class readLabeledData():
logger = utils.getLogger()
def __init__(self,filename):
self.__filename = filename
def getLabledTuples(self):
uniqueTimeSentences = []
sentenceTimeTuples = {}
try:
with open(self.__filename,'rb') as csvtrainfile:
datareader = csv.DictReader(csvtrainfile)
for row in datareader:
sentence = row['labeledSentence']
sentence = utils.process(row['labeledSentence'])
reftime = row['refTime']
begintime = row['beginTime']
endtime = row['endTime']
if sentence not in uniqueTimeSentences:
uniqueTimeSentences.append(sentence)
sentenceTimeTuples[sentence] = (reftime,begintime,endtime)
except IOError as e:
readUnLabeledData.logger.error("Can not open file:{}, {}".format(self.__filename,e))
return sentenceTimeTuples
def apply_metric(self, metric_func):
logger = getLogger(os.path.join(self.ckpt_dir, self.metric+'_scores.log'), log_file=True)
all_scores_dir = os.path.join(self.ckpt_dir, 'all_scores')
if not os.path.isdir(all_scores_dir):
os.makedirs(all_scores_dir)
if not os.path.isdir(os.path.join(all_scores_dir, 'scores_arrays')):
os.makedirs(os.path.join(all_scores_dir, 'scores_arrays'))
if self.tt_list == '':
conditions = os.listdir(self.est_path)
else:
with open(self.tt_list, 'r') as f:
conditions = [t.strip().split('/')[-1].replace('.ex', '') for t in f.readlines()]
for condition in conditions:
mix_scores_array = []
est_scores_array = []
score_name = condition + '_' + self.metric
f = open(os.path.join(all_scores_dir, score_name + '.txt'), 'w')
count = 0
for filename in os.listdir(os.path.join(self.est_path, condition)):
if not filename.endswith('_mix.wav'):
continue
count += 1
mix, _ = sf.read(os.path.join(self.est_path, condition, filename), dtype=np.float32)
sph, _ = sf.read(os.path.join(self.est_path, condition, filename.replace('_mix', '_sph')), dtype=np.float32)
sph_est, _ = sf.read(os.path.join(self.est_path, condition, filename.replace('_mix', '_sph_est')), dtype=np.float32)
mix_score = metric_func(sph, mix)
est_score = metric_func(sph, sph_est)
f.write('utt {}: mix {:.4f}, est {:.4f}\n'.format(filename, mix_score, est_score))
f.flush()
mix_scores_array.append(mix_score)
est_scores_array.append(est_score)
mix_scores_array = np.array(mix_scores_array, dtype=np.float32)
est_scores_array = np.array(est_scores_array, dtype=np.float32)
f.write('========================================\n')
f.write('{} results: ({} utts)\n'.format(self.metric, count))
f.write('mix : {:.4f} +- {:.4f}\n'.format(np.mean(mix_scores_array), np.std(mix_scores_array)))
f.write('est : {:.4f} +- {:.4f}\n'.format(np.mean(est_scores_array), np.std(est_scores_array)))
f.close()
np.save(os.path.join(all_scores_dir, 'scores_arrays', score_name + '_mix.npy'), mix_scores_array)
np.save(os.path.join(all_scores_dir, 'scores_arrays', score_name + '_est.npy'), est_scores_array)
message = 'Evaluating {}: {} utts: '.format(condition, count) + \
'{} [ mix: {:.4f}, est: {:.4f} | delta: {:.4f} ]'.format(self.metric,
np.mean(mix_scores_array), np.mean(est_scores_array),
np.mean(est_scores_array)-np.mean(mix_scores_array))
logger.info(message)
class readUnLabeledData():
logger = utils.getLogger()
def __init__(self,filename,numTerminalTags):
self.__filename = filename
(self.__processedSentenceTimeTuples,self.__processedSentenceOriginalSentenceMap) = self.__returnProcessedTuples(numTerminalTags)
#utils.print_ds(self.__processedSentenceTimeTuples)
def __returnProcessedTuples(self,numTerminalTags):
uniqueTimeSentences = []
sentenceTimeTuples = {}
processedOldSentenceMap = {}
try:
with open(self.__filename,'rb' ) as csvtestfile:
datareader = csv.DictReader(csvtestfile)
for row in datareader:
#print row
originalSentence = utils.process(row['sentence'])
processedSentence = self._process(originalSentence)
for i in range(numTerminalTags):
processedSentence = "<~start~>_<~START~> "+processedSentence+" <~end~>_<~END~>"
reftime = row['refTime']
if processedSentence not in uniqueTimeSentences:
uniqueTimeSentences.append(processedSentence)
sentenceTimeTuples[processedSentence] = (reftime)
processedOldSentenceMap[processedSentence] = originalSentence
except IOError as e:
readUnLabeledData.logger.error("Can not open file :{}, {}".format(self.__filename,e))
return (sentenceTimeTuples, processedOldSentenceMap)
