break
if ntot == 0:
xsum = sum(data)
xxsum = sum(data**2)
ntot += len(data)
else:
xsum += sum(data)
xxsum += sum(data**2)
ntot += len(data)
mnstd = numpy.array(
[xsum / ntot,
numpy.sqrt(xxsum / ntot - (xsum / ntot)**2)],
dtype="float32")
if xType in ("htkscp", "arkscp"):
trainXLst = nnet.loadscpf(trainXfile)
trainXall = numpy.array([], dtype=theano.config.floatX)
else:
print >> sys.stderr, "Right now, only htkscp and arkscp are supported as xType."
sys.exit(1)
# trainXLst = dict()
# trainXall = nnet.loadCFile(trainXfile, xType)
if yType in ("htkscp", "arkscp"):
trainYLst = nnet.loadscpf(trainYfile)
trainYall = numpy.array([], dtype=theano.config.floatX)
else:
trainYLst = dict()
trainYall = nnet.loadCFile(trainYfile, yType)
# prepare development data
xType = "htkscp"
bbsize = 0 # big batch size (0 means all data)
for ag in sys.argv[3:]:
if re.match("--xType=", ag):
xType = re.sub("--xType=", "", ag)
elif re.match("--bb=", ag):
bbsize = int(re.sub("--bb=", "", ag))
else:
print >> sys.stderr, "Error: Unknown option: " + ag
sys.exit(1)
# prepare training data
if xType in ("htkscp", "arkscp"):
trainXLst = nnet.loadscpf(datfilename) #读取文件内容
trainXall = numpy.array([], dtype=theano.config.floatX) #将trainXall变成需求的格式
else:
print >> sys.stderr, "Right now, only htkscp and arkscp are supported as xType."
sys.exit(1)
bblot = 0
ntot = 0
while True:
data = nnet.prepX(trainXLst, trainXall, numpy.array([]), xType, bbsize,
bblot, 0, 0, False)
bblot += 1
if len(data) == 0:
break
if ntot == 0:
xsum = sum(data)
if debug:
nn.debuginfo()
# Symbolic definition of transformation by Nnet
x = T.fmatrix("x") # fmatrix is float32
nnout = nn.forward(x)
featrans = theano.function(inputs=[x], outputs=nnout)
# prepare training data
if mvnormf != "":
mnstd=nnet.loadMVFile(mvnormf)
else:
mnstd=numpy.array([])
if xType in ("htkscp", "arkscp"):
trainXLst = nnet.loadscpf(datfile)
trainXall = numpy.array([], dtype=theano.config.floatX)
else:
print >> sys.stderr, "Right now, only htkscp and arkscp are supported as xType."
sys.exit(1)
idlist = nnet.scp2idlist(trainXLst, xType)
for ut in range(len(idlist)):
# set bb size to 1 and process one utterance by one utterance.
# no shuffling.
data = nnet.prepX(trainXLst, trainXall, mnstd,
xType, 1, ut, spl, trmspl, False)
mbnum=len(data)/mbsize
if mbnum*mbsize < len(data):
mbnum+=1