class MMUnsCluster:
def __init__(self, list_of_shows):
#self.cls_parse = ParseTextToJson(file_names_dict = {}, list_file_name = list_of_shows)
#self.programs_list = self.cls_parse._get_programs_list()
self.list_of_shows = list_of_shows
self.cls_names = TargetRel(None)
self.name_map = self.cls_names.name_map_all()
self.id_map = self.cls_names.target_id_map(self.name_map)
def initialize_dict(self):
speaker_cluster = {}
with open(self.list_of_shows, 'r') as f_:
for line in f_:
pr_name1, pr_name2, program_date, program_idx = line.strip().split('_')
program_name = pr_name1+'_'+pr_name2
if program_name in speaker_cluster:
if program_idx in speaker_cluster[program_name]:
speaker_cluster[program_name][program_idx][program_date] = {'interval':[], 'cluster_names':[]}
else:
speaker_cluster[program_name][program_idx] = {program_date: {'interval':[], 'cluster_names':[]}}
else:
speaker_cluster[program_name] = {program_idx: {program_date: {'interval':[], 'cluster_names':[]}}}
return speaker_cluster
def parse_diarization_file(self, file_name):
"""
prog_name:
prog_id:
prog_date:
'interval':[ [],[],... ]
'cluster_names':[ ... ]
'overlap_list': [{name:time, name:time,...},{name:time},...] sonradan ekleniyor
'assigned_names': [name, name,...] sonradan ekleniyor
'sealed': [True, False,...] sonradan ekleniyor
"""
N = 0
speaker_cluster = self.initialize_dict()
with open(file_name,'r') as f_:
prev_beg_time = -1
for line in f_:
N+=1
line_split = line.strip().split()
program_detail = line_split[0]
person_name = line_split[-1]
time_interval = map(float, parse_time_interval(line_split[2:4]))
pr_name1, pr_name2, program_date, program_idx = program_detail.split('_')
program_name = pr_name1+'_'+pr_name2
speaker_cluster[program_name][program_idx][program_date]['interval'].append([time_interval[0],sum(time_interval)])
speaker_cluster[program_name][program_idx][program_date]['cluster_names'].append(person_name)
if speaker_cluster[program_name][program_idx][program_date]['interval'][-1][0] == prev_beg_time:
print 'buldum speaker'
prev_beg_time = time_interval[0]
return speaker_cluster,N
def parse_written_name_file(self, file_name):
written_names = self.initialize_dict()
with open(file_name,'r') as f_:
prev_beg_time = -1
for line in f_:
line_split = line.strip().split()
program_detail = line_split[0]
person_name = line_split[-1]
time_interval = map(float, parse_time_interval(line_split[1:3]))
pr_name1, pr_name2, program_date, program_idx = program_detail.split('_')
program_name = pr_name1+'_'+pr_name2
written_names[program_name][program_idx][program_date]['interval'].append(time_interval)
written_names[program_name][program_idx][program_date]['cluster_names'].append(person_name)
if written_names[program_name][program_idx][program_date]['interval'][-1][0] == prev_beg_time:
print 'buldum written'
prev_beg_time = time_interval[0]
return written_names
"""
def parse_speaker_identification_files(self, file_name):
speaker_identities = self.initialize_dict()
with open(file_name,'r') as f_:
prev_beg_time = -1
for line in f_:
line_split = line.strip().split()
program_detail = line_split[0]
person_name = line_split[-1]
probability = float(person_name[person_name.index('<')+1:-1])
person_name = person_name[:person_name.index('<')]
time_interval = map(float, parse_time_interval(line_split[1:3]))
pr_name1, pr_name2, program_date, program_idx = program_detail.split('_')
program_name = pr_name1+'_'+pr_name2
speaker_identities[program_name][program_idx][program_date]['interval'].append(time_interval)