def getProcessedSentenceTuples(self):
return self.__processedSentenceTimeTuples
def getOrignialSentence(self,processedSentence):
return utils.getDictValue(self.__processedSentenceOriginalSentenceMap, processedSentence,None)
def _process(self,sentence):
sentence = sentence.strip()
words = sentence.split()
processedSentence = []
for word in words:
try:
int(word)
processedSentence.append("DIGIT")
except ValueError:
processedSentence.append(word.lower())
return " ".join(processedSentence)
class BeginEndTime:
logger = utils.getLogger()
def __init__(self, wordLabelTuples, referenceDate):
self.tagMap = {}
durationTuples = []
prefix = ""
for t in wordLabelTuples:
(label, word) = (t[0], t[1])
#Remove the angled Braces around the label
label = label[1:len(label) - 1]
#Get all TDPS and TDSS (prefix,suffix) tuples
if label == "TDSS":
if prefix != "":
durationTuple = (prefix, word)
durationTuples.append(durationTuple)
prefix = ""
continue
else:
BeginEndTime.logger.debug(
"Duration suffix without a prefix :{}".format(t))
if label == "TDPN" or label == "TDPS":
if prefix != "":
durationTuple = (prefix, '')
durationTuples.append(durationTuple)
prefix = word
continue
self.tagMap = utils.addKeyValue(self.tagMap, label, word)
if prefix != "":
durationTuples.append((prefix, ''))
if len(durationTuples) != 0:
self.tagMap["TIME_PERIOD"] = durationTuples
print "printing tagmap..."
print self.tagMap
self.tagMap = self.__validate(self.tagMap)
self.beginTime = self.__getBeginTime(self.tagMap, referenceDate)
self.endTime = self.__getEndTime(self.beginTime, self.tagMap)
# Convert hours to 24 hour format
def _convertTo24(self, tagMap, hourkey, ampmKey):
newValues = []
if hourkey in tagMap.keys():
values = tagMap[hourkey]
for val in values:
val = int(val)
if ampmKey in tagMap.keys():
ampm = 0 # 0 for am, 1 for pm
if "am" in tagMap[ampmKey]:
ampm = 0
else:
if "pm" in tagMap[ampmKey]:
ampm = 1
if ampm == 1:
if (val < 12):
val += 12
newValues.append(val)
tagMap[hourkey] = newValues
return tagMap
#Convert year digit to a standard 4 digit format
def _convertTo4digitYear(self, tagMap, yearKey):
if yearKey not in tagMap.keys():
return tagMap
newValues = []
t = date.today()
#Using the current century as the base
yr_2 = str(t.year)[:2]
values = tagMap[yearKey]
for val in values:
newVal = val
if len(val) == 2:
year = yr_2 + val
newVal = int(year)
newValues.append(newVal)
tagMap[yearKey] = newValues
return tagMap
# calculate time stamp for (prefix,suffix) tuples
def __calculateDurationOffset(self, timeTuple, endTime):
offsetTime = SpecialDaysCalendar.getTimeStruct(timeTuple, endTime)
if offsetTime != None:
return offsetTime
prefixstring = timeTuple[0]
suffixString = timeTuple[1]
prefix = None
if timeDurationOffsetAliases.has_key(timeTuple):
aliasTimeTuple = timeDurationOffsetAliases[timeTuple]
prefixstring = aliasTimeTuple[0]
suffixString = aliasTimeTuple[1]
try:
if prefixstring != "":
if prefixstring in timeStringToNumberAlias.keys():
prefix = timeStringToNumberAlias[prefixstring]
else:
prefix = int(prefixstring)
else:
prefix = int(timeTuple[0])
except ValueError as v:
BeginEndTime.logger.debug(
"TDP can not be converted to integer".format(v))
return endTime
else:
if suffixString in timeStringAlias.keys():
suffixString = timeStringAlias[suffixString] + "s"
offset = None
if suffixString == "days":
#Including current day as well
offset = timedelta(days=(prefix - 1))
if suffixString == "hours":
offset = timedelta(hours=prefix)
if suffixString == "seconds":
offset = timedelta(seconds=prefix)
if suffixString == "months":
offset = timedelta(months=prefix)
if suffixString == "minutes":
offset = timedelta(minutes=prefix)
stTime = datetime.datetime(endTime['year'], endTime['month'],
endTime['day'], endTime['hour'],
endTime['minute'],
endTime['second'])
newDate = stTime + offset
endTime['year'] = newDate.year
endTime['month'] = newDate.month
endTime['day'] = newDate.day
endTime['hour'] = newDate.hour
endTime['minute'] = newDate.minute
endTime['second'] = newDate.second
else:
BeginEndTime.logger.debug(
"Invalid suffix:{}".format(suffixString))
return endTime
def __calculateDayOffset(self, beginTime, daystring, isbeginDay=True):
if daystring == "today" or daystring == "eod" or daystring == "midnight" or daystring == "day":
