def make_questions(context_skeleton, qformat, generic=True, HHEd_fix=False):
context_dict = context_skeleton.added_contexts
if qformat not in ["NN", "HMM"]:
print "Error: Invalid question format! " + qformat
print "Must be either HMM or NN!"
sys.exit()
questions = []
if generic == False:
for key in context_dict:
qtype = getattr(context_skeleton, key)
#Setting it removes duplicates and sorting them is important for HMM question creation
#(sorting them is also prettier for the NN)
vals = list(set(context_dict[key]))
vals.sort()
if qtype == None:
pass
elif qtype == "bool":
for val in vals:
#If this is current phoneme id we apply the HHEd_fix
if HHEd_fix == True and key == "cp":
val = "-"+str(val)+"+"
if qformat == "HMM":
questions.append("QS \""+key+"-"+str(val)+"\" {*|"+key+":"+str(val)+"|*}")
elif qformat == "NN":
questions.append("LQ 0 \""+key+"-"+str(val)+"\" {*|"+key+":"+str(val)+"|*}")
elif "float" in qtype:
if qformat == "HMM":
questions += make_hmm_relational_qs(vals, key, qtype)
elif qformat == "NN":
for val in vals:
#HHEd's pattern matching for both NN and HMM's uses '.' as a special
#character. We therefore use the int version of the float value in the
#search pattern. But we have to check for xx first.
if val == "xx":
questions.append("LQ 1 0.0 \""+key+"-xx\" {*|"+key+":xx|*}")
else:
questions.append("LQ 1 "+str(val)+" \""+key+"-"+strintify(float(val))+"\" {*|"+key+":"+strintify(float(val))+"|*}")
elif "int" in qtype:
if qformat == "HMM":
questions += make_hmm_relational_qs(vals, key, qtype)
elif qformat == "NN":
for val in vals:
#The NN relies on floats for the actual value so we use that there
#and the int in the naming. But we have to check for xx first.
if val == "xx":
questions.append("LQ 1 0.0 \""+key+"-xx\" {*|"+key+":xx|*}")
else:
questions.append("LQ 1 "+strfloatify(val)+" \""+key+"-"+str(val)+"\" {*|"+key+":"+str(val)+"|*}")
else:
print "Error: Odd question type, should not exist - "+qtype
sys.exit()
else:
print "Error: Not implemented yet! (Outputting generic question set)"
return questions
def make_hmm_relational_qs(values, key, qtype):
questions = []
#Add xx question if appropriate else ignore
if "xx" in values:
if "xx" in qtype:
questions.append("QS \"" + key + "-xx\" {*|" + key + ":xx|*}")
else:
raise SiReError(
"xx in values but not in qtype {0} for key {1} - why?".format(
qtype, key))
values.remove("xx")
for i, val in enumerate(values):
if "float" in qtype:
val = strintify(float(val))
questions.append("QS \"" + key + "-" + str(val) + "\" {*|" + key +
":" + str(val) + "|*}")
#If val is more than one we make a less than question
#If we count 0 then we start at 0
if "0" in qtype:
start = 0
else:
start = 1
if int(val) > start:
#Prep the less than string
s = "QS \"" + key + "<=" + str(val) + "\" {"
#Make the less than string
#Get tens and remainder
tens = int(val) / 10
remainder = int(val) % 10
if tens > 0:
#Make singles
for n in range(start, 10):
s += "*|" + key + ":" + str(n) + "|*,"
#Make tens
for n in range(1, tens):
s += "*|" + key + ":" + str(n) + "?|*,"
for n in range(remainder + 1):
if n != remainder:
s += "*|" + key + ":" + str(tens) + str(n) + "|*,"
else:
s += "*|" + key + ":" + str(tens) + str(n) + "|*}"
questions.append(s)
else:
#Just make singles
for n in range(start, int(val) + 1):
s += "*|" + key + ":" + str(n) + "|*"
if n != int(val):
s += ","
else:
s += "}"
questions.append(s)
return questions
def get_context_string(self, HHEd_fix=False):
s = ""
s += str(self.added_contexts["start"]) + " "
s += str(self.added_contexts["end"]) + " "
for key in self.added_contexts:
if key in ["start", "end"]:
continue
if "float" in getattr(self, key):
s += "|"+key+":"+context_utils.strintify(float(self.added_contexts[key]))
else:
s += "|"+key+":"+str(self.added_contexts[key])
if HHEd_fix == True:
s = s.replace("cp:"+self.added_contexts["cp"], "cp:-"+self.added_contexts["cp"]+"+")
return s+"|"
