def __init__(self, args, no_encoder_attn=False, kernel_size=0):
super().__init__()
self.embed_dim = args.decoder_embed_dim
self.conv_dim = args.decoder_conv_dim
if args.decoder_glu:
self.linear1 = Linear(self.embed_dim, 2 * self.conv_dim)
self.act = nn.GLU()
else:
self.linear1 = Linear(self.embed_dim, self.conv_dim)
self.act = None
if args.decoder_conv_type == "lightweight":
self.conv = LightweightConv(
self.conv_dim,
kernel_size,
padding_l=kernel_size - 1,
weight_softmax=args.weight_softmax,
num_heads=args.decoder_attention_heads,
weight_dropout=args.weight_dropout,
)
elif args.decoder_conv_type == "dynamic":
self.conv = DynamicConv(
self.conv_dim,
kernel_size,
padding_l=kernel_size - 1,
weight_softmax=args.weight_softmax,
num_heads=args.decoder_attention_heads,
weight_dropout=args.weight_dropout,
)
else:
raise NotImplementedError
self.linear2 = Linear(self.conv_dim, self.embed_dim)
self.dropout_module = FairseqDropout(
args.dropout, module_name=self.__class__.__name__)
self.relu_dropout_module = FairseqDropout(
args.relu_dropout, module_name=self.__class__.__name__)
self.input_dropout_module = FairseqDropout(
args.input_dropout, module_name=self.__class__.__name__)
self.normalize_before = args.decoder_normalize_before
self.conv_layer_norm = LayerNorm(self.embed_dim)
if no_encoder_attn:
self.encoder_attn = None
self.encoder_attn_layer_norm = None
else:
self.encoder_attn = MultiheadAttention(
self.embed_dim,
args.decoder_attention_heads,
dropout=args.attention_dropout,
encoder_decoder_attention=True,
)
self.encoder_attn_layer_norm = LayerNorm(self.embed_dim)
self.fc1 = Linear(self.embed_dim, args.decoder_ffn_embed_dim)
self.fc2 = Linear(args.decoder_ffn_embed_dim, self.embed_dim)
self.final_layer_norm = LayerNorm(self.embed_dim)
self.need_attn = True
def __init__(self,
args,
no_encoder_attn=False,
add_bias_kv=False,
add_zero_attn=False):
super().__init__()
self.embed_dim = args.decoder_embed_dim
self.self_attn = MultiheadAttention(
embed_dim=self.embed_dim,
num_heads=args.decoder_attention_heads,
dropout=args.attention_dropout,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
self_attention=True,
)
self.dropout = args.dropout
self.activation_fn = utils.get_activation_fn(
activation=getattr(args, "activation_fn", "relu"))
self.activation_dropout = getattr(args, "activation_dropout", 0)
if self.activation_dropout == 0:
# for backwards compatibility with models that use args.relu_dropout
self.activation_dropout = getattr(args, "relu_dropout", 0)
self.normalize_before = args.decoder_normalize_before
# use layerNorm rather than FusedLayerNorm for exporting.
# char_inputs can be used to determint this.
# TODO remove this once we update apex with the fix
export = getattr(args, "char_inputs", False)
self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export)
if no_encoder_attn:
self.encoder_attn = None
self.encoder_attn_layer_norm = None
else:
self.encoder_attn = MultiheadAttention(
self.embed_dim,
args.decoder_attention_heads,
kdim=getattr(args, "encoder_embed_dim", None),
vdim=getattr(args, "encoder_embed_dim", None),
dropout=args.attention_dropout,
encoder_decoder_attention=True,
)
self.encoder_attn_layer_norm = LayerNorm(self.embed_dim,
export=export)
self.fc1 = Linear(self.embed_dim, args.decoder_ffn_embed_dim)
self.fc2 = Linear(args.decoder_ffn_embed_dim, self.embed_dim)
self.final_layer_norm = LayerNorm(self.embed_dim, export=export)
self.need_attn = True
self.onnx_trace = False
def build_self_attention(self, embed_dim, args):
return MultiheadAttention(
embed_dim,
args.encoder_attention_heads,
dropout=args.attention_dropout,
self_attention=True,
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
)
def build_encoder_attention(self, embed_dim, args):
return MultiheadAttention(
embed_dim,
args.decoder_attention_heads,
kdim=getattr(args, "encoder_embed_dim", None),
vdim=getattr(args, "encoder_embed_dim", None),
dropout=args.attention_dropout,
encoder_decoder_attention=True,
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
)
