def __init__(self,
size_vocab,
size,
depth=1,
recur_depth=1,
bidirectional=False,
filter_length=6,
filter_size=64,
stride=2,
drop_i=0.75,
drop_s=0.25):
super(Encoder, self).__init__()
util.autoassign(locals())
self.h0 = torch.autograd.Variable(torch.zeros(self.depth, 1,
self.size))
self.Conv = conv.Convolution1D(self.size_vocab,
self.filter_length,
self.filter_size,
stride=self.stride)
# self.RNN = nn.GRU(self.filter_size, self.size, self.depth, batch_first=True)
self.RNN = stacked_gru.StackedGRU(self.filter_size,
self.size,
self.depth,
bidirectional=bidirectional,
residual=True,
batch_first=True)
def __init__(self,
size_vocab,
size,
depth=1,
recur_depth=1,
filter_length=6,
filter_size=64,
stride=2,
drop_i=0.75,
drop_s=0.25,
residual=False,
seed=1):
super(Encoder, self).__init__()
util.autoassign(locals())
#self.h0 = torch.autograd.Variable(torch.zeros(self.depth, 1, self.size))
self.Conv = conv.Convolution1D(self.size_vocab,
self.filter_length,
self.filter_size,
stride=self.stride)
self.RHN = rhn.StackedRHNH0(self.filter_size,
self.size,
depth=self.depth,
recur_depth=self.recur_depth,
drop_i=self.drop_i,
drop_s=self.drop_s,
residual=self.residual,
seed=self.seed)
def __init__(self,
size_vocab,
size,
nb_conv_layer=1,
depth=1,
filter_length=6,
filter_size=[64],
stride=2,
dropout_p=0.0,
relu=False,
maxpool=False,
bidirectional=False):
super(SpeechEncoderBottom, self).__init__()
util.autoassign(locals())
layers = []
size_in = self.size_vocab
for i_conv in range(0, self.nb_conv_layer):
layers.append(
conv.Convolution1D(size_in,
self.filter_length,
self.filter_size[i_conv],
stride=self.stride,
maxpool=self.maxpool))
if self.relu:
layers.append(nn.ReLU(True))
size_in = self.filter_size[i_conv]
self.Conv = nn.Sequential(*layers)
if self.depth > 0:
# TODO: LSTM/GRU?
if self.bidirectional:
self.h0 = torch.autograd.Variable(
torch.zeros(self.depth * 2, 1, self.size))
self.c0 = torch.autograd.Variable(
torch.zeros(self.depth * 2, 1, self.size))
else:
self.h0 = torch.autograd.Variable(
torch.zeros(self.depth, 1, self.size))
self.c0 = torch.autograd.Variable(
torch.zeros(self.depth, 1, self.size))
self.Dropout = nn.Dropout(p=self.dropout_p)
# TODO: LSTM/GRU?
#self.RNN = nn.GRU(self.filter_size[self.nb_conv_layer - 1],
# self.size, self.depth, batch_first=True,
# bidirectional=self.bidirectional)
self.RNN = nn.LSTM(self.filter_size[self.nb_conv_layer - 1],
self.size,
self.depth,
batch_first=True,
bidirectional=self.bidirectional)
def __init__(self,
size_vocab,
size,
depth=1,
filter_length=6,
filter_size=64,
stride=2,
dropout_p=0.0):
super(SpeechEncoderBottomStack, self).__init__()
util.autoassign(locals())
self.Conv = conv.Convolution1D(self.size_vocab,
self.filter_length,
self.filter_size,
stride=self.stride)
if self.depth > 0:
self.Dropout = nn.Dropout(p=self.dropout_p)
self.RNN = GRUStack(self.filter_size, self.size, self.depth)
def __init__(self,
size_vocab,
size,
depth=1,
filter_length=6,
filter_size=64,
stride=2,
dropout_p=0.0):
super(SpeechEncoderBottom, self).__init__()
util.autoassign(locals())
self.Conv = conv.Convolution1D(self.size_vocab,
self.filter_length,
self.filter_size,
stride=self.stride)
if self.depth > 0:
self.h0 = torch.autograd.Variable(
torch.zeros(self.depth, 1, self.size))
self.Dropout = nn.Dropout(p=self.dropout_p)
self.RNN = nn.GRU(self.filter_size,
self.size,
self.depth,
batch_first=True)
def __init__(self,
size_vocab,
size,
depth=1,
filter_length=6,
filter_size=64,
stride=2,
residual=False):
super(Encoder, self).__init__()
util.autoassign(locals())
self.h0 = torch.autograd.Variable(torch.zeros(self.depth, 1,
self.size))
self.Conv = conv.Convolution1D(self.size_vocab,
self.filter_length,
self.filter_size,
stride=self.stride,
padding=0)
self.RNN = nn.GRU(self.filter_size,
self.size,
self.depth,
batch_first=True)
def __init__(self,
size_feature,
size,
depth=1,
filter_length=6,
filter_size=64,
stride=2,
size_attn=512,
dropout_p=0.0):
super(Encoder, self).__init__()
util.autoassign(locals())
self.h0 = torch.autograd.Variable(torch.zeros(self.depth, 1,
self.size))
self.Dropout = nn.Dropout(p=self.dropout_p)
self.Conv = conv.Convolution1D(self.size_feature,
self.filter_length,
self.filter_size,
stride=self.stride,
padding=0)
self.RNN = nn.GRU(self.filter_size,
self.size,
self.depth,
batch_first=True)
self.Attn = attention.SelfAttention(self.size, size=self.size_attn)