if isbeginDay:
return beginTime
else:
beginTime['hour'] = 23
beginTime['minute'] = 59
beginTime['second'] = 59
return beginTime
if daystring == "now":
return beginTime
if daystring == "weekend":
if isbeginDay == True:
daystring = "friday"
else:
daystring = "sunday"
if daystring == "weekend":
if isbeginDay == True:
daystring = "monday"
else:
daystring = "friday"
days_ahead = 0
if daystring == "tomorrow":
days_ahead = 1
else:
try:
weekday = timeStringToNumberAlias[daystring]
days_ahead = weekday - beginTime["weekday"]
if days_ahead <= 0:
days_ahead += 7
except KeyError:
BeginEndTime.logger.debug(
"Invalid daystring:{}".format(daystring))
if isbeginDay == False:
#Default offset of one day
days_ahead = 1
else:
return beginTime
d = datetime.datetime(beginTime['year'], beginTime['month'],
beginTime['day'], beginTime['hour'],
beginTime['minute'], beginTime['second'])
newDate = d + datetime.timedelta(days_ahead)
beginTime['year'] = newDate.year
beginTime['month'] = newDate.month
beginTime['day'] = newDate.day
beginTime['hour'] = newDate.hour
beginTime['minute'] = newDate.minute
beginTime['second'] = newDate.second
beginTime['weekday'] = newDate.weekday()
if isbeginDay == False:
beginTime["hour"] = 23
beginTime["minute"] = 59
beginTime["second"] = 59
return beginTime
def __calculateHourOffset(self, beginTime, hourstring):
if hourstring == "now":
return beginTime
if hourstring == "tonight" or hourstring == "midnight":
beginTime['hour'] = 23
beginTime['minute'] = 59
beginTime['second'] = 59
if hourstring == "noon":
beginTime['hour'] = 12
beginTime['minute'] = 0
beginTime['second'] = 0
return beginTime
def __validate(self, tagMap):
for key in tagMap.keys():
validValues = []
for value in tagMap[key]:
if key[len(key) - 1] == "N":
try:
intvalue = int(value)
if key == 'BTPMN' or key == "ETPMN":
if intvalue > AllowedValues[
'maxMonth'] or intvalue < AllowedValues[
'minMonth']:
raise ValueError(
"Invalid Month :{}".format(value))
else:
validValues.append(value)
elif key == 'BTPmN' or key == 'ETPmN':
if intvalue > AllowedValues[
'maxMinute'] or intvalue < AllowedValues[
'minMinute']:
raise ValueError(
"Invalid Minute :{}".format(value))
else:
validValues.append(value)
elif key == 'BTPHN' or key == 'ETPHN':
if intvalue > AllowedValues[
'maxHour'] or intvalue < AllowedValues[
'minHour']:
raise ValueError(
"Invalid Hour :{}".format(value))
else:
validValues.append(value)
elif key == 'BTPDN' or key == 'ETPDN':
if intvalue > AllowedValues[
'maxDay'] or intvalue < AllowedValues[
'minDay']:
raise ValueError(
"Invalid Day {}".format(value))
else:
validValues.append(value)
elif key == 'BTPYN' or key == 'ETPYN':
if len(value) == 2:
if intvalue > AllowedValues[
'smaxYear'] or intvalue < AllowedValues[
'sminYear']:
raise ValueError(
"Invalid Year {}".format(value))
else:
validValues.append(value)
elif len(value) == 4:
if intvalue > AllowedValues[
'maxYear'] or intvalue < AllowedValues[
'minYear']:
raise ValueError(
"Invalid Year {}".format(value))
else:
validValues.append(value)
else:
raise ValueError(
"Invalid Year {}".format(value))
else:
validValues.append(value)
except ValueError as v:
BeginEndTime.logger.debug(
"Invalid Tagmap entries {}".format(v))
else:
validValues.append(value)
tagMap[key] = validValues
return tagMap
def __calculateRepeat(self, beginTime, repeat, repeatFrequency=30):
beginTimes = []
d = datetime.datetime(beginTime['year'], beginTime['month'],
beginTime['day'], beginTime['hour'],
beginTime['minute'], beginTime['second'])
for i in range(repeatFrequency):
days_ahead = 0
if repeat == "daily":
days_ahead = i
newDate = d + datetime.timedelta(days_ahead)
beginTimenew = copy.deepcopy(beginTime)
beginTimenew['year'] = newDate.year
beginTimenew['month'] = newDate.month
beginTimenew['day'] = newDate.day
beginTimenew['hour'] = newDate.hour
beginTimenew['minute'] = newDate.minute
beginTimenew['second'] = newDate.second
beginTimes.append(beginTimenew)
return beginTimes
def __validateDayMonthHour(self, beginTime):
try:
timestring = int(
datetime.datetime(beginTime['year'], beginTime['month'],
beginTime['day'], beginTime['hour'],
beginTime['minute']).strftime('%s'))
return True
except:
return False
