shalmaneser-rosy 1.2.0.rc4

data/lib/rosy/RosyIterator.rb

@@ -0,0 +1,478 @@
+# RosyIterator
+# KE May 2005
+#
+# RosyIterator is a class that 
+# * reads the "xwise" parameters in the experiment file to 
+#   determine the portions in which data is to be fed to classifiers,
+#   and offers an iterator that iterates through every group to 
+#   be trained/tested on
+# * constructs views matching the given "xwise" group.
+#   
+# RosyIterator incorporates the following services:
+# - choosing the right DB table, depending on 
+#   whether training/test data is being accessed,
+#   and with or without a splitlog
+# - making and adding all currently available Dynamic Gold objects
+#   (i.e. objects that are capable of mapping the gold column to 
+#   something else)
+# - initializing a view, potentially modified depending on the assignment step:
+#   argrec -> use dynamic gold, mapping gold labels to "FE" or "NONE"
+#   arglab -> use only those rows that have "FE" assigned from the argrec step
+#
+# Setting "xwise": An "xwise" entry in the hash passed on to RosyIterator.new()
+# overrides all other settings. If that isn't given, the "xwise_" + step 
+# (xwise_argrec, xwise_arglab, xwise_onestep) from the experiment file is read.
+# If that hasn't been set either, the default is frame-wise.
+
+require 'common/ruby_class_extensions'
+
+require 'rosy/View'
+require "common/RosyConventions"
+require "rosy/RosyPruning"
+require "rosy/RosySplit"
+require "rosy/RosyTrainingTestTable"
+
+class RosyIterator
+
+  ###
+  # new
+  #
+  # open the correct database table, 
+  # initialize Dynamic Gold objects
+
+
+  def initialize(ttt_obj, # RosyTrainingTestTable object 
+		 exp,     # RosyConfigData object: experiment file
+		 dataset, # string: train/test
+		 var_hash = {}) # further arguments:
+    # step: string: argrec/arglab/onestep, or nil (= no manipulation of the view)
+    # testID: string: ID of test set, or nil
+    # splitID string: splitlog ID, or nil if no split is to be used
+    # xwise: string: containing any subset of frame/target_pos/target joined by spaces,
+    #                overrides @exp.get("xwise_" + @step) if non-nil
+    # prune: boolean: if pruning has been chosen in the experiment file,
+    #                make a value restriction that omits pruned instances
+
+    @exp = exp
+    @dataset = dataset
+    @ttt_obj = ttt_obj
+    @splitID = var_hash["splitID"]
+    @step = var_hash["step"]
+    @testID = var_hash["testID"]
+    
+    # object variables we are going to use below
+    @db_table = nil  # DB table we are working on
+    @allcolnames = nil   # names of all columns of first and potentially second table
+
+    @dyn_gold_objects = nil # list of dynamic gold-producing object
+    @standard_dyngold_id = nil # ID of standard dyngold obj to use
+
+    @standard_value_restrictions = [] #value restrictions to use with each view
+
+    @second_table = nil      # read view from 2 tables? if so, DBTable object for 2nd table
+    @use_cols_from_second_table = nil # array: names of columns from 2nd table
+    @second_table_colprefix = nil     # string: prefix for columns from 2nd table
+
+    @xwise = nil # array: read data one X at a time (forms groups)
+    @groups = nil # distinct values for X from xwise
+    @current_group = nil # current group (will be set by iterator each_group)
+
+    ##
+    # open the right database table
+    if @dataset == "train" or @splitID
+      @db_table = @ttt_obj.existing_train_table()
+
+    else
+      unless @testID
+	raise "cannot open the test table without test ID"
+      end
+      @db_table = @ttt_obj.existing_test_table(@testID)
+    end
+    @allcolnames = @db_table.list_column_names()
+
+    ##
+    # make dynamic gold objects
+    @dyn_gold_objects = Array.new
+    @dyn_gold_objects << DynGoldBinary.new(@exp.get("noval"))
+
+    ###
+    # what is the standard gold column to be returned?