speaker_identities[program_name][program_idx][program_date]['cluster_names'].append(person_name)
if speaker_identities[program_name][program_idx][program_date]['interval'][-1][0] == prev_beg_time:
print 'buldum written'
prev_beg_time = time_interval[0]
return speaker_identities
"""
def _overlap_modes(self,interval_sp, interval_wr, wr_names):
overlap_list = []
for _ in xrange(len(interval_sp)):
overlap_list.append({})
interval_sp = np.array(interval_sp)
interval_wr = np.array(interval_wr)
for i, inter_sp in enumerate(interval_sp):
beg_sp,end_sp = inter_sp[0],inter_sp[1]
for inter_wr, name in zip(interval_wr,wr_names):
beg_wr,end_wr = inter_wr[0],inter_wr[1]
if beg_wr < beg_sp and end_wr <= beg_sp:
continue
if beg_wr >= end_sp and end_wr > end_sp:
break
if beg_wr <= beg_sp and end_wr <= end_sp:
overlap_list[i][name] = overlap_list[i].get(name,0) + end_wr - beg_sp
elif beg_wr >= beg_sp and end_wr <= end_sp:
overlap_list[i][name] = overlap_list[i].get(name,0) + end_wr - beg_wr
elif beg_wr >= beg_sp and end_wr >= end_sp:
overlap_list[i][name] = overlap_list[i].get(name,0) + end_sp - beg_wr
elif beg_wr <= beg_sp and end_wr >= end_sp:
overlap_list[i][name] = overlap_list[i].get(name,0) + end_sp - beg_sp
return overlap_list
def overlap_modes(self, speaker_cluster, written_names, list_name = 'overlap_list'):
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
#if program_name == "LCP_CaVousRegarde" and program_date == "2011-12-20" and program_idx == "204600":
# pass
interval_sp = speaker_cluster[program_name][program_idx][program_date]['interval']
interval_wr = written_names[program_name][program_idx][program_date]['interval']
names_wr = written_names[program_name][program_idx][program_date]['cluster_names']
overlap_list = self._overlap_modes(interval_sp, interval_wr, names_wr)
speaker_cluster[program_name][program_idx][program_date][list_name] = overlap_list
return speaker_cluster
def assign_raw_supervised(self, speaker_cluster, mode_name = "GMM", all_assign = False):
N_assigned = 0
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
sealed = speaker_cluster[program_name][program_idx][program_date]['sealed']
assigned_names = speaker_cluster[program_name][program_idx][program_date]['assigned_names']
candidate_names_GMM = speaker_cluster[program_name][program_idx][program_date][mode_name]
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if not sealed[i] and candidate_names_GMM[i]:
cand_names = candidate_names_GMM[i].keys()
cand_name = cand_names[np.argmax(candidate_names_GMM[i].values())]
probability = float(cand_name[cand_name.index('<')+1:-1])
pers_name = cand_name[:cand_name.index('<')]
if probability > 0.5 or all_assign:
sealed[i] = True
assigned_names[i] = pers_name
N_assigned += 1
print N_assigned, "cluster is assigned in ", mode_name
return speaker_cluster
def assign_raw_supervised_cluster_based(self, speaker_cluster, supervised_dict):
N_assigned = 0
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
sealed = speaker_cluster[program_name][program_idx][program_date]['sealed']
assigned_names = speaker_cluster[program_name][program_idx][program_date]['assigned_names']
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if not sealed[i]:
cand_names_GMM,cand_probs_GMM,cand_times_GMM = extract_unique_names_probs(supervised_dict[cluster_name]['GMM'])
cand_names_SVM,cand_probs_SVM,cand_times_SVM = extract_unique_names_probs(supervised_dict[cluster_name]['SVM'])
i_GMM,i_SVM = np.argmax(cand_times_GMM),np.argmax(cand_times_SVM)
n_GMM,n_SVM = cand_names_GMM[i_GMM],cand_names_SVM[i_SVM]