# if time1 and time2 are equal it returns 0
# returns -1 if time1 is greater than time2 , returns 1 otherwise
# Compares to the minute level
# if one of them is none then it returns the other timestamp
# If one of them is contains invalid combination of day month year then
#it returns 1 if time2 is valid and -1 if time1 is valid
#if both are invalid raise an exception
def __compareTimes(self, time1, time2):
if time1 == None and time2 == None:
return 0
if time1 == None:
return 1
validTime1 = self.__validateDayMonthHour(time1)
validTime2 = self.__validateDayMonthHour(time2)
if validTime1 == False and validTime2 == False:
raise ValueError("invalid time1 and time2 given for comparisons")
if validTime1 == False:
return 1
if validTime2 == False:
return -1
if time2 == None:
return -1
if time1["year"] == time2["year"] and time1["month"] == time2[
"month"] and time1["hour"] == time2["hour"] and time1[
"minute"] == time2["minute"]:
return 0
if time1["year"] > time2["year"]:
return -1
else:
if time1['year'] == time2['year']:
if time1['month'] > time2['month']:
return -1
else:
if time1['month'] == time2['month']:
if time1['hour'] > time2['hour']:
return -1
else:
if time1['hour'] == time2['hour']:
if time1['minute'] > time2['minute']:
return -1
else:
return 1
else:
return 1
else:
return 1
else:
return 1
def __getBeginTime(self, tagMap, refDate):
beginTime = DefaultTime.getFromDateTime(refDate)
tagMap = self._convertTo24(tagMap, 'BTPHN', 'BTPAP')
tagMap = self._convertTo4digitYear(tagMap, 'BTPYN')
if 'BTPMS' in tagMap.keys():
monthstrings = tagMap['BTPMS']
months = []
for monthstring in monthstrings:
index = 0
try:
if monthstring not in timeStringAlias.keys():
BeginEndTime.logger.debug(
"Invalid month string, not in standard form:{}".
format(monthstring))
continue
stdmonthstring = timeStringAlias[monthstring]
mon = timeStringToNumberAlias[stdmonthstring]
months.append(mon)
#if index == 0:
# months = [mon]
# index += 1
#else:
# months.append(mon)
except KeyError:
BeginEndTime.logger.debug(
"Invalid monthstring {}".format(monthstring))
if len(months) > 0:
minMonth = min(months)
beginTime['month'] = minMonth
if 'BTPDS' in tagMap.keys():
index = 0
for daystring in tagMap['BTPDS']:
if daystring not in timeStringAlias.keys():
BeginEndTime.logger.debug(
"Invalid day string, not in standard form:{}".format(
daystring))
continue
stddaystring = timeStringAlias[daystring]
newBeginTime = self.__calculateDayOffset(
beginTime, stddaystring, True)
if index == 0:
beginTime = newBeginTime
index += 1
continue
if self.__compareTimes(newBeginTime, beginTime) == 1:
beginTime = newBeginTime
if 'BTPHS' in tagMap.keys():
index = 0
for hourstring in tagMap['BTPHS']:
if hourstring not in timeStringAlias.keys():
BeginEndTime.logger.debug(
"Invalid hour string, not in standard form:{}".format(
hourstring))
continue
stdhourstring = timeStringAlias[hourstring]
newBeginTime = self.__calculateHourOffset(
beginTime, stdhourstring)
if index == 0:
beginTime = newBeginTime
index += 1
continue
if self.__compareTimes(newBeginTime, beginTime) == 1:
beginTime = newBeginTime
for key in tagMap.keys():
if key in tagTimeAlias.keys():
if key[0] == 'B':
vals = []
for val in tagMap[key]:
try:
intval = int(val)
vals.append(intval)
except ValueError as v:
BeginEndTime.logger.debug(
"ERROR Converting {} for key {}, error:".
format(tagMap[key], key, v))
if len(set(vals)) > 1:
BeginEndTime.logger.debug(
"More than one string for month/year/day detected key:{} value:{}"
.format(key, vals))
if len(vals) > 0:
minval = min(vals)
beginTime[tagTimeAlias[key]] = minval
# Handling repetitions
#if 'BTPR' in tagMap.keys():
# repeatString = tagMap['BTPR']
# beginTime = self.__calculateRepeat(beginTime,repeatString)
return beginTime
def __getEndTime(self, beginTime, tagMap):
endTime = copy.deepcopy(beginTime)
tagMap = self._convertTo24(tagMap, 'ETPHN', 'ETPAP')
tagMap = self._convertTo4digitYear(tagMap, 'ETPYN')
if 'TIME_PERIOD' in tagMap.keys():
for timeTuple in tagMap['TIME_PERIOD']:
newEndTime = self.__calculateDurationOffset(timeTuple, endTime)
if self.__compareTimes(endTime, newEndTime) == 1:
endTime = newEndTime
if 'ETPMS' in tagMap.keys():
index = 0
for monthstring in tagMap['ETPMS']:
try:
if monthstring not in timeStringAlias.keys():
BeginEndTime.logger.debug(
"Invalid month string, not in standard form:{}".
