def __init__(self, args):
super(TransformerLayer, self).__init__()
self.layernorm_positioning = args.layernorm_positioning
if hasattr(args, "attention_head_size"):
attention_head_size = args.attention_head_size
else:
attention_head_size = args.hidden_size // args.heads_num
has_bias = bool(1 - args.remove_transformer_bias)
# Multi-headed self-attention.
self.self_attn = MultiHeadedAttention(args.hidden_size,
args.heads_num,
attention_head_size,
args.dropout,
has_bias=has_bias)
self.dropout_1 = nn.Dropout(args.dropout)
self.layer_norm_1 = LayerNorm(args.hidden_size, has_bias=has_bias)
# Feed forward layer.
if args.feed_forward == "gated":
self.feed_forward = GatedFeedForward(args.hidden_size,
args.feedforward_size,
args.hidden_act, has_bias)
else:
self.feed_forward = PositionwiseFeedForward(
args.hidden_size, args.feedforward_size, args.hidden_act,
has_bias)
self.dropout_2 = nn.Dropout(args.dropout)
self.layer_norm_2 = LayerNorm(args.hidden_size, has_bias=has_bias)
def __init__(self, args):
super(GptBlock, self).__init__()
# Multi-headed self-attention.
self.self_attn = MultiHeadedAttention(args.hidden_size, args.heads_num,
args.dropout)
self.layer_norm_1 = LayerNorm(args.hidden_size)
# Feed forward layer.
self.feed_forward = PositionwiseFeedForward(args.hidden_size,
args.feedforward_size,
args.hidden_act)
self.layer_norm_2 = LayerNorm(args.hidden_size)
def __init__(self, args):
super(TransformerDecoderLayer, self).__init__()
self.self_attn = MultiHeadedAttention(args.hidden_size, args.heads_num,
args.dropout)
self.context_attn = MultiHeadedAttention(args.hidden_size,
args.heads_num, args.dropout)
self.layer_norm_1 = LayerNorm(args.hidden_size)
self.layer_norm_2 = LayerNorm(args.hidden_size)
self.layer_norm_3 = LayerNorm(args.hidden_size)
self.feed_forward = PositionwiseFeedForward(args.hidden_size,
args.feedforward_size,
args.hidden_act)
def __init__(self, args):
super(ISynthesizer, self).__init__()
self.att = None
self.dropout_1 = nn.Dropout(args.dropout)
self.layer_norm_1 = LayerNorm(args.hidden_size)
# Feed forward layer.
self.feed_forward = PositionwiseFeedForward(args.hidden_size,
args.feedforward_size)
self.dropout_2 = nn.Dropout(args.dropout)
self.layer_norm_2 = LayerNorm(args.hidden_size)
if self.__class__.__name__ == 'ISynthesizer':
raise Exception("ISynthesizer cannot be instantiated.")
def __init__(self, args):
super(RelationAwareTransformerLayer, self).__init__()
# Multi-headed self-attention.
self.self_attn = RelationAwareMultiHeadedAttention(
args.hidden_size, args.heads_num, args.dropout
)
self.dropout_1 = nn.Dropout(args.dropout)
self.layer_norm_1 = LayerNorm(args.hidden_size)
# Feed forward layer.
self.feed_forward = PositionwiseFeedForward(
args.hidden_size, args.feedforward_size
)
self.dropout_2 = nn.Dropout(args.dropout)
self.layer_norm_2 = LayerNorm(args.hidden_size)
def __init__(self, args):
super(TransformerDecoderLayer, self).__init__()
self.layernorm_positioning = args.layernorm_positioning
# Multi-headed self-attention.
self.self_attn = MultiHeadedAttention(args.hidden_size, args.heads_num,
args.dropout)
self.dropout_1 = nn.Dropout(args.dropout)
self.layer_norm_1 = LayerNorm(args.hidden_size)
# Multi-headed context-attention.
self.context_attn = MultiHeadedAttention(args.hidden_size,
args.heads_num, args.dropout)
self.dropout_2 = nn.Dropout(args.dropout)
self.layer_norm_2 = LayerNorm(args.hidden_size)
# Feed forward layer.
self.feed_forward = PositionwiseFeedForward(args.hidden_size,
args.feedforward_size,
args.hidden_act)
self.dropout_3 = nn.Dropout(args.dropout)
self.layer_norm_3 = LayerNorm(args.hidden_size)