p_GMM,p_SVM = cand_probs_GMM[i_GMM],cand_probs_SVM[i_SVM]
pers_name = n_GMM if p_GMM > p_SVM else n_SVM
pp = np.max((p_GMM,p_SVM))
if pp > 0:
sealed[i] = True
assigned_names[i] = pers_name
N_assigned += 1
print N_assigned, "cluster is assigned in GMM"
return speaker_cluster
def get_dict_of_supervised(self, speaker_cluster):
supervised_dict = {}
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
candidate_names_GMM = speaker_cluster[program_name][program_idx][program_date]['GMM']
candidate_names_SVM = speaker_cluster[program_name][program_idx][program_date]['SVM']
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if supervised_dict.get(cluster_name,None) is None:
supervised_dict[cluster_name] = {'GMM':{},'SVM':{}}
for pers_name in candidate_names_GMM[i]:
supervised_dict[cluster_name]['GMM'][pers_name] = supervised_dict[cluster_name]['GMM'].get(pers_name,0) + candidate_names_GMM[i][pers_name]
for pers_name in candidate_names_SVM[i]:
supervised_dict[cluster_name]['SVM'][pers_name] = supervised_dict[cluster_name]['SVM'].get(pers_name,0) + candidate_names_SVM[i][pers_name]
return supervised_dict
def temporal_count(self, speaker_cluster):
temporal_dict = {}
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
previous_cluster_name = -1
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if temporal_dict.get(cluster_name,None) is None:
temporal_dict[cluster_name] = {}
if i == 0:
previous_cluster_name = cluster_name
continue
temporal_dict[previous_cluster_name][cluster_name] = temporal_dict[previous_cluster_name].get(cluster_name,0) + 1
previous_cluster_name = cluster_name
return temporal_dict
def temporal_assignment(self, speaker_cluster, temporal_dict, result_dict):
N_assigned = 0
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
sealed = speaker_cluster[program_name][program_idx][program_date]['sealed']
assigned_names = speaker_cluster[program_name][program_idx][program_date]['assigned_names']
previous_cluster_name = -1
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if not sealed[i] and i != 0 and temporal_dict.get(previous_cluster_name,False):
cl_names = temporal_dict[previous_cluster_name].keys()
cl_freqs = temporal_dict[previous_cluster_name].values()
cl_name = cl_names[np.argmax(cl_freqs)]
if result_dict[cl_name]: #and len(result_dict[cl_name]) == 1:
sealed[i] = True
names = result_dict[cl_name].keys()
freqs = result_dict[cl_name].values()
assigned_names[i] = names[np.argmax(freqs)]
#assigned_names[i] = result_dict[cl_name].keys()[0]
N_assigned += 1
previous_cluster_name = cluster_name
print N_assigned, "cluster is assigned in temporal assignment"
return speaker_cluster
def get_results_untill_now(self, speaker_cluster, with_labels = False):
result_dict = {}
correct = 0
labels = None
over_all = 0
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
sealed = speaker_cluster[program_name][program_idx][program_date]['sealed']
assigned_names = speaker_cluster[program_name][program_idx][program_date]['assigned_names']
if with_labels:
labels = speaker_cluster[program_name][program_idx][program_date]['labels']
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if result_dict.get(cluster_name,None) is None:
result_dict[cluster_name] = {}
if sealed[i]:
result_dict[cluster_name][assigned_names[i]] = result_dict[cluster_name].get(assigned_names[i],0) + 1
if labels and labels[i]:
over_all += 1
target = labels[i].keys()[np.argmax(labels[i].values())]
correct = correct + 1 if target == assigned_names[i] else correct
print "there are ", correct, " in ", over_all
return result_dict