format(monthstring))
continue
stdmonthstring = timeStringAlias[monthstring]
newMon = timeStringToNumberAlias[stdmonthstring]
if index == 0:
endTime['month'] = newMon
if index > 0:
if newMon > endTime['month']:
endTime['month'] = newMon
if tagMap.has_key('ETPDS') == False and tagMap.has_key(
'TIME_PERIOD') == False:
(weekday, maxdays) = calendar.monthrange(
endTime['year'], endTime['month'])
endTime['day'] = maxdays
index += 1
except KeyError:
BeginEndTime.logger.debug(
"Invalid monthstring:{}".format(monthstring))
if 'ETPDS' in tagMap.keys():
for daystring in tagMap['ETPDS']:
if daystring not in timeStringAlias.keys():
BeginEndTime.logger.debug(
"Invalid day string, not in standard form:{}".format(
daystring))
continue
stddaystring = timeStringAlias[daystring]
newEndTime = self.__calculateDayOffset(endTime, stddaystring,
False)
if self.__compareTimes(newEndTime, endTime) == -1:
endTime = newEndTime
if 'ETPHS' in tagMap.keys():
for hourstring in tagMap['ETPHS']:
if hourstring not in timeStringAlias.keys():
BeginEndTime.logger.debug(
"Invalid hour string, not in standard form:{}".format(
hourstring))
continue
hourstring = timeStringAlias[hourstring]
stdhourstring = timeStringAlias[hourstring]
newEndTime = self.__calculateHourOffset(endTime, stdhourstring)
if self.__compareTimes(newEndTime, endTime) == -1:
endTime = newEndTime
for key in tagMap.keys():
if key in tagTimeAlias.keys():
if key[0] == 'E':
vals = []
for val in tagMap[key]:
try:
intval = int(val)
vals.append(intval)
except ValueError as v:
BeginEndTime.logger.debug(
"ERROR Converting {} for key {}, error:".
format(tagMap[key], key, v))
if (len(set(vals)) > 1):
BeginEndTime.logger.debug(
"More than one value for month/day/year key:{} value:{}"
.format(key, vals))
if len(vals) > 0:
maxval = max(vals)
endTime[tagTimeAlias[key]] = maxval
if endTime["hour"] == 0 and endTime["minute"] == 0 and endTime[
"minute"] == 0:
endTime['hour'] = 23
endTime['minute'] = 59
endTime['second'] = 59
if self.__compareTimes(endTime, beginTime) == 1:
BeginEndTime.logger.debug("EndTime less than Start Time")
#Default One day offset
endTime = copy.deepcopy(beginTime)
endTime = self.__calculateDayOffset(endTime, "day", False)
if (endTime['year'] - beginTime['year']) > 1:
BeginEndTime.logger.debug(
"Endtime and beginTime are off by more than 1 year")
#endTime['year'] = beginTime['year']
return endTime
def getBeginTime(self):
return self.beginTime
def getEndTime(self):
return self.endTime
class labeledSentencesProcessor():
logger = utils.getLogger()
def __init__(self,sentences,numTerminalTags=1,numValidTags=2):
self._baseLabels = [
"<ETPYN>", "<ETPDS>","<ETPDN>", "<ETPMS>", "<ETPMN>", "<ETPHS>", "<ETPHN>", "<ETPmN>","<ETPmS>", "<ETPAP>","<ETPZ>",
"<BTPYS>", "<BTPYN>", "<BTPDS>","<BTPDN>", "<BTPMS>", "<BTPMN>", "<BTPHS>", "<BTPHN>", "<BTPmN>", "<BTPmS>", "<BTPAP>",
"<BTPZ>", "<TDPS>" , "<TDPN>" , "<TDSS>" , "<TDSN>" , "<TPRD>"
]
(self._labelWords, self._filesentences) = self.__parseLabeledData(sentences, numTerminalTags,numValidTags)
def _isbaseLabel(self,word_label):
for label in self._baseLabels:
#Remove the <> while comparing because training data does not have <>
label = label[1:len(label)-1]
if label in word_label:
return True
return False
def _returnWordLabel(self,word,addlabel,label):
monthL = ['january', 'february', 'march', 'april', 'may', 'june', 'july', 'august', 'september', 'october', 'november', 'december'];
monthS = ['jan', 'feb', 'mar', 'apr', 'may', 'june', 'july', 'aug', 'sept', 'oct', 'nov', 'dec'];
dayL = ['monday', 'tuesday', 'wednesday', 'thursday', 'friday', 'saturday', 'sunday']
dayS = ['mon', 'tue', 'wed', 'thu', 'fri', 'sat', 'sun']
try:
int(word)
if addlabel == True:
if label is None:
label = "<DIGIT>"
return ("DIGIT",label)
else:
label = "<NA>"
return ("DIGIT","<NA>")
except ValueError:
if addlabel == True:
if label is None:
label = "<"+word.upper()+">"
else:
return (word.lower(),"<NA>")
return(word.lower(),label)
def _process(self,word_label,index,numValidTags,words):
addlabel = False
if numValidTags == -1:
addlabel = True
else:
for i in range(index-(numValidTags),index+numValidTags+1,1):
if i>=0 and i<len(words):
if self._isbaseLabel(words[i]) ==True:
