def __init__(self, args):
super(TransformerDecoderLayer, 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)
with_scale = bool(1 - args.remove_attention_scale)
# Multi-headed self-attention.
self.self_attn = MultiHeadedAttention(args.hidden_size,
args.heads_num,
attention_head_size,
args.dropout,
has_bias=has_bias,
with_scale=with_scale)
self.dropout_1 = nn.Dropout(args.dropout)
# Multi-headed context-attention.
self.context_attn = MultiHeadedAttention(args.hidden_size,
args.heads_num,
attention_head_size,
args.dropout,
has_bias=has_bias,
with_scale=with_scale)
self.dropout_2 = nn.Dropout(args.dropout)
# 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_3 = nn.Dropout(args.dropout)
# Layer Normalization
if args.layernorm == "t5":
self.layer_norm_1 = T5LayerNorm(args.hidden_size)
self.layer_norm_2 = T5LayerNorm(args.hidden_size)
self.layer_norm_3 = T5LayerNorm(args.hidden_size)
else:
self.layer_norm_1 = LayerNorm(args.hidden_size)
self.layer_norm_2 = LayerNorm(args.hidden_size)
self.layer_norm_3 = LayerNorm(args.hidden_size)
def __init__(self, args):
super(TransformerDecoder, self).__init__()
self.layers_num = args.layers_num
self.layernorm_positioning = args.layernorm_positioning
self.relative_position_embedding = args.relative_position_embedding
self.transformer_decoder = nn.ModuleList(
[TransformerDecoderLayer(args) for _ in range(self.layers_num)]
)
if "deepspeed_checkpoint_activations" in args:
self.deepspeed_checkpoint_activations = args.deepspeed_checkpoint_activations
self.deepspeed_checkpoint_layers_num = args.deepspeed_checkpoint_layers_num
else:
self.deepspeed_checkpoint_activations = False
has_bias = bool(1 - args.remove_transformer_bias)
if self.layernorm_positioning == "pre":
if args.layernorm == "t5":
self.layer_norm = T5LayerNorm(args.hidden_size)
else:
self.layer_norm = LayerNorm(args.hidden_size)
if self.relative_position_embedding:
self.self_pos_emb = RelativePositionEmbedding(bidirectional=False, heads_num=args.heads_num,
num_buckets=args.relative_attention_buckets_num)
def __init__(self, args):
super(TransformerEncoder, self).__init__()
self.mask = args.mask
self.layers_num = args.layers_num
self.parameter_sharing = args.parameter_sharing
self.factorized_embedding_parameterization = args.factorized_embedding_parameterization
self.layernorm_positioning = args.layernorm_positioning
self.relative_position_embedding = args.relative_position_embedding
self.has_residual_attention = args.has_residual_attention
if "deepspeed_checkpoint_activations" in args:
self.deepspeed_checkpoint_activations = args.deepspeed_checkpoint_activations
self.deepspeed_checkpoint_layers_num = args.deepspeed_checkpoint_layers_num
else:
self.deepspeed_checkpoint_activations = False
has_bias = bool(1 - args.remove_transformer_bias)
if self.factorized_embedding_parameterization:
self.linear = nn.Linear(args.emb_size, args.hidden_size)
if self.parameter_sharing:
self.transformer = TransformerLayer(args)
else:
self.transformer = nn.ModuleList(
[TransformerLayer(args) for _ in range(self.layers_num)]
)
if self.layernorm_positioning == "pre":
if args.layernorm == "t5":
self.layer_norm = T5LayerNorm(args.hidden_size)
else:
self.layer_norm = LayerNorm(args.hidden_size)
if self.relative_position_embedding:
self.relative_pos_emb = RelativePositionEmbedding(bidirectional=True, heads_num=args.heads_num,
num_buckets=args.relative_attention_buckets_num)