def main(argv):
install_path = os.path.abspath(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
utils_dirname = os.path.join(install_path,'lib')
sys.path.append(utils_dirname)
from dataset_walker import dataset_walker
list_dir = os.path.join(install_path,'config')
parser = argparse.ArgumentParser(description='Evaluate output from a belief tracker.')
parser.add_argument('--dataset', dest='dataset', action='store', metavar='DATASET', required=True,
help='The dataset to analyze')
parser.add_argument('--dataroot',dest='dataroot',action='store', metavar='PATH', required=True,
help='Will look for corpus in <destroot>/<dataset>/...')
parser.add_argument('--trackfile',dest='scorefile',action='store',metavar='JSON_FILE',required=True,
help='File containing score JSON')
parser.add_argument('--scorefile',dest='csv',action='store',metavar='CSV_FILE',required=True,
help='File to write with CSV scoring data')
parser.add_argument('--ontology',dest='ontology',action='store',metavar='JSON_FILE',required=True,
help='JSON Ontology file')
parser.add_argument('--rocdump',dest='rocdump',action='store',metavar='FILE_STEM',
help='If present, use this file stem to write out ROC plot data: filestem.<schedule>.<slot>.<type>.csv, where type is either roc (which contains the ROC curve coordinates) or scores (which contains the raw scores used to compute the ROC curves).')
args = parser.parse_args()
sessions = dataset_walker(args.dataset,dataroot=args.dataroot,labels=True)
tracker_output = json.load(open(args.scorefile))
ontology = json.load(open(args.ontology))
slots_informable = ontology["informable"].keys()
slots_requestable = ontology["requestable"]
csvfile = open(args.csv,'w')
# what stats are there?
stats = []
stat_classes = [Stat_Accuracy, Stat_Probs, Stat_MRR, Stat_Updates, Stat_ROC]
for schedule in SCHEDULES:
for label_scheme in LABEL_SCHEMES:
for component in ['goal','requested', 'method', 'all']:
if component == 'goal' :
for slot in slots_informable + ['all','joint','joint_independent'] :
for stat_class in stat_classes:
stats.append((('goal', slot), (schedule, label_scheme), stat_class()))
elif component == 'requested' :
if label_scheme != "a" :
continue
for slot in slots_requestable + ['all'] :
for stat_class in stat_classes:
stats.append((('requested', slot), (schedule, label_scheme), stat_class()))
elif component == 'method' :
for stat_class in stat_classes:
stats.append((('method',), (schedule, label_scheme), stat_class()))
elif component == 'all' :
for stat_class in stat_classes:
stats.append((('all',), (schedule, label_scheme), stat_class()))
turn_counter = 0.0
for session_num, (session_tracker, session) in enumerate(zip(tracker_output['sessions'], sessions)):
for _, _, stat_class in stats:
stat_class.newDialog()
session_id = session.log['session-id']
try:
# these are the set of slots 'mentioned so far', i.e. for schedule2
S = defaultdict(lambda : set([]))
S_requested = set([])
session_length = len(session)
goal_labels_b, method_labels_b = misc.LabelsB(session, ontology)
method_schedule_2 = False # whether schedule 2 is active for method
for turn_num, ((log_turn,label_turn),_tracker_turn) in enumerate(zip(session,session_tracker['turns'])):
turn_counter += 1.0
S_new = misc.S(log_turn, ontology)
for slot in S_new :
S[slot] = S[slot].union(S_new[slot])
# remove just informed slots from S_requested
S_requested = S_requested.difference(misc.SysInformed(log_turn))
# add in ones from slu hyps
S_requested = S_requested.union(set(misc.S_requested(log_turn)))
tracker_goal_labels = _tracker_turn["goal-labels"]
for slot in slots_informable:
if slot in tracker_goal_labels :
tracker_goal_labels[slot] = normalise_dist(tracker_goal_labels[slot].items(), (session_id, turn_num, "goal."+slot))
else :
tracker_goal_labels[slot] = [(None, 1.0)]
# prepare for joint goals scoring:
tracker_goal_joint_labels = "independent"
if "goal-labels-joint" in _tracker_turn :
tracker_goal_joint_labels = _tracker_turn["goal-labels-joint"]
if tracker_goal_joint_labels != "independent" :
# tracker_goal_joint_labels must be a list of joint hyps
tracker_goal_joint_labels = [(hyp["slots"], hyp["score"]) for hyp in tracker_goal_joint_labels]
tracker_goal_joint_labels = normalise_dist(tracker_goal_joint_labels, (session_id, turn_num, "goal.joint"))
# also gather the correct joint label
true_goal_joint = None
for slot in label_turn["goal-labels"]:
if true_goal_joint == None :
true_goal_joint = {}
true_goal_joint[slot] = label_turn["goal-labels"][slot]
true_goal_joint_b = None
for slot in goal_labels_b[turn_num]:
if true_goal_joint_b == None :
true_goal_joint_b = {}
true_goal_joint_b[slot] = goal_labels_b[turn_num][slot]
tracker_requested_slots = _tracker_turn["requested-slots"]
for slot in tracker_requested_slots:
dist = [(True, tracker_requested_slots[slot]), (False,1.0-tracker_requested_slots[slot])]
tracker_requested_slots[slot] = normalise_dist(dist, (session_id, turn_num, "requested."+slot))
tracker_method_label = normalise_dist(_tracker_turn["method-label"].items(), (session_id, turn_num,"method"))
# for method schedule 2, work out whether any slu-hyp has been given
# which informs the method:
if not method_schedule_2 :
