def __init__(self, args, dictionary):
super().__init__(dictionary)
self.dropout = args.dropout
self.embed_dim = args.decoder_embed_dim
self.output_embed_dim = args.decoder_embed_dim
self.padding_idx = dictionary.pad_idx
self.max_tgt_positions = args.max_tgt_positions
self.embedding = generate_embedding(len(dictionary), self.embed_dim,
dictionary.pad_idx)
self.embed_scale = 1.0 if args.no_scale_embedding else math.sqrt(
self.embed_dim)
self.embed_positions = PositionalEmbedding(
self.embed_dim,
padding_idx=self.padding_idx,
init_size=self.max_tgt_positions + self.padding_idx + 1)
self.layers = nn.ModuleList([])
# Generate N identical Decoder Layers
self.layers.extend([
TransformerDecoderLayer(args) for _ in range(args.decoder_layers)
])
self.embed_out = nn.Linear(self.output_embed_dim, len(dictionary))
nn.init.normal_(self.embed_out.weight,
mean=0,
std=self.output_embed_dim**-0.5)
def __init__(self, args, dictionary):
super().__init__(dictionary)
self.dropout = args.dropout
self.embed_dim = args.encoder_embed_dim
self.padding_idx = dictionary.pad_idx
self.max_src_positions = args.max_src_positions
self.embedding = generate_embedding(len(dictionary), self.embed_dim,
dictionary.pad_idx)
self.embed_scale = 1.0 if args.no_scale_embedding else math.sqrt(
self.embed_dim)
self.embed_positions = PositionalEmbedding(
self.embed_dim,
padding_idx=self.padding_idx,
init_size=self.max_src_positions + self.padding_idx + 1)
self.layers = nn.ModuleList([])
# Generate N identical Encoder Layers
self.layers.extend([
TransformerEncoderLayer(args) for _ in range(args.encoder_layers)
])