def M2_assignment(self, speaker_cluster):
one_to_one = 0
all_sayac = 0
M2_dict = {}
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
assigned_names = [None]*len(speaker_cluster[program_name][program_idx][program_date]['cluster_names'])
sealed = [False]*len(speaker_cluster[program_name][program_idx][program_date]['cluster_names'])
speaker_cluster[program_name][program_idx][program_date]['assigned_names'] = assigned_names
speaker_cluster[program_name][program_idx][program_date]['sealed'] = sealed
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
overlap_map = speaker_cluster[program_name][program_idx][program_date]['overlap_list'][i]
all_sayac += 1
if not cluster_name in M2_dict:
M2_dict[cluster_name] = []
if len(overlap_map) == 1:
one_to_one += 1
assigned_names[i] = overlap_map.keys()[0]
sealed[i] = True
M2_dict[cluster_name].append(assigned_names[i])
print one_to_one, ' is assigned in M2', all_sayac
return speaker_cluster, M2_dict
def propagate_M2(self, speaker_cluster, M2_dict):
N_assigned = 0
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
sealed = speaker_cluster[program_name][program_idx][program_date]['sealed']
assigned_names = speaker_cluster[program_name][program_idx][program_date]['assigned_names']
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if not sealed[i] and M2_dict[cluster_name] and len(M2_dict[cluster_name]) == 1:
sealed[i] = True
assigned_names[i] = M2_dict[cluster_name][0]
N_assigned += 1
print N_assigned, "cluster is assigned in M2_propagate"
return speaker_cluster
def getSpeakerDoc_NameDoc(self, speaker_cluster):
speaker_document = {}
name_document = {}
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
overlap_map = speaker_cluster[program_name][program_idx][program_date]['overlap_list'][i]
speaker_document[cluster_name] = speaker_document.get(cluster_name, {})
for wr_name in overlap_map:
speaker_document[cluster_name][wr_name] = speaker_document[cluster_name].get(wr_name, 0) + overlap_map[wr_name]
name_document[wr_name] = name_document.get(wr_name, {})
name_document[wr_name][cluster_name] = name_document[wr_name].get(cluster_name,0) + overlap_map[wr_name]
return (speaker_document, name_document)
def stableMatching(self, speaker_document, name_document, name_pers = True):
# name based matching
speaker_pref = {}
speaker_matched = {}
total_N_cluster = len(speaker_document)
for cluster_name in speaker_document:
speaker_matched[cluster_name] = None
if speaker_document[cluster_name]:
k_list = np.array(speaker_document[cluster_name].keys())
v_list = np.array(speaker_document[cluster_name].values())
for i,person_name in enumerate(k_list):
idf = np.sum(name_document[person_name].values())/total_N_cluster
v_list[i] = 0 if idf == 0 else v_list[i]/idf
ind = v_list.argsort()[::-1]
k_list = k_list[ind]
speaker_pref[cluster_name] = copy.copy(list(k_list))
else:
speaker_pref[cluster_name] = []
name_pref = {}
name_matched = {}
total_N_name = len(name_document)
for person_name in name_document:
name_matched[person_name] = None
if name_document[person_name]:
k_list = np.array(name_document[person_name].keys())
v_list = np.array(name_document[person_name].values())
for i,cluster_name in enumerate(k_list):
idf = np.sum(speaker_document[cluster_name].values())/total_N_name
v_list[i] = 0 if idf == 0 else v_list[i]/idf
ind = v_list.argsort()[::-1]
k_list = k_list[ind]
name_pref[person_name] = copy.copy(list(k_list))
else:
name_pref[person_name] = []
if name_pers:
ischange = True
while ischange:
ischange = False
for person_name in name_pref:
if name_matched[person_name]:
