def __init__(self, vocabSize, embSize, dSize):
super(TextEnc, self).__init__()
self.EmbLayer = module.Embed(
vocabSize, embSize) # (1 , T) -> (V, T), (e, V) * (V, T) -> (e, T)
self.seq = nn.ModuleList()
self.Conv1st = module.Cv(inChannel=embSize,
outChannel=dSize * 2,
kernelSize=1,
padding="SAME",
dilation=1,
activationF="ReLU")
self.seq.append(self.Conv1st)
self.seq.append(nn.Dropout(param.dr))
self.Conv2nd = module.Cv(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=1,
padding="SAME",
dilation=1,
activationF=None)
self.seq.append(self.Conv2nd)
self.seq.append(nn.Dropout(param.dr))
# HcTwice
for _ in range(2):
for d in range(4):
self.seq.append(
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=3**d))
self.seq.append(nn.Dropout(param.dr))
for _ in range(2):
self.seq.append(
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=1))
self.seq.append(nn.Dropout(param.dr))
for _ in range(2):
self.seq.append(
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=1,
padding="SAME",
dilation=1))
def __init__(self, fbinSize, dSize):
super(AudioEnc, self).__init__()
self.seq = nn.ModuleList()
self.Conv1st = module.Cv(inChannel=fbinSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF="ReLU")
self.seq.append(self.Conv1st)
self.seq.append(nn.Dropout(param.dr))
self.Conv2nd = module.Cv(inChannel=dSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF="ReLU")
self.seq.append(self.Conv2nd)
self.seq.append(nn.Dropout(param.dr))
self.Conv3rd = module.Cv(inChannel=dSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF=None)
self.seq.append(self.Conv3rd)
self.seq.append(nn.Dropout(param.dr))
for _ in range(2):
for d in range(4):
self.seq.append(
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=3**d))
self.seq.append(nn.Dropout(param.dr))
for _ in range(2):
self.seq.append(
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=3))
def __init__(self, upsamfbinSize, fbinSize, c, dSize):
super(SSRN, self).__init__()
self.seq = nn.ModuleList()
self.Conv1st = module.Cv(inChannel=fbinSize,
outChannel=c,
kernelSize=1,
padding="SAME",
dilation=1,
activationF=None)
self.seq.append(self.Conv1st)
self.seq.append(nn.Dropout(param.dr))
for d in range(2):
self.seq.append(
module.Hc(inChannel=c,
outChannel=c,
kernelSize=3,
padding="SAME",
dilation=3**d))
self.seq.append(nn.Dropout(param.dr))
for _ in range(2):
self.seq.append(
module.Dc(inChannel=c,
outChannel=c,
kernelSize=2,
padding="SAME",
dilation=1,
activationF=None))
self.seq.append(nn.Dropout(param.dr))
for _ in range(2):
self.seq.append(
module.Hc(inChannel=c,
outChannel=c,
kernelSize=3,
padding="SAME",
dilation=1))
self.seq.append(nn.Dropout(param.dr))
self.seq.append(
module.Cv(inChannel=c,
outChannel=2 * c,
kernelSize=1,
padding="SAME",
dilation=1,
activationF=None))
self.seq.append(nn.Dropout(param.dr))
for _ in range(2):
self.seq.append(
module.Hc(inChannel=2 * c,
outChannel=2 * c,
kernelSize=3,
padding="SAME",
dilation=1))
self.seq.append(nn.Dropout(param.dr))
self.seq.append(
module.Cv(inChannel=2 * c,
outChannel=upsamfbinSize,
kernelSize=1,
padding="SAME",
dilation=1,
activationF=None))
self.seq.append(nn.Dropout(param.dr))
for _ in range(2):
self.seq.append(
module.Cv(inChannel=upsamfbinSize,
outChannel=upsamfbinSize,
kernelSize=1,
padding="SAME",
dilation=1,
activationF="ReLU"))
self.seq.append(nn.Dropout(param.dr))
self.ConvLast = module.Cv(inChannel=upsamfbinSize,
outChannel=upsamfbinSize,
kernelSize=1,
padding="SAME",
dilation=1,
activationF="sigmoid")
self.seq.append(self.ConvLast)
def __init__(self, upsamfbinSize, fbinSize, c, dSize):
super(SSRN, self).__init__()
self.Conv1st = module.Cv(inChannel=fbinSize,
outChannel=c,
kernelSize=1,
padding="SAME",
dilation=1,
activationF=None)
self.Hc1 = nn.Sequential(
module.Hc(inChannel=c,
outChannel=c,
kernelSize=3,
padding="SAME",
dilation=1),
module.Hc(inChannel=c,
outChannel=c,
kernelSize=3,
padding="SAME",
dilation=3))
self.DcHcTwice1 = nn.Sequential(
module.Dc(inChannel=c,
outChannel=c,
kernelSize=2,
padding="SAME",
dilation=1,
activationF=None),
module.Hc(inChannel=c,
outChannel=c,
kernelSize=3,
padding="SAME",
dilation=1),
module.Hc(inChannel=c,
outChannel=c,
kernelSize=3,
padding="SAME",
dilation=3))
self.DcHcTwice2 = nn.Sequential(
module.Dc(inChannel=c,
outChannel=c,
kernelSize=2,
padding="SAME",
dilation=1,
activationF=None),
module.Hc(inChannel=c,
outChannel=c,
