shalmaneser-rosy 1.2.0.rc4

data/lib/rosy/RosyEval.rb

@@ -0,0 +1,465 @@
+# RosyEval
+# KE May 05
+#
+# Evaluation for Rosy:
+# Precision, Recall, F-score
+# Output to evaluation file,
+# plus optional output of evaluation log file.
+#
+# Builds on the general Salsa Eval package
+
+# Salsa packages
+require "common/Eval"
+require "common/ruby_class_extensions"
+
+# Rosy packages
+require "rosy/RosyIterator"
+require "rosy/RosySplit"
+require "rosy/RosyTask"
+require "rosy/RosyPruning"
+
+# Frprep packages
+require "common/prep_config_data"
+
+#######################################################################
+# This class is a subclass of the general evaluation class
+# Eval, which makes evaluation results readable via
+# readable object variables
+#
+# step: can be argrec, arglab, onestep, as usual, but also 
+#       - "all":
+#          evaluate argrec and arglab together.
+#          When argrec == NONE, use the argrec value, else use the arglab value
+#       - "prune": 
+#          evaluate the pruning column as if it were an argrec assignment
+#
+# When step == argrec or prune, evaluate _only_ the target class FE
+# Otherwise, evaluate all target classes
+class RosyEval < Eval
+  def initialize(exp,      # RosyConfigData object: experiment file
+		 ttt_obj,  # RosyTrainingTestTable object 
+		 step,     # string: argrec, arglab, onestep, all, prune
+		 splitID,  # string: splitlog ID, or nil
+		 testID,   # string: test ID, or nil
+		 outfilename, # string: name of file to print output to
+		 logfilename, # string: name of file to print eval log to (may be nil)
+                 dont_adjoin_frprep_exp) # string: if non-nil, don't re-adjoin frprep experiment obj
+    @exp = exp
+    @step = step
+
+    if outfilename
+      $stderr.puts "Rosy evaluation: printing results to " + outfilename
+    end
+    if logfilename 
+     $stderr.puts "and printing an evaluation log to " + logfilename
+    end
+
+    ##
+    # add preprocessing information to the experiment file object
+    unless dont_adjoin_frprep_exp
+      if splitID
+        # use split data
+        preproc_expname = @exp.get("preproc_descr_file_train")
+      else
+        # use test data
+        preproc_expname = @exp.get("preproc_descr_file_test")
+      end
+      if not(preproc_expname)
+        $stderr.puts "Please set the name of the preprocessing exp. file name"
+        $stderr.puts "in the experiment file."
+        exit 1
+      elsif not(File.readable?(preproc_expname))
+        $stderr.puts "Error in the experiment file:"
+        $stderr.puts "Parameter preproc_descr_file_train has to be a readable file."
+        exit 1
+      end
+      preproc_exp = FrPrepConfigData.new(preproc_expname)
+      @exp.adjoin(preproc_exp)
+    end
+
+    ##
+    # evaluate which labels?
+    if ["argrec", "prune"].include? @step
+      # evaluate only the label "FE"
+      super(outfilename, logfilename, "FE")
+    else
+      # evaluate all target classes
+      super(outfilename, logfilename)
+    end
+
+    ##
+    # what are classifier columns?
+    case @step
+    when "all"
+      # read one argrec and one arglab classifier run column
+      @classif_column_argrec = ttt_obj.existing_runlog("argrec", "test", testID,splitID)
+      @classif_column_arglab = ttt_obj.existing_runlog("arglab", "test", testID,splitID)
+      @columns = ["gold", @classif_column_argrec, @classif_column_arglab]
+
+      if @classif_column_argrec.nil? or @classif_column_arglab.nil?
+        # no run found for the given specifications
+        $stderr.puts "Couldn't determine the run to evaluate."
+        $stderr.puts "There were either none or too many possible runs given your specification.\n"
+        $stderr.puts "Here is a list of all runs the system knows for this experiment ID:\n\n"
+        $stderr.puts ttt_obj.runlog_to_s("test", testID, splitID)
+        exit 1
+      end
+      
+    when "prune"
+      # read pruning column, evaluate as a kind of argrec assignment
+      unless Pruning.prune?(@exp)
+        raise "Error: Pruning evaluation without pruning column. Skipping."