addlabel = True
if "_" in word_label:
(word, label) = word_label.split("_")
label = "<"+label+">"
if word == '' and label == '<>':
return ("<UDERSCORE>", "<NA>")
else:
if word == '' or label == '':
raise KeyError
return self._returnWordLabel(word, addlabel, label)
else:
return self._returnWordLabel(word_label, addlabel, None)
def __parseLabeledData(self,insentences,numTerminalTags,numValidTags):
sentences = []
labelWords = {}
for line in insentences:
line = line.strip()
words = line.split()
tokens = []
index = 0
for fullword in words:
(word,label) = self._process(fullword,index,numValidTags,words)
labelWords = utils.addKeyValue(labelWords,label,word)
token = word+"_"+label
tokens.append(token)
index = index + 1
processedline = " ".join(tokens)
for i in range(numTerminalTags):
processedline = "<~start~>_<~START~> "+processedline+" <~end~>_<~END~>"
sentences.append(processedline)
return (labelWords,sentences)
def gettrainLabels(self):
return self.getLabelWords().keys()
def getSentences(self):
return self._filesentences
def getLabelWordsMap(self):
return self._labelWords
def getBaseLabels(self):
return self._baseLabels
def getUnProcessedLines(self):
return self._unprocessedLines
def train(self, args):
with open(args.tr_list, 'r') as f:
self.tr_list = [line.strip() for line in f.readlines()]
self.tr_size = len(self.tr_list)
self.cv_file = args.cv_file
self.ckpt_dir = args.ckpt_dir
self.logging_period = args.logging_period
self.resume_model = args.resume_model
self.time_log = args.time_log
self.lr = args.lr
self.lr_decay_factor = args.lr_decay_factor
self.lr_decay_period = args.lr_decay_period
self.clip_norm = args.clip_norm
self.max_n_epochs = args.max_n_epochs
self.batch_size = args.batch_size
self.buffer_size = args.buffer_size
self.loss_log = args.loss_log
self.unit = args.unit
self.segment_size = args.segment_size
self.segment_shift = args.segment_shift
self.gpu_ids = tuple(map(int, args.gpu_ids.split(',')))
if len(self.gpu_ids) == 1 and self.gpu_ids[0] == -1:
# cpu only
self.device = torch.device('cpu')
else:
# gpu
self.device = torch.device('cuda:{}'.format(self.gpu_ids[0]))
if not os.path.isdir(self.ckpt_dir):
os.makedirs(self.ckpt_dir)
logger = getLogger(os.path.join(self.ckpt_dir, 'train.log'),
log_file=True)
# create data loaders for training and cross validation
tr_loader = AudioLoader(self.tr_list,
self.sample_rate,
self.unit,
self.segment_size,
self.segment_shift,
self.batch_size,
self.buffer_size,
self.in_norm,
mode='train')
cv_loader = AudioLoader(self.cv_file,
self.sample_rate,
unit='utt',
segment_size=None,
segment_shift=None,
batch_size=1,
buffer_size=10,
in_norm=self.in_norm,
mode='eval')
# create a network
net = Net()
logger.info('Model summary:\n{}'.format(net))
net = net.to(self.device)
if len(self.gpu_ids) > 1:
net = DataParallel(net, device_ids=self.gpu_ids)
# calculate model size
param_count = numParams(net)
logger.info('Trainable parameter count: {:,d} -> {:.2f} MB\n'.format(
param_count, param_count * 32 / 8 / (2**20)))
# net feeder
feeder = NetFeeder(self.device, self.win_size, self.hop_size)
# training criterion and optimizer
criterion = LossFunction()
optimizer = Adam(net.parameters(), lr=self.lr, amsgrad=False)
scheduler = lr_scheduler.StepLR(optimizer,
step_size=self.lr_decay_period,
gamma=self.lr_decay_factor)
# resume model if needed
if self.resume_model:
logger.info('Resuming model from {}'.format(self.resume_model))
ckpt = CheckPoint()
ckpt.load(self.resume_model, self.device)
state_dict = {}
for key in ckpt.net_state_dict:
if len(self.gpu_ids) > 1:
state_dict['module.' + key] = ckpt.net_state_dict[key]
else:
state_dict[key] = ckpt.net_state_dict[key]
net.load_state_dict(state_dict)
optimizer.load_state_dict(ckpt.optim_state_dict)
ckpt_info = ckpt.ckpt_info
logger.info(
'model info: epoch {}, iter {}, cv_loss - {:.4f}\n'.format(
ckpt.ckpt_info['cur_epoch'] + 1,
ckpt.ckpt_info['cur_iter'] + 1, ckpt.ckpt_info['cv_loss']))
else:
logger.info('Training from scratch...\n')
ckpt_info = {
'cur_epoch': 0,
'cur_iter': 0,
'tr_loss': None,
'cv_loss': None,
'best_loss': float('inf')
}
start_iter = 0
# train model
while ckpt_info['cur_epoch'] < self.max_n_epochs:
accu_tr_loss = 0.