kernelSize=3,
padding="SAME",
dilation=1),
module.Hc(inChannel=c,
outChannel=c,
kernelSize=3,
padding="SAME",
dilation=3))
self.Conv2nd = module.Cv(inChannel=c,
outChannel=2 * c,
kernelSize=1,
padding="SAME",
dilation=1,
activationF=None)
self.HcTwice = nn.Sequential(
module.Hc(inChannel=2 * c,
outChannel=2 * c,
kernelSize=3,
padding="SAME",
dilation=1),
module.Hc(inChannel=2 * c,
outChannel=2 * c,
kernelSize=3,
padding="SAME",
dilation=1))
self.Conv3rd = module.Cv(inChannel=2 * c,
outChannel=upsamfbinSize,
kernelSize=1,
padding="SAME",
dilation=1,
activationF=None)
self.ConvTwice = nn.Sequential(
module.Cv(inChannel=upsamfbinSize,
outChannel=upsamfbinSize,
kernelSize=1,
padding="SAME",
dilation=1,
activationF="ReLU"),
module.Cv(inChannel=upsamfbinSize,
outChannel=upsamfbinSize,
kernelSize=1,
padding="SAME",
dilation=1,
activationF="ReLU"))
self.ConvLast = module.Cv(inChannel=upsamfbinSize,
outChannel=upsamfbinSize,
kernelSize=1,
padding="SAME",
dilation=1,
activationF="sigmoid")
def __init__(self, fbinSize, dSize):
super(AudioDec, self).__init__()
self.Conv1st = module.Cv(inChannel=dSize * 2,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF=None)
self.Hc1 = nn.Sequential(
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=1),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=3),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=9),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=27))
self.Hc2 = nn.Sequential(
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=1),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=1))
self.ConvThrice = nn.Sequential(
module.Cv(inChannel=dSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF="ReLU"),
module.Cv(inChannel=dSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF="ReLU"),
module.Cv(inChannel=dSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF="ReLU"))
self.ConvLast = module.Cv(inChannel=dSize,
outChannel=fbinSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF="sigmoid")
def __init__(self, vocabSize, embSize, dSize):
super(TextEnc, self).__init__()
self.EmbLayer = module.Embed(vocabSize, embSize)
self.Conv1st = module.Cv(inChannel=embSize,
outChannel=dSize * 2,
kernelSize=1,
padding="SAME",
dilation=1,
activationF="ReLU")
self.Conv2nd = module.Cv(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=1,
padding="SAME",
dilation=1,
activationF=None)
self.HcTwice1 = nn.Sequential(
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=1),
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=3),
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=9),
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=27))
self.HcTwice2 = nn.Sequential(
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=1),
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=3),
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=9),
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=27))
self.Hc3 = nn.Sequential(
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=1),
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=3,
padding="SAME",
dilation=1))
self.Hc4 = nn.Sequential(
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=1,
padding="SAME",
dilation=1),
module.Hc(inChannel=dSize * 2,
outChannel=dSize * 2,
kernelSize=1,
padding="SAME",
dilation=1))
def __init__(self, fbinSize, dSize):
super(AudioEnc, self).__init__()
self.Conv1st = module.Cv(inChannel=fbinSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF="ReLU")
self.Conv2nd = module.Cv(inChannel=dSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF="ReLU")
self.Conv3rd = module.Cv(inChannel=dSize,
outChannel=dSize,
kernelSize=1,
padding="causal",
dilation=1,
activationF=None)
self.HcTwice1 = nn.Sequential(
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=1),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=3),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=9),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=27))
self.HcTwice2 = nn.Sequential(
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=1),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=3),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=9),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=27))
self.Hc3 = nn.Sequential(
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=3),
module.Hc(inChannel=dSize,
outChannel=dSize,
kernelSize=3,
padding="causal",
dilation=3))