def build_self_attention(self,
embed_dim,
args,
add_bias_kv=False,
add_zero_attn=False):
return MultiheadAttention(
embed_dim,
args.decoder_attention_heads,
dropout=args.attention_dropout,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
self_attention=not getattr(args, "cross_self_attention", False),
q_noise=self.quant_noise,
qn_block_size=self.quant_noise_block_size,
)
def build_self_attention(
self,
embed_dim,
num_attention_heads,
dropout,
self_attention,
q_noise,
qn_block_size,
):
return MultiheadAttention(
embed_dim,
num_attention_heads,
dropout=dropout,
self_attention=True,
q_noise=q_noise,
qn_block_size=qn_block_size,
)
def __init__(self, args):
super().__init__()
self.embed_dim = args.encoder_embed_dim
self.self_attn = MultiheadAttention(
self.embed_dim,
args.encoder_attention_heads,
dropout=args.attention_dropout,
self_attention=True,
)
self.self_attn_layer_norm = LayerNorm(self.embed_dim)
self.dropout = args.dropout
self.activation_fn = utils.get_activation_fn(
activation=getattr(args, "activation_fn", "relu"))
self.activation_dropout = getattr(args, "activation_dropout", 0)
if self.activation_dropout == 0:
# for backwards compatibility with models that use args.relu_dropout
self.activation_dropout = getattr(args, "relu_dropout", 0)
self.normalize_before = args.encoder_normalize_before
self.fc1 = Linear(self.embed_dim, args.encoder_ffn_embed_dim)
self.fc2 = Linear(args.encoder_ffn_embed_dim, self.embed_dim)
self.final_layer_norm = LayerNorm(self.embed_dim)
def __init__(
self,
embedding_dim: float = 768,
ffn_embedding_dim: float = 3072,
num_attention_heads: float = 8,
dropout: float = 0.1,
attention_dropout: float = 0.1,
activation_dropout: float = 0.1,
activation_fn: str = "relu",
layer_norm_first: bool = False,
) -> None:
super().__init__()
# Initialize parameters
self.embedding_dim = embedding_dim
self.dropout = dropout
self.activation_dropout = activation_dropout
# Initialize blocks
self.activation_fn = utils.get_activation_fn(activation_fn)
self.self_attn = MultiheadAttention(
self.embedding_dim,
num_attention_heads,
dropout=attention_dropout,
self_attention=True,
)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(self.activation_dropout)
self.dropout3 = nn.Dropout(dropout)
self.layer_norm_first = layer_norm_first
# layer norm associated with the self attention layer
self.self_attn_layer_norm = LayerNorm(self.embedding_dim)
self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)
self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)
# layer norm associated with the position wise feed-forward NN
self.final_layer_norm = LayerNorm(self.embedding_dim)
class TransformerDecoderLayer(nn.Module):
"""Decoder layer block.
In the original paper each operation (multi-head attention, encoder
attention or FFN) is postprocessed with: `dropout -> add residual ->
layernorm`. In the tensor2tensor code they suggest that learning is more
robust when preprocessing each layer with layernorm and postprocessing with:
`dropout -> add residual`. We default to the approach in the paper, but the
tensor2tensor approach can be enabled by setting
*args.decoder_normalize_before* to ``True``.
Args:
args (argparse.Namespace): parsed command-line arguments
no_encoder_attn (bool, optional): whether to attend to encoder outputs
(default: False).
"""
def __init__(self,
args,
no_encoder_attn=False,
add_bias_kv=False,
add_zero_attn=False):
super().__init__()
self.embed_dim = args.decoder_embed_dim
self.self_attn = MultiheadAttention(
embed_dim=self.embed_dim,
num_heads=args.decoder_attention_heads,
dropout=args.attention_dropout,
add_bias_kv=add_bias_kv,
add_zero_attn=add_zero_attn,
self_attention=True,
)
self.dropout = args.dropout
self.activation_fn = utils.get_activation_fn(
activation=getattr(args, "activation_fn", "relu"))
self.activation_dropout = getattr(args, "activation_dropout", 0)
if self.activation_dropout == 0:
# for backwards compatibility with models that use args.relu_dropout
self.activation_dropout = getattr(args, "relu_dropout", 0)