continue
for candidate in name_pref[person_name]:
if speaker_matched[candidate]:
if person_name in speaker_pref[candidate]:
enemy = speaker_matched[candidate]
if speaker_pref[candidate].index(person_name) < speaker_pref[candidate].index(enemy):
speaker_matched[candidate] = person_name
name_matched[person_name] = candidate
ischange = True
break
else:
speaker_matched[candidate] = person_name
name_matched[person_name] = candidate
ischange = True
break
else:
ischange = True
while ischange:
ischange = False
for cluster_name in speaker_pref:
if speaker_matched[cluster_name]:
continue
for candidate in speaker_pref[cluster_name]:
if name_matched[candidate]:
if cluster_name in name_pref[candidate]:
enemy = name_matched[candidate]
if name_pref[candidate].index(cluster_name) < name_pref[candidate].index(enemy):
name_matched[candidate] = cluster_name
speaker_matched[cluster_name] = candidate
ischange = True
break
else:
name_matched[candidate] = cluster_name
speaker_matched[cluster_name] = candidate
ischange = True
break
return speaker_matched
def method_M1(self, speaker_cluster, name_pers = True):
speaker_cluster,_ = self.M2_assignment(speaker_cluster)
#speaker_cluster = self.propagate_M2(speaker_cluster, M2_dict)
sp_doc, nm_doc = self.getSpeakerDoc_NameDoc(speaker_cluster)
speaker_matched = self.stableMatching(sp_doc, nm_doc, name_pers = name_pers)
N_assigned = 0
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
sealed = speaker_cluster[program_name][program_idx][program_date]['sealed']
assigned_names = speaker_cluster[program_name][program_idx][program_date]['assigned_names']
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if not sealed[i] and speaker_matched.get(cluster_name,None) is not None:
sealed[i] = True
assigned_names[i] = speaker_matched[cluster_name]
N_assigned += 1
print N_assigned, "cluster is assigned in method M1"
return speaker_cluster
def method_M3(self, speaker_cluster, speaker_document, name_document, name_pers = True):
## name perspective
speaker_matched = {}
if name_pers:
total_N_name = len(name_document)
for person_name in name_document:
tf_idf = np.zeros(len(name_document[person_name]))
cluster_names = name_document[person_name].keys()
if not cluster_names:
continue
for i, cluster_name in enumerate(cluster_names):
tf = speaker_document[cluster_name][person_name]
idf = np.sum(speaker_document[cluster_name].values())/total_N_name
tf_idf[i] = 0 if idf == 0 else tf/idf
speaker_matched[cluster_names[tf_idf.argmax()]] = person_name
else:
total_N_cluster = len(speaker_document)
for cluster_name in speaker_document:
#if cluster_name == 'MS41_BFMTV_BFMStory_2012-01-10_175800':
# pass
tf_idf = np.zeros(len(speaker_document[cluster_name]))
person_names = speaker_document[cluster_name].keys()
if not person_names:
continue
for i, person_name in enumerate(person_names):
tf = name_document[person_name][cluster_name]
idf = np.sum(name_document[person_name].values())/total_N_cluster
tf_idf[i] = 0 if idf == 0 else tf/idf
speaker_matched[cluster_name] = person_names[tf_idf.argmax()]
N_assigned = 0
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
sealed = speaker_cluster[program_name][program_idx][program_date]['sealed']
assigned_names = speaker_cluster[program_name][program_idx][program_date]['assigned_names']
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
#if cluster_name == 'MS41_BFMTV_BFMStory_2012-01-10_175800':
# pass
if not sealed[i] and speaker_matched.get(cluster_name,None) is not None:
sealed[i] = True
assigned_names[i] = speaker_matched[cluster_name]
N_assigned += 1