+      end
+      @classif_column = Pruning.colname(@exp)
+      @columns = ["gold", @classif_column]
+
+    else
+      # read the classifier run column for the current step
+      @classif_column = ttt_obj.existing_runlog(@step, "test", testID,splitID)
+      @columns = ["gold", @classif_column]
+
+      if @classif_column.nil?
+        # no run found for the given specifications
+        $stderr.puts "Couldn't determine the run to evaluate."
+        $stderr.puts "There were either none or too many possible runs given your specification.\n"
+        $stderr.puts "Here is a list of all runs the system knows for this experiment ID:\n\n"
+        $stderr.puts ttt_obj.runlog_to_s("test", testID, splitID)
+        exit 1
+      end
+    end
+    
+    ##
+    # make object for iterating through groups and making views
+    case @step
+    when "all"
+      # all: no step in particular
+      @iterator = RosyIterator.new(ttt_obj, exp, "test", 
+                                   "step" => nil, 
+                                   "testID" => testID, 
+                                   "splitID" => splitID,
+                                   "xwise" => "frame")
+    when "prune"
+      # prune: use argrec
+      @iterator = RosyIterator.new(ttt_obj, exp, "test", 
+                                   "step" => "argrec", 
+                                   "testID" => testID, 
+                                   "splitID" => splitID)
+
+    else
+      # use the given step
+      @iterator = RosyIterator.new(ttt_obj, exp, "test", 
+                                   "step" => @step, 
+                                   "testID" => testID, 
+                                   "splitID" => splitID)
+    end
+
+    ##
+    # xwise
+    if @step == "all"
+      # argrec and arglab may have different xwises,
+      # which would create trouble.
+      # just use "frame" instead
+      @xwise = ["frame"]
+    else
+      # evaluate as you have trained and tested
+      @xwise = @iterator.get_xwise_column_names()
+    end
+
+    ##
+    # split? then include FE labels from unparsed sentences
+    # in count of gold labels
+    if splitID
+      # get a FailedParses object for this split
+      @failed_parses_split = FailedParses.new()
+      fp_filename = File.new_filename(@exp.instantiate("rosy_dir",
+                                                  "exp_ID" => @exp.get("experiment_ID")),
+                                 @exp.instantiate("failed_file",
+                                                  "exp_ID" => @exp.get("experiment_ID"),
+                                                  "split_ID" => splitID,
+                                                  "dataset" => "test"))
+      @failed_parses_split.load(fp_filename)
+    end
+
+    # announce the task
+    $stderr.puts "---------"
+    $stderr.print "Rosy experiment #{@exp.get("experiment_ID")}: Evaluating "
+    if splitID
+      $stderr.puts "on split dataset #{splitID}"
+    else
+      $stderr.puts "on test dataset #{testID}"
+    end
+    $stderr.puts "---------"
+  end
+
+  ###
+  protected
+
+  ###
+  # each_group
+  #
+  # yield each group name in turn
+  def each_group()
+
+    @view = nil
+
+    # for the sake of the failed parses module:
+    # it can split the failed parses by frame, target and target_pos,
+    # but if our "xwise" splits the data along any further columns,
+    # the failed parses module cannot know how to split up its failed parses.
+    # so see whether we've got any column names besides the three named above
+    # in our xwise,
+    # and if so, count the groups and split the failed parses evenly between them
+    normal_xwise_cols = ["frame", "target", "target_pos"] & @xwise
+    extra_xwise_cols = @xwise - normal_xwise_cols
+
+    # num_groups_for_normalxwise: hash: normal_xwise_values(string) -> num. of
+    #  groups with these normal xwise values(integer)
+    # where the key normal_xwise_values is a conjunction of
+    # strings <col_name>=<value> joined by commas,
+    # and the column names are in alphabetical order
+    num_groups_for_normalxwise = Hash.new(0)
+
+    unless extra_xwise_cols.empty?