+    if @step == "argrec"
+      # argument recognition: distinguish just "FE", "NONE" as gold
+      @standard_dyngold_id = "binary_gold"
+    end
+      
+    ##
+    # if splitID has been set, 
+    # make additional restrictions on the column values
+    if @splitID
+      # get split table name
+      @second_table = @ttt_obj.existing_split_table(@splitID, @dataset, RosySplit.split_index_colname())
+
+      # additional value restriction:
+      # only use rows whose sentence ID also appears in the split table
+      # (i.e. rows included in the split)
+      @standard_value_restrictions << RosySplit.make_join_restriction(@splitID, 
+                                                                      @db_table,
+                                                                      @dataset,
+                                                                      @ttt_obj)
+
+      # additional column names:
+      # those of the second table (but remove duplicates)
+      @allcolnames.concat @ttt_obj.existing_split_table(@splitID, @dataset, RosySplit.split_index_colname()).list_column_names()
+      @allcolnames.uniq!
+
+
+      # if we're using a split, read the phase 2 features and the classification results
+      # from the split table rather than from the main table:
+      # @use_cols_from_second_table is a list of column names (strings) 
+      #     to take from the 2nd table
+      # @second_table_colprefix is a string: all columns starting with this prefix
+      #     are taken from the 2nd table
+      @use_cols_from_second_table = [ RosySplit.split_index_colname() ]
+      @second_table_colprefix = @exp.get("classif_column_name")
+    end
+    
+    ###
+    # Any (row) value restrictions to be imposed
+    # on all views we generate?
+    if @step == "arglab"
+      # argument labeling: use as input only those lines
+      # for which argrec-label is "FE"
+
+      if @exp.get("assume_argrec_perfect")
+	# assume perfect argrec step:
+	# take all rows where gold is not "noval"
+	@standard_value_restrictions << ValueRestriction.new(@db_table.table_name + ".gold", 
+                                                             @exp.get("noval"), 
+                                                             "posneg" => "!=")
+      else
+	# use argrec step as is:
+	# take all rows where the argrec result is "FE"
+
+        case @dataset
+        when "train"
+          run_column_name = @ttt_obj.existing_runlog("argrec", "train", nil, @splitID)
+        when "test"
+          run_column_name = @ttt_obj.existing_runlog("argrec", "test", @testID, @splitID)
+        else
+          raise "Shouldn't be here"
+        end
+
+        if run_column_name.nil? 
+          $stderr.puts "Missing: argrec classification results on #{@dataset} data."
+          $stderr.puts "I have logs of the following runs: "
+          $stderr.puts @ttt_obj.runlog_to_s()
+          raise "Problem"
+        end
+
+        # run column where? split table, or the table we are mainly working with?
+        if @second_table
+          run_column_name = @second_table.table_name + "." + run_column_name
+        else
+          run_column_name = @db_table.table_name + "." + run_column_name
+        end
+
+	@standard_value_restrictions << ValueRestriction.new(run_column_name, "FE")
+      end
+    end
+
+    # pruning?
+    if var_hash["prune"] and    # pruning requested in RosyIterator initialization
+        ["argrec", "onestep"].include? @step and # pruning only affects argument recognition
+        Pruning.prune?(@exp)    # pruning has been set in the experiment file
+       @standard_value_restrictions << Pruning.restriction_removing_pruned(@exp)
+    end
+
+    ##
+    # access "xwise" information
+    # are we training by frame or by target POS or target lemma?
+
+    # xwise-value in var_hash overrides others
+    @xwise = var_hash["xwise"]
+    unless @xwise
+      if @step
+	# read xwise from experiment file,
+	# if we know what training/test step we're in
+	@xwise = @exp.get("xwise_" + @step)
+      end
+    end
+    if @xwise.nil?
+      # default: read one frame at a time
+      @xwise = "frame"
+    end
+
+    # xwise is a string consisting of any subset of 
+    # "frame", "target_pos", "target" joined by spaces.
+    # transform to an array by splitting at spaces
+    @xwise = @xwise.split()
+    @xwise.each { |xwise_entry|
+      unless @ttt_obj.feature_names.include? xwise_entry
+        # sanity check: valid xwise value?
