def __init__(self, embedding, positional_embedding, layer, num_layers,
dropout):
super(TransformerEncoder, self).__init__()
self.embedding = embedding
self.positional_embedding = positional_embedding
# embed_size = embedding.embedding_dim
hidden_size = layer.hidden_size
# self.input_projection = nn.Linear(embed_size, hidden_size)
self.layers = clone(layer, num_layers)
self.layer_norm = nn.LayerNorm(hidden_size)
# self.output_projection = nn.Linear(hidden_size, embed_size)
self.dropout = dropout
def __init__(self, embedding, positional_embedding, layer, num_layers, dropout, share_decoder_embedding=True):
super(ConvDecoder, self).__init__()
self.embedding = embedding
self.positional_embedding = positional_embedding
embed_size = embedding.embedding_dim
vocab_size = embedding.num_embeddings
hidden_size = layer.hidden_size
self.input_projection = nn.Linear(embed_size, hidden_size)
self.layers = clone(layer, num_layers)
self.layer_norm = nn.LayerNorm(hidden_size)
self.output_projection = nn.Linear(hidden_size, embed_size)
self.dropout = dropout
self.generator = nn.Linear(embed_size, vocab_size)
if share_decoder_embedding:
self.generator.weight = embedding.weight
def __init__(self,
input_size,
hidden_size,
num_layers=1,
bias=True,
batch_first=False,
dropout=0,
bidirectional=True):
super(ResidualGRU, self).__init__()
self.lstm_layers = clone(
nn.LSTM(input_size=hidden_size,
hidden_size=hidden_size,
num_layers=1,
bias=bias,
batch_first=batch_first,
bidirectional=bidirectional), num_layers)
self.input_projection = nn.Linear(input_size, hidden_size)
self.input_size = input_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.bias = bias
self.batch_first = batch_first
self.dropout = dropout
self.bidirectional = bidirectional