accu_n_frames = 0
net.train()
for n_iter, egs in enumerate(tr_loader):
n_iter += start_iter
mix = egs['mix']
sph = egs['sph']
n_samples = egs['n_samples']
mix = mix.to(self.device)
sph = sph.to(self.device)
n_samples = n_samples.to(self.device)
n_frames = countFrames(n_samples, self.win_size, self.hop_size)
start_time = timeit.default_timer()
# prepare features and labels
feat, lbl = feeder(mix, sph)
loss_mask = lossMask(shape=lbl.shape,
n_frames=n_frames,
device=self.device)
# forward + backward + optimize
optimizer.zero_grad()
with torch.enable_grad():
est = net(feat)
loss = criterion(est, lbl, loss_mask, n_frames)
loss.backward()
if self.clip_norm >= 0.0:
clip_grad_norm_(net.parameters(), self.clip_norm)
optimizer.step()
# calculate loss
running_loss = loss.data.item()
accu_tr_loss += running_loss * sum(n_frames)
accu_n_frames += sum(n_frames)
end_time = timeit.default_timer()
batch_time = end_time - start_time
if self.time_log:
with open(self.time_log, 'a+') as f:
print(
'Epoch [{}/{}], Iter [{}], tr_loss = {:.4f} / {:.4f}, batch_time (s) = {:.4f}'
.format(ckpt_info['cur_epoch'] + 1,
self.max_n_epochs, n_iter, running_loss,
accu_tr_loss / accu_n_frames, batch_time),
file=f)
f.flush()
else:
print(
'Epoch [{}/{}], Iter [{}], tr_loss = {:.4f} / {:.4f}, batch_time (s) = {:.4f}'
.format(ckpt_info['cur_epoch'] + 1, self.max_n_epochs,
n_iter, running_loss,
accu_tr_loss / accu_n_frames, batch_time),
flush=True)
if (n_iter + 1) % self.logging_period == 0:
avg_tr_loss = accu_tr_loss / accu_n_frames
avg_cv_loss = self.validate(net, cv_loader, criterion,
feeder)
net.train()
ckpt_info['cur_iter'] = n_iter
is_best = True if avg_cv_loss < ckpt_info[
'best_loss'] else False
ckpt_info[
'best_loss'] = avg_cv_loss if is_best else ckpt_info[
'best_loss']
latest_model = 'latest.pt'
best_model = 'best.pt'
ckpt_info['tr_loss'] = avg_tr_loss
ckpt_info['cv_loss'] = avg_cv_loss
if len(self.gpu_ids) > 1:
ckpt = CheckPoint(ckpt_info, net.module.state_dict(),
optimizer.state_dict())
else:
ckpt = CheckPoint(ckpt_info, net.state_dict(),
optimizer.state_dict())
logger.info('Saving checkpoint into {}'.format(
os.path.join(self.ckpt_dir, latest_model)))
if is_best:
logger.info('Saving checkpoint into {}'.format(
os.path.join(self.ckpt_dir, best_model)))
logger.info(
'Epoch [{}/{}], ( tr_loss: {:.4f} | cv_loss: {:.4f} )\n'
.format(ckpt_info['cur_epoch'] + 1, self.max_n_epochs,
avg_tr_loss, avg_cv_loss))
model_path = os.path.join(self.ckpt_dir, 'models')
if not os.path.isdir(model_path):
os.makedirs(model_path)
ckpt.save(os.path.join(model_path, latest_model), is_best,
os.path.join(model_path, best_model))
lossLog(os.path.join(self.ckpt_dir, self.loss_log), ckpt,
self.logging_period)
accu_tr_loss = 0.
