,~q,~10f,~10f).~n',[Iteration,QueryID,Query,TrueQueryProb,CurrentProb]), once(update_query_cleanup(QueryID)), ( (Type == '='; (Type == '<', CurrentProb>QueryProb); (Type=='>',CurrentProb SquaredError is (CurrentProb-TrueQueryProb)**2; SquaredError = 0.0 ), bb_get(llh_test_queries,Old_LLH_Test_Queries), New_LLH_Test_Queries is Old_LLH_Test_Queries+log(CurrentProb), bb_put(llh_test_queries,New_LLH_Test_Queries) ), AllSquaredErrors), close(Handle), bb_delete(llh_test_queries,LLH_Test_Queries), length(AllSquaredErrors,Length), ( Length>0 -> ( sum_list(AllSquaredErrors,SumAllSquaredErrors), min_list(AllSquaredErrors,MinError), max_list(AllSquaredErrors,MaxError), MSE is SumAllSquaredErrors/Length );( MSE=0.0, MinError=0.0, MaxError=0.0 ) ), logger_set_variable(mse_testset,MSE), logger_set_variable(mse_min_testset,MinError), logger_set_variable(mse_max_testset,MaxError), logger_set_variable(llh_test_queries,LLH_Test_Queries), format_learning(2,' (~8f)~n',[MSE]). %======================================================================== %= Calculates the sigmoid function respectivly the inverse of it %= warning: applying inv_sigmoid to 0.0 or 1.0 will yield +/-inf %= %= +Float, -Float %======================================================================== sigmoid(T,Slope,Sig) :- IN <== T, OUT is 1/(1+exp(-IN*Slope)), Sig <== OUT. inv_sigmoid(T,Slope,InvSig) :- InvSig is -log(1/T-1)/Slope. %======================================================================== %= Perform one iteration of gradient descent %= %= assumes that everything is initialized, if the current values %= of query_probability/2 and query_gradient/4 are not up to date %= they will be recalculated %= finally, the values_correct/0 is retracted to signal that the %= probabilities of the examples have to be recalculated %======================================================================== save_old_probabilities. % vsc: avoid silly search gradient_descent :- problog_flag(sigmoid_slope,Slope), % current_iteration(Iteration), findall(FactID,tunable_fact(FactID,_GroundTruth),L), length(L,N), lbfgs_initialize(N,X,0,Solver), forall(tunable_fact(FactID,_GroundTruth), set_fact( FactID, Slope, X) ), lbfgs_run(Solver,_BestF), lbfgs_finalize(Solver). set_fact(FactID, Slope, P ) :- X <== P[FactID], sigmoid(X, Slope, Pr), (Pr > 0.99 -> NPr = 0.99 ; Pr < 0.01 -> NPr = 0.01 ; Pr = NPr ), set_fact_probability(FactID, NPr). set_tunable(I,Slope,P) :- X <== P[I], sigmoid(X,Slope,Pr), set_fact_probability(I,Pr). :- include(problog/lbdd). %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % start calculate gradient %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% user:evaluate(LLH_Training_Queries, X,Grad,N,_,_) :- %Handle = user_error, N1 is N-1, forall(between(0,N1,I),(Grad[I]<==0.0)), go( X,Grad, LLs), sum_list( LLs, LLH_Training_Queries). test :- S =.. [f,0-0.9,1-0.8,2-0.6,3-0.7,4-0.5,5-0.4,6-0.7,7-0.2], functor(S,_,N), N1 is N-1, problog_flag(sigmoid_slope,Slope), X <== array[N] of floats, Grad <== array[N] of floats, forall(between(0,N1,I),(Grad[I]<==0.0)), forall(between(1,N,I),(arg(I,S,_-V),inv_sigmoid(V,Slope,V0),I1 is I-1,X[I1]<==V0)), findall( LL, compute_gradient(Grad, X, Slope,LL), LLs ), sum_list( LLs, LLH_Training_Queries), writeln(LLH_Training_Queries:LLs ),forall(between(0,N1,I),(G<==Grad[I],writeln(I=G))). go( X,Grad, LLs) :- problog_flag(sigmoid_slope,Slope), findall( LL, compute_gradient(Grad, X, Slope,LL), LLs ). compute_gradient( Grad, X, Slope, LL) :- user:example(QueryID,_Query,QueryProb), recorded(QueryID,BDD,_), BDD = bdd(_,_,MapList), bind_maplist(MapList, Slope, X), query_probabilities( BDD, BDDProb), LL is (BDDProb-QueryProb)*(BDDProb-QueryProb), retractall( query_probability_intern( QueryID, _) ), assert( query_probability_intern( QueryID,BDDProb )), forall( query_gradients(BDD,I,IProb,GradValue), gradient_pair(BDDProb, QueryProb, Grad, GradValue, I, IProb) ). gradient_pair(BDDProb, QueryProb, Grad, GradValue, I, Prob) :- G0 <== Grad[I], GN is G0-GradValue*Prob*(1-Prob)*2*(QueryProb-BDDProb), Grad[I] <== GN. wrap( X, Grad, GradCount) :- tunable_fact(FactID,GroundTruth), Z<==X[FactID], W<==Grad[FactID], WC<==GradCount[FactID], WC > 0, format('ex(~d, ~q, ~4f, ~4f).