mact = log_turn["output"]["dialog-acts"]
for slu_hyp in log_turn["input"]["live"]["slu-hyps"] :
user_act = slu_hyp["slu-hyp"]
method_label = misc.MethodLabel(user_act, mact)
if method_label != "none" :
method_schedule_2 = True
break
for component, (schedule, label_scheme), stat_class in stats:
if component[0] == "goal" and (component[1] == "joint" or component[1] == "joint_independent"):
if schedule == 2:
# calculate schedule2 applicability
applies = False
for slot in slots_informable:
if len(S[slot]) > 0:
applies = True
break
if not applies :
continue
this_true_label = true_goal_joint
if label_scheme == "b" :
this_true_label = true_goal_joint_b
if tracker_goal_joint_labels == "independent" or component[1] == "joint_independent" :
stat_class.add(tracker_goal_labels, this_true_label, (session_id, turn_num, component, schedule, label_scheme), independent=True)
else :
stat_class.add(tracker_goal_joint_labels, this_true_label, (session_id, turn_num, component, schedule, label_scheme))
if (component[0] == "goal" or component[0] == "all") and (len(component)==1 or ("joint" not in component[1])) :
if component[0] == "all" or component[1] == "all" :
slots = slots_informable[:]
else :
slots = [component[1]]
for slot in slots:
if schedule ==2 and len(S[slot]) == 0 :
continue
dist = tracker_goal_labels[slot]
true_label = None
if slot in label_turn["goal-labels"] :
true_label = label_turn["goal-labels"][slot]
if label_scheme == "b" :
true_label = None
if slot in goal_labels_b[turn_num] :
true_label = goal_labels_b[turn_num][slot]
stat_class.add(dist, true_label, (session_id, turn_num, component, schedule, label_scheme))
if component[0] == "requested" or component[0] == "all" :
if component[0] == "all" or component[1] == "all":
slots = slots_requestable[:]
else :
slots = [component[1]]
for slot in slots:
if schedule ==2 and (slot not in S_requested):
continue
dist = [(False,1.0), (True,0.0)]
if slot in tracker_requested_slots :
dist = tracker_requested_slots[slot]
true_label = (slot in label_turn["requested-slots"])
stat_class.add(dist, true_label, (session_id, turn_num, component, schedule, label_scheme))
if component[0] == "method" or component[0] == "all":
if schedule == 2 and not method_schedule_2:
continue # no slu hyp informing the method has been given yet.
dist = tracker_method_label
true_label = label_turn["method-label"]
if label_scheme == "b" :
true_label = method_labels_b[turn_num]
stat_class.add(dist, true_label, (session_id, turn_num, component, schedule, label_scheme))
except KeyboardInterrupt :
raise
except:
traceback.print_exc(file=sys.stdout)
print "While scoring " + str(session_id)
# output to csv
print >>csvfile,( "state_component, stat, schedule, label_scheme, N, result")
for stat in stats:
component, (schedule, label_scheme), stat_class = stat
results = stat_class.results()
for stat_subname, N, result in results:
if result == None :
result = "-"
else :
result = "%.7f"%result
print >>csvfile,( "%s, %s, %i, %s, %i, %s"%(".".join(component), stat_subname, schedule, label_scheme, N, result))
if isinstance(stat_class, Stat_ROC) and (args.rocdump):
rocfile = args.rocdump + '.schedule' + str(schedule) + str(label_scheme)+'.' + (".".join(component)) + '.roc.csv'
scoresfile = args.rocdump + '.schedule' + str(schedule) + str(label_scheme)+'.' + (".".join(component)) + '.scores.csv'
stat_class.DumpROCToFile(rocfile)
stat_class.DumpScoresToFile(scoresfile)
print >>csvfile,'basic,total_wall_time,,,,%s' % (tracker_output['wall-time'])
print >>csvfile,'basic,sessions,,,,%s' % (len(sessions))
print >>csvfile,'basic,turns,,,,%i' % (int(turn_counter))
print >>csvfile,'basic,wall_time_per_turn,,,,%s' % (tracker_output['wall-time'] / turn_counter)
print >>csvfile,'basic,dataset,,,,%s' % (tracker_output['dataset'] )
csvfile.close()
self.labels_filename = labels_filename
f = open(applog_filename)
self.log = json.load(f)
f.close()
if (labels_filename != None):
f = open(labels_filename)
self.labels = json.load(f)
f.close()
else:
self.labels = None
def __iter__(self):
if (self.labels_filename != None):
for (log, labels) in zip(self.log['turns'], self.labels['turns']):
yield (log, labels)
else:
for log in self.log['turns']:
yield (log, None)
def __len__(self, ):
return len(self.log['turns'])
if __name__ == '__main__':
import misc
dataset = dataset_walker("HDCCN", dataroot="data", labels=True)
for call in dataset:
if call.log["session-id"] == "voip-f32f2cfdae-130328_192703":
for turn, label in call:
print(misc.S(turn))
self.labels_filename = labels_filename
f = open(applog_filename)
self.log = json.load(f)
f.close()
if (labels_filename != None):
f = open(labels_filename)
self.labels = json.load(f)
f.close()
else:
self.labels = None
def __iter__(self):
if (self.labels_filename != None):
for (log,labels) in zip(self.log['turns'],self.labels['turns']):
yield (log,labels)
else:
for log in self.log['turns']:
yield (log,None)
def __len__(self, ):
return len(self.log['turns'])
if __name__ == '__main__':
import misc
dataset = dataset_walker("HDCCN", dataroot="data", labels=True)
for call in dataset :
if call.log["session-id"]=="voip-f32f2cfdae-130328_192703" :
for turn, label in call :
print misc.S(turn)
