def __init__(self, num_layers, hidden_size, attn_type, copy_attn,
cnn_kernel_width, dropout, embeddings, copy_attn_type):
super(CNNDecoder, self).__init__()
self.cnn_kernel_width = cnn_kernel_width
self.embeddings = embeddings
# Decoder State
self.state = {}
input_size = self.embeddings.embedding_size
self.linear = nn.Linear(input_size, hidden_size)
self.conv_layers = nn.ModuleList([
GatedConv(hidden_size, cnn_kernel_width, dropout, True)
for i in range(num_layers)
])
self.attn_layers = nn.ModuleList(
[ConvMultiStepAttention(hidden_size) for i in range(num_layers)])
# CNNDecoder has its own attention mechanism.
# Set up a separate copy attention layer if needed.
assert not copy_attn, "Copy mechanism not yet tested in conv2conv"
if copy_attn:
self.copy_attn = GlobalAttention(hidden_size,
attn_type=copy_attn_type)
else:
self.copy_attn = None
def __init__(self, num_layers, hidden_size, attn_type, copy_attn,
cnn_kernel_width, dropout, embeddings):
super(CNNDecoder, self).__init__()
# Basic attributes.
self.decoder_type = 'cnn'
self.num_layers = num_layers
self.hidden_size = hidden_size
self.cnn_kernel_width = cnn_kernel_width
self.embeddings = embeddings
self.dropout = dropout
# Build the CNN.
input_size = self.embeddings.embedding_size
self.linear = nn.Linear(input_size, self.hidden_size)
self.conv_layers = nn.ModuleList()
for _ in range(self.num_layers):
self.conv_layers.append(
GatedConv(self.hidden_size, self.cnn_kernel_width,
self.dropout, True))
self.attn_layers = nn.ModuleList()
for _ in range(self.num_layers):
self.attn_layers.append(
onmt.modules.ConvMultiStepAttention(self.hidden_size))
# CNNDecoder has its own attention mechanism.
# Set up a separated copy attention layer, if needed.
self._copy = False
if copy_attn:
self.copy_attn = onmt.modules.GlobalAttention(hidden_size,
attn_type=attn_type)
self._copy = True