print N_assigned, "cluster is assigned in method M3"
return speaker_cluster
def write_as_hyp(self, speaker_cluster, file_name):
with open(file_name,'w') as file_:
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
assigned_names = speaker_cluster[program_name][program_idx][program_date]['assigned_names']
for i,cl_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
pred_name = 'Inconnu_' + cl_name if assigned_names[i] is None else assigned_names[i]
beg,last = speaker_cluster[program_name][program_idx][program_date]['interval'][i]
string_ = "%s_%s_%s %s %s %s %s\n"%(program_name,program_date,program_idx,beg,
last, 'speaker', pred_name)
file_.write(string_)
def assignM1(self, speaker_cluster, file_name):
speaker_document = {}
name_document = {}
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
overlap_map = speaker_cluster[program_name][program_idx][program_date]['overlap_list'][i]
speaker_document[cluster_name] = speaker_document.get(cluster_name, {})
for wr_name in overlap_map:
speaker_document[cluster_name][wr_name] = speaker_document[cluster_name].get(wr_name, 0) + overlap_map[wr_name]
name_document[wr_name] = name_document.get(wr_name, {})
name_document[wr_name][cluster_name] = name_document[wr_name].get(cluster_name,0) + overlap_map[wr_name]
for wr_name in name_document:
assign_cluster = None
largest = 0
for cluster_name in name_document[wr_name]:
if name_document[wr_name][cluster_name] > largest:
largest = name_document[wr_name][cluster_name]
assign_cluster = cluster_name
if assign_cluster is not None:
if speaker_document[assign_cluster].get('command',0) == -1:
continue
if len(name_document[wr_name]) == 1:
speaker_document[assign_cluster]['command'] = -1
speaker_document[assign_cluster]['assigned_name'] = wr_name
continue
if speaker_document[assign_cluster].get('command',0) < largest:
speaker_document[assign_cluster]['assigned_name'] = wr_name
speaker_document[assign_cluster]['command'] = largest
with open(file_name,'w') as file_:
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
for i,cl_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
pred_name = speaker_document[cl_name].get('assigned_name','unknownUmit')
beg,last = speaker_cluster[program_name][program_idx][program_date]['interval'][i]
string_ = "%s_%s_%s %s %s %s %s\n"%(program_name,program_date,program_idx,beg,
last, 'speaker', pred_name)
file_.write(string_)
return (speaker_document,name_document)
class MMSupCluster:
def __init__(self, list_of_shows):
self.list_of_shows = list_of_shows
self.cls_names = TargetRel(None)
self.name_map = self.cls_names.name_map_all()
self.id_map = self.cls_names.target_id_map(self.name_map)
def initialize_dict(self):
speaker_cluster = {}
with open(self.list_of_shows, 'r') as f_:
for line in f_:
pr_name1, pr_name2, program_date, program_idx = line.strip().split('_')
program_name = pr_name1+'_'+pr_name2
if program_name in speaker_cluster:
if program_idx in speaker_cluster[program_name]:
speaker_cluster[program_name][program_idx][program_date] = {'interval':[], 'cluster_names':[],'labels':[],'GMM':[],'SVM':[]}
else:
speaker_cluster[program_name][program_idx] = {program_date: {'interval':[], 'cluster_names':[],'labels':[],'GMM':[],'SVM':[]}}
else:
speaker_cluster[program_name] = {program_idx: {program_date: {'interval':[], 'cluster_names':[],'labels':[],'GMM':[],'SVM':[]}}}
return speaker_cluster
def parse_diarization_file(self, file_name):
"""
prog_name:
prog_id:
prog_date:
'interval':[ [],[],... ]
'cluster_names':[ ... ]
'GMM': [{'Ali':4,'Veli':2},{...}]
'SVM': ...