+        raise "Unknown value for parameter 'xwise' in experiment file.\n" +
+              "Allowed: any subset of the list of features listed in the experiment file.\n" +
+              "This is the granularity of training and testing\n" +
+               "What I got was: " + @xwise.join(" ")
+      end
+    }
+
+    # list all frames/ all target POSs/all frame+target-pairs
+    @groups = unique_values_of_columns(@xwise)
+    @current_group = nil
+  end
+
+  ####
+  # get_xwise_column_names
+  #
+  # get the column names used for determining the groups
+  #
+  # returns: an array of strings, ["frame"] or ["frame", "target"], 
+  # or ["target_pos"]
+  def get_xwise_column_names()
+    return @xwise
+  end
+
+  ####
+  # num_groups
+  # returns: integer
+  def num_groups()
+    return @groups.length()
+  end
+
+  ####
+  # each_group
+  #
+  # iterates through the "xwise" groups, sets
+  # internal values such that get_a_view_for_current_group()
+  # will get you the correct view
+  #
+  # yields: for each group, a pair of
+  # - the hash describing the group, as returned by unique_values_of_column
+  # - plus an ID for the group, made up of its hash values concatenated into a string
+  #   (values are connected by spaces)
+  def each_group()
+    @groups.each { |hash|
+      # hash is a hash column_name(string)-> value(object)
+      # this is the unique description of the current group
+      @current_group = hash
+      yield [hash, hash.values.join(" ")]
+    }
+  end
+
+  ####
+  # get_a_view_for_current_group
+  #
+  # constructs a new View object
+  # matching the last yielded group (of each_group)
+  #
+  # you give it: the names of the columns to be included in the view
+  # (or "*" for all columns) and a list of value restrictions
+  # on the rows (ValueRestriction objects, equalities or inequalities
+  # column_name = value, columnb_name != value), which may be omitted
+  # 
+  # returns: DBView object
+  # @param columns [Array] array:string, column names to include
+  #   or string: "*" for all columns
+  # @param value_restrictions [Array] array:ValueRestriction objects
+  def get_a_view_for_current_group(columns, value_restrictions = []) 
+    get_a_view_for_group(@current_group, columns, value_restrictions)
+  end
+
+  ####
+  # get_a_view_for_group
+  #
+  # constructs a new View object
+  # matching the a group given by its row hash
+  # (as yielded by each_group)
+  #
+  # you give it: the group description hash,
+  # the names of the columns to be included in the view
+  # (or "*" for all columns) and a list of value restrictions
+  # on the rows (ValueRestriction objects, equalities or inequalities
+  # column_name = value, columnb_name != value), which may be omitted
+  # 
+  # returns: DBView object
+  # @param group [Hash] column(string)->value(object)
+  #   describing the group
+  # @param columns [Array] array:string, column names to include
+  #   or string: "*" for all columns
+  # @param value_restrictions [Array]  of ValueRestriction objects
+  def get_a_view_for_group(group, columns, value_restrictions = [])
+
+    # value_restrictions needs to be an array
+    if value_restrictions.nil?
+      value_restrictions = []
+    end
+
+    # we need to add value restrictions that say
+    # that the group column names need to have the values for
+    # the given group.
+    # however, group column names may belong to either the first or
+    # the second table
+
+    # separate group column names into two groups
+    first_columns, second_columns = 
+         separate_into_1st_and_2nd_table_cols(group.keys)
+
+    # make separate value restrictions for the two groups
+    value_restrictions = value_restrictions + first_columns.map {|column_name|
+      ValueRestriction.new(column_name, group[column_name])
+    }
+    if second_columns
+      unless @second_table
+        raise "Cannot use second table columns without second table"
+      end
+      value_restrictions.concat second_columns.map { |column_name|
+        ValueRestriction.new(@second_table.table_name + "." + column_name, 
+                             group[column_name],
+                             "table_name_included" => true)
+      }
+    end
+  
+    # get a view with the given columns, given value restrictions
+    # plus add more value restrictions: must be the current group
+    return get_a_view(columns,value_restrictions)
+  end
+
+
+
+  ####
+  # get_a_view
+  #
+  # construct a new View object,
+  #
+  # you give it: the names of the columns to be included in the view
+  # (or "*" for all columns) and a list of value restrictions
+  # on the rows (ValueRestriction objects, equalities or inequalities
+  # column_name = value, columnb_name != value), which may be omitted
+  # 
+  # returns: DBView object
+  def get_a_view(columns, # array:strings, list of column names
+		           # or string "*" (all columns)
+		 value_restrictions = []) # array: ValueRestriction objects 
+                           # or [], nil for no restrictions
+
+    if value_restrictions.nil?
