def __init__(self,
idim,
dims,
dropout=0.1,
initialW=None,
initial_bias=None):
super(LinearSampling, self).__init__()
stvd = 1. / np.sqrt(dims)
self.dropout = dropout
with self.init_scope():
self.linear = L.Linear(idim,
dims,
initialW=initialW(scale=stvd),
initial_bias=initial_bias(scale=stvd))
self.pe = PositionalEncoding(dims, dropout)
def __init__(self,
channels,
idim,
dims,
dropout=0.1,
initialW=None,
initial_bias=None):
"""Initialize Conv2dSubsampling."""
super(Conv2dSubsampling, self).__init__()
self.dropout = dropout
with self.init_scope():
# Standard deviation for Conv2D with 1 channel and kernel 3 x 3.
n = 1 * 3 * 3
stvd = 1.0 / np.sqrt(n)
self.conv1 = L.Convolution2D(
1,
channels,
3,
stride=2,
pad=1,
initialW=initialW(scale=stvd),
initial_bias=initial_bias(scale=stvd),
)
n = channels * 3 * 3
stvd = 1.0 / np.sqrt(n)
self.conv2 = L.Convolution2D(
channels,
channels,
3,
stride=2,
pad=1,
initialW=initialW(scale=stvd),
initial_bias=initial_bias(scale=stvd),
)
stvd = 1.0 / np.sqrt(dims)
self.out = L.Linear(
idim,
dims,
initialW=initialW(scale=stvd),
initial_bias=initial_bias(scale=stvd),
)
self.pe = PositionalEncoding(dims, dropout)
def __init__(self, odim, args, initialW=None, initial_bias=None):
super(Decoder, self).__init__()
initialW = chainer.initializers.Uniform if initialW is None else initialW
initial_bias = chainer.initializers.Uniform if initial_bias is None else initial_bias
with self.init_scope():
self.output_norm = LayerNorm(args.adim)
self.pe = PositionalEncoding(args.adim, args.dropout_rate)
stvd = 1. / np.sqrt(args.adim)
self.output_layer = L.Linear(args.adim, odim, initialW=initialW(scale=stvd),
initial_bias=initial_bias(scale=stvd))
self.embed = L.EmbedID(odim, args.adim, ignore_label=-1,
initialW=chainer.initializers.Normal(scale=1.0))
for i in range(args.dlayers):
name = 'decoders.' + str(i)
layer = DecoderLayer(args.adim, d_units=args.dunits,
h=args.aheads, dropout=args.dropout_rate,
initialW=initialW,
initial_bias=initial_bias)
self.add_link(name, layer)
self.n_layers = args.dlayers
def __init__(self,
channels,
idim,
dims,
dropout=0.1,
initialW=None,
initial_bias=None):
super(Conv2dSubsampling, self).__init__()
n = 1 * 3 * 3
stvd = 1. / np.sqrt(n)
layer = L.Convolution2D(1,
channels,
3,
stride=2,
pad=1,
initialW=initialW(scale=stvd),
initial_bias=initial_bias(scale=stvd))
self.add_link('conv.0', layer)
n = channels * 3 * 3
stvd = 1. / np.sqrt(n)
layer = L.Convolution2D(channels,
channels,
3,
stride=2,
pad=1,
initialW=initialW(scale=stvd),
initial_bias=initial_bias(scale=stvd))
self.add_link('conv.2', layer)
stvd = 1. / np.sqrt(dims)
layer = L.Linear(idim,
dims,
initialW=initialW(scale=stvd),
initial_bias=initial_bias(scale=stvd))
self.add_link('out.0', layer)
self.dropout = dropout
with self.init_scope():
self.pe = PositionalEncoding(dims, dropout)