+      # we do have extra columns
+
+      # for each value sequence for normal_xwise_cols: find out how many values
+      # of extra xwise col.s there are
+      @iterator.each_group() { |group_descr_hash, group_name|
+
+        # make the hash key
+        key = normal_xwise_cols.sort.map { |col_name|
+          col_name + "=" + group_descr_hash[col_name]
+        }.join(",")
+
+        # record one occurrence of this hash key
+        num_groups_for_normalxwise[key] += 1
+      }
+    end
+
+    @iterator.each_group() { |group_descr_hash, group_name|
+
+      if @exp.get("verbose")
+        $stderr.puts group_name
+      end
+
+      # construct view for the current group
+      @view = @iterator.get_a_view_for_current_group(@columns)
+
+      ##
+      # get counts of FE labels from unparsed sentences:
+
+      # first take apart the group label to find
+      # the frame name, target name, target POS name in this group
+      # (all but one may be nil)
+      frame = target = target_pos = nil
+
+      # get a description of this group, array of pairs [column name, value]
+      # where column name is the name of one database column
+      @xwise.interleave(group_name.split()).each { |col_name, col_value|
+        case col_name
+        when "frame"
+          frame = col_value
+        when "target"
+          target = col_value
+        when "target_pos"
+          target_pos = col_value
+        else
+          # additional database columns: handled below
+        end
+      }
+
+      # do we have additional column names in "xwise", besides 'frame', 'target', 'target_pos'?
+      if extra_xwise_cols.empty?
+        split_between_groups = 1
+      else
+        key = normal_xwise_cols.sort.map { |col_name|
+          col_name + "=" + group_descr_hash[col_name]
+        }.join(",")
+        split_between_groups = num_groups_for_normalxwise[key]
+
+        # sanity check
+        if split_between_groups == 0
+          raise "shouldn't be here"
+        end
+      end
+
+      # failed_fes returns: hash that maps FE names [String] onto numbers of failed FEs [Int] 
+      if @failed_parses_split
+        @failed_parses_split.failed_fes(frame, target, target_pos).each_pair { |fe, count|
+          # add this number of gold labels we failed to find
+          # to the number of gold labels that Eval counts
+
+          # if argrec, map all non-NONE FEs to "FE"
+          if @step == "argrec" and fe != @exp.get("noval")
+            fe = "FE"
+          end
+          inject_gold_counts(group_name, fe, (count.to_f / split_between_groups.to_f).round)
+        }
+      end
+
+      # yield the name of the group to the Eval object for evaluation
+      yield group_name
+      @view.close()
+    }
+  end
+
+  ###
+  # each_instance
+  #
+  # given a group name, yield each instance of this group in turn,
+  # or rather: yield pairs [gold_class(string), assigned_class(string)]
+  #
+  # this method depends on each_group() having been called before and
+  # having initialized @view to the right view object
+  def each_instance(group) # string: group name
+    case @step
+    when "all"
+      # step "all":
+      # if the argrec label is "NONE", use that as the assigned label.
+      # else use the arglab-label
+      @view.each_hash { |row|
+        if row[@classif_column_argrec] == @exp.get("noval")
+          yield [ row["gold"], row[@classif_column_argrec] ]
+        else
+          yield [ row["gold"], row[@classif_column_arglab] ]
+        end
+      }
+    
+    when "prune"
+      # step "prune":
+      # if the pruning column has entry 1, regard as assignment "FE",
+      # else regard as assignment "NONE".