def make_hmm_relational_qs(values, key, qtype):
questions = []
#Add xx question if appropriate else ignore
if "xx" in values:
if "xx" in qtype:
questions.append("QS \""+key+"-xx\" {*|"+key+":xx|*}")
else:
print "Error: xx in values but not in qtype {0} for key {1} - why?".format(qtype,key)
sys.exit()
values.remove("xx")
for i, val in enumerate(values):
if "float" in qtype:
val = strintify(float(val))
questions.append("QS \""+key+"-"+str(val)+"\" {*|"+key+":"+str(val)+"|*}")
#If val is more than one we make a less than question
#If we count 0 then we start at 0
if "0" in qtype:
start = 0
else:
start = 1
if int(val) > start:
#Prep the less than string
s = "QS \""+key+"<="+str(val)+"\" {"
#Make the less than string
#Get tens and remainder
tens = int(val)/10
remainder = int(val)%10
if tens > 0:
#Make singles
for n in range(start, 10):
s += "*|"+key+":"+str(n)+"|*,"
#Make tens
for n in range(1, tens):
s += "*|"+key+":"+str(n)+"?|*,"
for n in range(remainder+1):
if n != remainder:
s += "*|"+key+":"+str(tens)+str(n)+"|*,"
else:
s += "*|"+key+":"+str(tens)+str(n)+"|*}"
questions.append(s)
else:
#Just make singles
for n in range(start, int(val)+1):
s += "*|"+key+":"+str(n)+"|*"
if n != int(val):
s += ","
else:
s += "}"
questions.append(s)
return questions
def get_context_string(self, HHEd_fix=False):
s = ""
s += str(self.added_contexts["start"]) + " "
s += str(self.added_contexts["end"]) + " "
for key in self.added_contexts:
if key in ["start", "end"]:
continue
if "float" in getattr(self, key):
if self.added_contexts[key] == "xx" and "xx" in getattr(self, key):
s += "|"+key+":"+str(self.added_contexts[key])
else:
s += "|"+key+":"+context_utils.strintify(float(self.added_contexts[key]))
else:
s += "|"+key+":"+str(self.added_contexts[key])
if HHEd_fix == True:
s = s.replace("cp:"+self.added_contexts["cp"], "cp:-"+self.added_contexts["cp"]+"+")
return s+"|"
def make_questions(context_skeleton, qformat, generic=True, HHEd_fix=False, utt=False):
context_dict = context_skeleton.added_contexts
#We write out each context not used just for checks.
#TODO: Make it possible to throw an exception if context not used.
#TODO: Currently we ignore this when making GV contexts.
if utt == False:
for context in vars(context_skeleton).keys():
if context not in context_dict and context != "added_contexts":
print context_dict.keys()
print "Warning! Context ({0}) not used!".format(context)
if qformat not in ["Nitech_NN", "HMM", "CSTR_NN"]:
raise SiReError("Invalid question format ({0})! Must be either HMM, Nitech_NN or CSTR_NN!".format(qformat))
questions = []
if generic == False:
for key in context_dict:
qtype = getattr(context_skeleton, key)
#Setting it removes duplicates and sorting them is important for HMM question creation
#(sorting them is also prettier for the NN)
vals = list(set(context_dict[key]))
vals.sort()
if qtype == None:
pass
elif qtype == "bool":
for val in vals:
#If this is current phoneme id we apply the HHEd_fix
if HHEd_fix == True and key == "cp":
val = "-"+str(val)+"+"
if qformat == "HMM" or qformat == "CSTR_NN":
questions.append("QS \""+key+"-"+str(val)+"\" {*|"+key+":"+str(val)+"|*}")
elif qformat == "Nitech_NN":
questions.append("LQ 0 \""+key+"-"+str(val)+"\" {*|"+key+":"+str(val)+"|*}")
elif "float" in qtype:
if qformat == "HMM":
questions += make_hmm_relational_qs(vals, key, qtype)
elif qformat == "Nitech_NN":
for val in vals:
#HHEd's pattern matching for both NN and HMM's uses '.' as a special
#character. We therefore use the int version of the float value in the
#search pattern. But we have to check for xx first.
if val == "xx":
questions.append("LQ 1 0.0 \""+key+"-xx\" {*|"+key+":xx|*}")
else:
questions.append("LQ 1 "+str(val)+" \""+key+"-"+strintify(float(val))+"\" {*|"+key+":"+strintify(float(val))+"|*}")
elif qformat == "CSTR_NN":
#For CSTR's DNN system we just need to specify the question and not one for each value of the question if float or int.