'labels': [[123,333], 567]
"""
N = 0
speaker_cluster = self.initialize_dict()
with open(file_name,'r') as f_:
prev_beg_time = -1
for line in f_:
N+=1
line_split = line.strip().split()
program_detail = line_split[0]
person_name = line_split[-1]
time_interval = map(float, parse_time_interval(line_split[2:4]))
pr_name1, pr_name2, program_date, program_idx = program_detail.split('_')
program_name = pr_name1+'_'+pr_name2
if not program_name in speaker_cluster or not program_idx in speaker_cluster[program_name] or not program_date in speaker_cluster[program_name][program_idx]:
continue
speaker_cluster[program_name][program_idx][program_date]['interval'].append([time_interval[0],sum(time_interval)])
speaker_cluster[program_name][program_idx][program_date]['cluster_names'].append(person_name)
if speaker_cluster[program_name][program_idx][program_date]['interval'][-1][0] == prev_beg_time:
print 'buldum speaker'
prev_beg_time = time_interval[0]
return speaker_cluster,N
def parse_mode_file(self, file_name):
mode_dict = self.initialize_dict()
with open(file_name,'r') as f_:
prev_beg_time = -1
for line in f_:
line_split = line.strip().split()
program_detail = line_split[0]
person_name = line_split[-1]
time_interval = map(float, parse_time_interval(line_split[1:3]))
pr_name1, pr_name2, program_date, program_idx = program_detail.split('_')
program_name = pr_name1+'_'+pr_name2
mode_dict[program_name][program_idx][program_date]['interval'].append(time_interval)
mode_dict[program_name][program_idx][program_date]['cluster_names'].append(person_name)
if mode_dict[program_name][program_idx][program_date]['interval'][-1][0] == prev_beg_time:
print 'buldum written'
prev_beg_time = time_interval[0]
return mode_dict
def _overlap_modes(self,interval_sp, interval_wr, wr_names):
overlap_list = []
for _ in xrange(len(interval_sp)):
overlap_list.append({})
interval_sp = np.array(interval_sp)
interval_wr = np.array(interval_wr)
for i, inter_sp in enumerate(interval_sp):
beg_sp,end_sp = inter_sp[0],inter_sp[1]
for inter_wr, name in zip(interval_wr,wr_names):
beg_wr,end_wr = inter_wr[0],inter_wr[1]
if beg_wr < beg_sp and end_wr <= beg_sp:
continue
if beg_wr >= end_sp and end_wr > end_sp:
break
if beg_wr <= beg_sp and end_wr <= end_sp:
overlap_list[i][name] = overlap_list[i].get(name,0) + end_wr - beg_sp
elif beg_wr >= beg_sp and end_wr <= end_sp:
overlap_list[i][name] = overlap_list[i].get(name,0) + end_wr - beg_wr
elif beg_wr >= beg_sp and end_wr >= end_sp:
overlap_list[i][name] = overlap_list[i].get(name,0) + end_sp - beg_wr
elif beg_wr <= beg_sp and end_wr >= end_sp:
overlap_list[i][name] = overlap_list[i].get(name,0) + end_sp - beg_sp
return overlap_list
def overlap_modes(self, speaker_cluster, mode_dict, mode_name = 'GMM'):
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
interval_sp = speaker_cluster[program_name][program_idx][program_date]['interval']
interval_wr = mode_dict[program_name][program_idx][program_date]['interval']
names_wr = mode_dict[program_name][program_idx][program_date]['cluster_names']
overlap_list = self._overlap_modes(interval_sp, interval_wr, names_wr)
speaker_cluster[program_name][program_idx][program_date][mode_name] = overlap_list
return speaker_cluster
def merge_clusters(self, speaker_cluster):
supervised_dict = {}
for program_name in speaker_cluster:
for program_idx in speaker_cluster[program_name]:
for program_date in speaker_cluster[program_name][program_idx]:
mode_GMM = speaker_cluster[program_name][program_idx][program_date]['GMM']
mode_SVM = speaker_cluster[program_name][program_idx][program_date]['SVM']
for i, cluster_name in enumerate(speaker_cluster[program_name][program_idx][program_date]['cluster_names']):
if supervised_dict.get(cluster_name,None) is None:
supervised_dict[cluster_name] = {'GMM':{},'SVM':{}}
for pers_name in mode_GMM[i]:
supervised_dict[cluster_name]['GMM'][pers_name] = supervised_dict[cluster_name]['GMM'].get(pers_name,0) + mode_GMM[i][pers_name]
for pers_name in mode_SVM[i]:
supervised_dict[cluster_name]['SVM'][pers_name] = supervised_dict[cluster_name]['SVM'].get(pers_name,0) + mode_SVM[i][pers_name]
return supervised_dict