accu_n_frames = 0
if n_iter + 1 == self.tr_size // self.batch_size:
start_iter = 0
ckpt_info['cur_iter'] = 0
break
ckpt_info['cur_epoch'] += 1
scheduler.step() # learning rate decay
return
def test(self, args):
with open(args.tt_list, 'r') as f:
self.tt_list = [line.strip() for line in f.readlines()]
self.model_file = args.model_file
self.ckpt_dir = args.ckpt_dir
self.est_path = args.est_path
self.write_ideal = args.write_ideal
self.gpu_ids = tuple(map(int, args.gpu_ids.split(',')))
if len(self.gpu_ids) == 1 and self.gpu_ids[0] == -1:
# cpu only
self.device = torch.device('cpu')
else:
# gpu
self.device = torch.device('cuda:{}'.format(self.gpu_ids[0]))
if not os.path.isdir(self.ckpt_dir):
os.makedirs(self.ckpt_dir)
logger = getLogger(os.path.join(self.ckpt_dir, 'test.log'),
log_file=True)
# create a network
net = Net()
logger.info('Model summary:\n{}'.format(net))
net = net.to(self.device)
# calculate model size
param_count = numParams(net)
logger.info('Trainable parameter count: {:,d} -> {:.2f} MB\n'.format(
param_count, param_count * 32 / 8 / (2**20)))
# training criterion and optimizer
criterion = LossFunction()
# net feeder
feeder = NetFeeder(self.device, self.win_size, self.hop_size)
# resynthesizer
resynthesizer = Resynthesizer(self.device, self.win_size,
self.hop_size)
# load model
logger.info('Loading model from {}'.format(self.model_file))
ckpt = CheckPoint()
ckpt.load(self.model_file, self.device)
net.load_state_dict(ckpt.net_state_dict)
logger.info('model info: epoch {}, iter {}, cv_loss - {:.4f}\n'.format(
ckpt.ckpt_info['cur_epoch'] + 1, ckpt.ckpt_info['cur_iter'] + 1,
ckpt.ckpt_info['cv_loss']))
net.eval()
for i in range(len(self.tt_list)):
# create a data loader for testing
tt_loader = AudioLoader(self.tt_list[i],
self.sample_rate,
unit='utt',
segment_size=None,
segment_shift=None,
batch_size=1,
buffer_size=10,
in_norm=self.in_norm,
mode='eval')
logger.info('[{}/{}] Estimating on {}'.format(
i + 1, len(self.tt_list), self.tt_list[i]))
est_subdir = os.path.join(
self.est_path,
self.tt_list[i].split('/')[-1].replace('.ex', ''))
if not os.path.isdir(est_subdir):
os.makedirs(est_subdir)
accu_tt_loss = 0.
accu_n_frames = 0
for k, egs in enumerate(tt_loader):
mix = egs['mix']
sph = egs['sph']
n_samples = egs['n_samples']
n_frames = countFrames(n_samples, self.win_size, self.hop_size)
mix = mix.to(self.device)
sph = sph.to(self.device)
feat, lbl = feeder(mix, sph)
with torch.no_grad():
loss_mask = lossMask(shape=lbl.shape,
n_frames=n_frames,
device=self.device)
est = net(feat)
loss = criterion(est, lbl, loss_mask, n_frames)
accu_tt_loss += loss.data.item() * sum(n_frames)
accu_n_frames += sum(n_frames)
sph_idl = resynthesizer(lbl, mix)
sph_est = resynthesizer(est, mix)
# save estimates
mix = mix[0].cpu().numpy()
sph = sph[0].cpu().numpy()
sph_est = sph_est[0].cpu().numpy()
sph_idl = sph_idl[0].cpu().numpy()
mix, sph, sph_est, sph_idl = wavNormalize(
mix, sph, sph_est, sph_idl)
sf.write(os.path.join(est_subdir, '{}_mix.wav'.format(k)), mix,
self.sample_rate)
sf.write(os.path.join(est_subdir, '{}_sph.wav'.format(k)), sph,
self.sample_rate)
sf.write(os.path.join(est_subdir, '{}_sph_est.wav'.format(k)),
sph_est, self.sample_rate)
if self.write_ideal:
sf.write(
os.path.join(est_subdir, '{}_sph_idl.wav'.format(k)),
sph_idl, self.sample_rate)
avg_tt_loss = accu_tt_loss / accu_n_frames
logger.info('loss: {:.4f}'.format(avg_tt_loss))
return
import argparse
import pprint
import torch
from utils.models import Model
from utils.utils import getLogger
logger = getLogger(__name__)
def main():
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
# parse the configuarations
parser = argparse.ArgumentParser(
description='Additioal configurations for testing',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
'--gpu_ids',
type=str,
default='-1',
help=
'IDs of GPUs to use (please use `,` to split multiple IDs); -1 means CPU only'
)
parser.add_argument('--tt_list',
type=str,
required=True,
help='Path to the list of testing files')
parser.add_argument('--ckpt_dir',