+      @view.each_hash { |row|
+        if row[@classif_column] == "1"
+          yield [ row["gold"], "FE" ]
+        else
+          yield [ row["gold"], @exp.get("noval") ]
+        end
+      }
+
+    else
+      # argrec, arglab, onestep:
+      # just yield pairs [goldlabel, classif_column_label]
+      # as given in the view
+
+      @view.each_hash { |row|
+        yield [row["gold"], row[@classif_column]]
+      }
+    end
+      
+  end
+end
+
+###########################################################33
+# This is the class to be called from rosy.rb
+###########################################################33
+class RosyEvalTask < RosyTask
+
+  def initialize(exp,      # RosyConfigData object: experiment description
+		 opts,     # hash: runtime argument option (string) -> value (string)
+		 ttt_obj)  # RosyTrainingTestTable object
+
+    #####
+    # In enduser mode, this whole task is unavailable
+    in_enduser_mode_unavailable()
+
+    @exp = exp
+    @ttt_obj = ttt_obj
+
+    ##
+    # check runtime options
+    @step = "both"
+    @splitID = nil
+    @testID = default_test_ID()    
+
+    opts.each do |opt,arg|
+      case opt
+      when "--step"
+	unless ["argrec", "arglab", "both", "onestep"].include? arg
+	  raise "Classification step must be one of: argrec, arglab, both, onestep. I got: " + arg.to_s
+	end
+	@step = arg
+      when "--logID"
+	@splitID = arg
+      when "--testID"
+	@testID = arg
+      else
+	# this is an option that is okay but has already been read and used by rosy.rb
+      end
+    end
+  end
+
+  def perform()
+    dont_adjoin_frprep_exp = nil
+    original_step = @step
+
+    if ["both", "argrec", "onestep"].include? original_step and
+        Pruning.prune?(@exp)
+      # evaluate pruning
+      $stderr.puts "Rosy evaluating pruning"
+      @step = "prune"
+      perform_aux()
+      dont_adjoin_frprep_exp = "dont_adjoin_frprep_exp"
+    end
+
+    if original_step == "both"
+      # both? then do first argrec, then arglab
+      $stderr.puts "Rosy evaluating step argrec"
+      @step = "argrec"
+      perform_aux(dont_adjoin_frprep_exp)
+
+      
+      $stderr.puts "Rosy evaluating step arglab"
+      @step = "arglab"
+      perform_aux("dont_adjoin_frprep_exp")
+
+# KE Jan 30, 2006: evaluation "all" deactivated until we've
+# figured out how to evaluate accuracy for the NONE class
+#      $stderr.puts "Rosy overall evaluation"
+#      @step = "all"
+#      perform_aux("dont_adjoin_frprep_exp")
+
+    else
+      # not both? then just do one
+      @step = original_step
+      perform_aux(dont_adjoin_frprep_exp)
+    end
+  end
+
+  ###############3
+  private
+
+  # perform_aux: do the actual work of the perform() method
+  # moved here because of the possibility of having @step=="both",
+  # which makes it necessary to perform two eval steps one after the other
+  def perform_aux(dont_adjoin_frprep_exp = nil)  # string passed on to RosyEval initialize method
+    # construct names for evaluation output file
+    # and evaluation log file (which classifies each instances as correct/incorrect/unassigned)
+    if @splitID
+      outfilename_id = "split" + @splitID
+    else
+      outfilename_id = "test" + @testID 
+    end
+    @outfilename = File.new_filename(@exp.instantiate("rosy_dir",
+                                                      "exp_ID" => @exp.get("experiment_ID")),
+                                     @exp.instantiate("eval_file",
+                                                      "exp_ID" => @exp.get("experiment_ID"),
+                                                      "test_ID" => outfilename_id,
+                                                      "step" => @step))
+
+    if @exp.get("print_eval_log")
+      @logfilename = File.new_filename(@exp.instantiate("rosy_dir",
+                                                        "exp_ID" => @exp.get("experiment_ID")),
+                                       @exp.instantiate("log_file",
+                                                        "exp_ID" => @exp.get("experiment_ID"),
+                                                        "test_ID" => outfilename_id,
+                                                        "step" => @step))
+    else
+      @logfilename = nil
+    end
+    @eval_obj = RosyEval.new(@exp, @ttt_obj, @step, @splitID, @testID, 
+                             @outfilename, @logfilename,
+                             dont_adjoin_frprep_exp)
+    @eval_obj.compute()
+  end
+end