#Note that floats needs to be strintified in the actual question.
questions.append("CQS \""+key+"\" {*|"+key+":(\d+)|*}")
elif "int" in qtype:
if qformat == "HMM":
questions += make_hmm_relational_qs(vals, key, qtype)
elif qformat == "Nitech_NN":
for val in vals:
#The NN relies on floats for the actual value so we use that there
#and the int in the naming. But we have to check for xx first.
if val == "xx":
questions.append("LQ 1 0.0 \""+key+"-xx\" {*|"+key+":xx|*}")
else:
questions.append("LQ 1 "+strfloatify(int(val))+" \""+key+"-"+str(val)+"\" {*|"+key+":"+str(val)+"|*}")
elif qformat == "CSTR_NN":
#For CSTR's DNN system we just need to specify the question and not one for each value of the question if float or int.
#Note that floats needs to be strintified in the actual question.
questions.append("CQS \""+key+"\" {*|"+key+":(\d+)|*}")
else:
raise SiReError("Odd question type, should not exist - {0}".format(qtype))
else:
raise SiReError("Not implemented yet! (Outputting generic question set)")
return questions
def make_questions(context_skeleton,
qformat,
generic=True,
HHEd_fix=False,
utt=False):
context_dict = context_skeleton.added_contexts
#We write out each context not used just for checks.
#TODO: Make it possible to throw an exception if context not used.
#TODO: Currently we ignore this when making GV contexts.
if utt == False:
for context in vars(context_skeleton).keys():
if context not in context_dict and context != "added_contexts":
print context_dict.keys()
print "Warning! Context ({0}) not used!".format(context)
if qformat not in ["Nitech_NN", "HMM", "CSTR_NN"]:
raise SiReError(
"Invalid question format ({0})! Must be either HMM, Nitech_NN or CSTR_NN!"
.format(qformat))
questions = []
if generic == False:
for key in context_dict:
qtype = getattr(context_skeleton, key)
#Setting it removes duplicates and sorting them is important for HMM question creation
#(sorting them is also prettier for the NN)
vals = list(set(context_dict[key]))
vals.sort()
if qtype == None:
pass
elif qtype == "bool":
for val in vals:
#If this is current phoneme id we apply the HHEd_fix
if HHEd_fix == True and key == "cp":
val = "-" + str(val) + "+"
if qformat == "HMM" or qformat == "CSTR_NN":
questions.append("QS \"" + key + "-" + str(val) +
"\" {*|" + key + ":" + str(val) +
"|*}")
elif qformat == "Nitech_NN":
questions.append("LQ 0 \"" + key + "-" + str(val) +
"\" {*|" + key + ":" + str(val) +
"|*}")
elif "float" in qtype:
if qformat == "HMM":
questions += make_hmm_relational_qs(vals, key, qtype)
elif qformat == "Nitech_NN":
for val in vals:
#HHEd's pattern matching for both NN and HMM's uses '.' as a special
#character. We therefore use the int version of the float value in the
#search pattern. But we have to check for xx first.
if val == "xx":
questions.append("LQ 1 0.0 \"" + key +
"-xx\" {*|" + key + ":xx|*}")
else:
questions.append("LQ 1 " + str(val) + " \"" + key +
"-" + strintify(float(val)) +
"\" {*|" + key + ":" +
strintify(float(val)) + "|*}")
elif qformat == "CSTR_NN":
#For CSTR's DNN system we just need to specify the question and not one for each value of the question if float or int.
#Note that floats needs to be strintified in the actual question.
questions.append("CQS \"" + key + "\" {*|" + key +
":(\d+)|*}")
elif "int" in qtype:
if qformat == "HMM":
questions += make_hmm_relational_qs(vals, key, qtype)
elif qformat == "Nitech_NN":
for val in vals:
#The NN relies on floats for the actual value so we use that there
#and the int in the naming. But we have to check for xx first.
if val == "xx":
questions.append("LQ 1 0.0 \"" + key +
"-xx\" {*|" + key + ":xx|*}")
else:
questions.append("LQ 1 " + strfloatify(int(val)) +
" \"" + key + "-" + str(val) +
"\" {*|" + key + ":" + str(val) +
"|*}")
elif qformat == "CSTR_NN":
#For CSTR's DNN system we just need to specify the question and not one for each value of the question if float or int.
#Note that floats needs to be strintified in the actual question.
questions.append("CQS \"" + key + "\" {*|" + key +
":(\d+)|*}")
else:
raise SiReError(
"Odd question type, should not exist - {0}".format(qtype))
else:
raise SiReError(
"Not implemented yet! (Outputting generic question set)")
return questions