+      value_restrictions = []
+    end
+    return get_a_view_aux(columns, value_restrictions, 
+                          "gold" => "gold", 
+                          "dynamic_feature_list" => @dyn_gold_objects,
+                          "standard_dyngold_id" => @standard_dyngold_id,
+                          "sentence_id_feature" => "sentid")
+  end
+
+  ####
+  # unique_values_of_columns
+  #
+  # construct a new View object
+  # for the given column and
+  # get all unique values for it
+  #
+  # returns: a list of hashes, one for each unique set of values
+  def unique_values_of_columns(columns) # array:string, several column names
+    retv = Array.new
+
+    view = get_a_view_aux(columns, [],
+                          "distinct" => true)
+
+    view.each_hash() { |row|
+      retv << row
+    }
+    view.close()
+    return retv
+  end
+
+  #############################################
+  private
+
+  ###
+  # given a list of column names, 
+  # separate them into first table and second table columns
+  #
+  # columns may be either an array of string (column names)
+  # or the string "*" for "all columns"
+  def separate_into_1st_and_2nd_table_cols(columns)
+
+    if @use_cols_from_second_table or @second_table_colprefix
+      # if there are columns I'm supposed to take from the second
+      # table rather than the first, let's do that
+      if columns == "*"
+        # we have simply been told to use all columns
+        columns = @allcolnames
+      end
+
+      # second table columns either start with @second_table_colprefix
+      # or are in the list @use_columns_from_second_table
+      second_columns, first_columns = columns.distribute { |colname|
+        (@second_table_colprefix and colname =~ /^#{@second_table_colprefix}/) or
+          (@use_cols_from_second_table and @use_cols_from_second_table.include?(colname))
+      }
+
+    else
+      # no columns to take from a 2nd table
+      first_columns = columns
+      second_columns = nil
+    end
+
+    return [first_columns, second_columns]
+  end
+
+  ###
+  # access DB table:
+  # figure out which table, set of columns from that table,
+  # set of columns from secondary table
+  #
+  # columns: either array of strings or "*"
+  #
+  def get_a_view_aux(columns,
+                     value_restrictions,
+                     var_hash)
+
+    # distinguish main table and split table columns
+    first_columns, second_columns = separate_into_1st_and_2nd_table_cols(columns)
+
+    # make pairs of a DB table and the columns from that table
+    tables_and_cols = [SelectTableAndColumns.new(@db_table, first_columns)]
+    if @second_table
+      tables_and_cols << SelectTableAndColumns.new(@second_table, second_columns)
+    end
+
+
+    # and get a view
+    return DBView.new(tables_and_cols, 
+		      value_restrictions + @standard_value_restrictions, 
+                      @ttt_obj.database,
+		      var_hash)
+  end
+
+end
+
+
+###############
+# class DynGoldBinary
+#
+# dynamic gold class:
+# maps all FEs to "FE", and
+# maps @noval to @noval.
+#
+# ID to hand to View in each_hash/each_array/each_sentence if you want
+# to use this dynamic gold class:
+# "binary_gold"
+class DynGoldBinary
+  def initialize(noval)
+    @noval = noval
+  end
+
+  def make(gold)
+    if gold == @noval
+      return @noval
+    else
+      return "FE"
+    end
+  end
+
+  def id()
+    return "binary_gold"
+  end
+end