~n',[FactID,GroundTruth,Z,W]), % Grad[FactID] <== WN, fail. wrap( _X, _Grad, _GradCount). %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % stop calculate gradient %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% user:progress(FX,_X,_G, _X_Norm,_G_Norm,_Step,_N,_CurrentIteration,_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) :- assertz(current_iteration(Iteration)), problog_flag(sigmoid_slope,Slope), forall(tunable_fact(FactID,_GroundTruth), set_tunable(FactID,Slope,X)), 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]). %======================================================================== %= initialize the logger module and set the flags for learning %= don't change anything here! use set_problog_flag/2 instead %======================================================================== init_flags :- % prolog_file_name(queries,Queries_Folder), % get absolute file name for './queries' prolog_file_name(output,Output_Folder), % get absolute file name for './output' % problog_define_flag(bdd_directory, problog_flag_validate_directory, 'directory for BDD scripts', Queries_Folder,learning_general), problog_define_flag(output_directory, problog_flag_validate_directory, 'directory for logfiles etc', Output_Folder,learning_general,flags:learning_output_dir_handler), problog_define_flag(log_frequency, problog_flag_validate_posint, 'log results every nth iteration', 1, learning_general), % problog_define_flag(rebuild_bdds, problog_flag_validate_nonegint, 'rebuild BDDs every nth iteration', 0, learning_general), % problog_define_flag(reuse_initialized_bdds,problog_flag_validate_boolean, 'Reuse BDDs from previous runs',false, learning_general), % problog_define_flag(check_duplicate_bdds,problog_flag_validate_boolean,'Store intermediate results in hash table',true,learning_general), problog_define_flag(init_method,problog_flag_validate_dummy,'ProbLog predicate to search proofs',(Query,Tree,problog:problog_kbest_as_bdd(Query,100,Tree)),learning_general,flags:learning_libdd_init_handler), problog_define_flag(alpha,problog_flag_validate_number,'weight of negative examples (auto=n_p/n_n)',auto,learning_general,flags:auto_handler), problog_define_flag(sigmoid_slope,problog_flag_validate_posnumber,'slope of sigmoid function',1.0,learning_general), % problog_define_flag(continuous_facts,problog_flag_validate_boolean,'support parameter learning of continuous distributions',1.0,learning_general), problog_define_flag(learning_rate,problog_flag_validate_posnumber,'Default learning rate (If line_search=false)',examples,learning_line_search,flags:examples_handler), problog_define_flag(line_search, problog_flag_validate_boolean,'estimate learning rate by line search',false,learning_line_search), problog_define_flag(line_search_never_stop, problog_flag_validate_boolean,'make tiny step if line search returns 0',true,learning_line_search), problog_define_flag(line_search_tau, problog_flag_validate_indomain_0_1_open,'tau value for line search',0.618033988749,learning_line_search), problog_define_flag(line_search_tolerance,problog_flag_validate_posnumber,'tolerance value for line search',0.05,learning_line_search), problog_define_flag(line_search_interval, problog_flag_validate_dummy,'interval for line search',(0,100),learning_line_search,flags:linesearch_interval_handler). init_logger :- logger_define_variable(iteration, int), logger_define_variable(duration,time), logger_define_variable(mse_trainingset,float), logger_define_variable(mse_min_trainingset,float), logger_define_variable(mse_max_trainingset,float), logger_define_variable(mse_testset,float), logger_define_variable(mse_min_testset,float), logger_define_variable(mse_max_testset,float), logger_define_variable(gradient_mean,float), logger_define_variable(gradient_min,float), logger_define_variable(gradient_max,float), logger_define_variable(ground_truth_diff,float), logger_define_variable(ground_truth_mindiff,float), logger_define_variable(ground_truth_maxdiff,float), logger_define_variable(learning_rate,float), logger_define_variable(alpha,float), logger_define_variable(llh_training_queries,float), logger_define_variable(llh_test_queries,float). :- initialization(init_flags). :- initialization(init_logger).