self.normalize_before = args.decoder_normalize_before
# use layerNorm rather than FusedLayerNorm for exporting.
# char_inputs can be used to determint this.
# TODO remove this once we update apex with the fix
export = getattr(args, "char_inputs", False)
self.self_attn_layer_norm = LayerNorm(self.embed_dim, export=export)
if no_encoder_attn:
self.encoder_attn = None
self.encoder_attn_layer_norm = None
else:
self.encoder_attn = MultiheadAttention(
self.embed_dim,
args.decoder_attention_heads,
kdim=getattr(args, "encoder_embed_dim", None),
vdim=getattr(args, "encoder_embed_dim", None),
dropout=args.attention_dropout,
encoder_decoder_attention=True,
)
self.encoder_attn_layer_norm = LayerNorm(self.embed_dim,
export=export)
self.fc1 = Linear(self.embed_dim, args.decoder_ffn_embed_dim)
self.fc2 = Linear(args.decoder_ffn_embed_dim, self.embed_dim)
self.final_layer_norm = LayerNorm(self.embed_dim, export=export)
self.need_attn = True
self.onnx_trace = False
def prepare_for_onnx_export_(self):
self.onnx_trace = True
def forward(self, input):
"""
Args:
input (Tuple):
input[0] (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)`
input[1] (Tensor): encoder output of shape `(batch, src_len, embed_dim)`
input[2] (ByteTensor/FloatTensor): encoder padding mask -
binary ByteTensor of shape `(batch, src_len)` where padding elements
are indicated by ``1``.
Returns:
output (Tuple):
output[0] (Tensor): encoded output of shape `(batch, src_len, embed_dim)`
output[1] (ByteTensor/FloatTensor): encoder padding mask
output[2] (LongTensor): previous decoder outputs
"""
# Note: incremental state is not yet supported
mt_task = False
if isinstance(input, tuple):
x = input[0]
encoder_out = input[1]
encoder_padding_mask = input[2]
incremental_state = None
mt_task = True
else:
x = input
encoder_out = None
encoder_padding_mask = None
incremental_state = None
if incremental_state is None:
self_attn_mask = self.buffered_future_mask(x)
else:
self_attn_mask = None
# TODO: add back prev_self_attn_state, prev_attn_state,
# self_attn_padding_mask
prev_self_attn_state = None
prev_attn_state = None
self_attn_padding_mask = None
residual = x
x = self.maybe_layer_norm(self.self_attn_layer_norm, x, before=True)
if prev_self_attn_state is not None:
if incremental_state is None:
incremental_state = {}
prev_key, prev_value = prev_self_attn_state
saved_state = {"prev_key": prev_key, "prev_value": prev_value}
self.self_attn._set_input_buffer(incremental_state, saved_state)
x, attn = self.self_attn(
query=x,
key=x,
value=x,
key_padding_mask=self_attn_padding_mask,
incremental_state=incremental_state,
need_weights=False,
attn_mask=self_attn_mask,
)
x = F.dropout(x, p=self.dropout, training=self.training)
x = residual + x
x = self.maybe_layer_norm(self.self_attn_layer_norm, x, after=True)
if self.encoder_attn is not None:
residual = x
x = self.maybe_layer_norm(self.encoder_attn_layer_norm,
x,
before=True)
if prev_attn_state is not None:
if incremental_state is None:
incremental_state = {}
prev_key, prev_value = prev_attn_state
saved_state = {"prev_key": prev_key, "prev_value": prev_value}
self.encoder_attn._set_input_buffer(incremental_state,
saved_state)
x, attn = self.encoder_attn(
query=x,
key=encoder_out,
value=encoder_out,
key_padding_mask=encoder_padding_mask,
incremental_state=incremental_state,
static_kv=True,
need_weights=(not self.training and self.need_attn),
)
x = F.dropout(x, p=self.dropout, training=self.training)
x = residual + x
x = self.maybe_layer_norm(self.encoder_attn_layer_norm,
x,
after=True)
residual = x
x = self.maybe_layer_norm(self.final_layer_norm, x, before=True)
x = self.activation_fn(self.fc1(x))
x = F.dropout(x, p=self.activation_dropout, training=self.training)
x = self.fc2(x)
x = F.dropout(x, p=self.dropout, training=self.training)
x = residual + x
x = self.maybe_layer_norm(self.final_layer_norm, x, after=True)
if mt_task:
return (x, encoder_out, encoder_padding_mask)
return x
def buffered_future_mask(self, tensor):
dim = tensor.size(0)
if (not hasattr(self, "_future_mask") or self._future_mask is None
or self._future_mask.device != tensor.device):
self._future_mask = torch.triu(
utils.fill_with_neg_inf(tensor.new(dim, dim)), 1)
if self._future_mask.size(0) < dim:
self._future_mask = torch.triu(
utils.fill_with_neg_inf(self._future_mask.resize_(dim, dim)),
1)
return self._future_mask[:dim, :dim]
def maybe_layer_norm(self, layer_norm, x, before=False, after=False):
assert before ^ after
if after ^ self.normalize_before:
return layer_norm(x)
else:
return x
def make_generation_fast_(self, need_attn=False, **kwargs):
self.need_attn = need_attn
class LightConvDecoderLayer(nn.Module):
"""Decoder layer block.
Args:
args (argparse.Namespace): parsed command-line arguments
no_encoder_attn (bool, optional): whether to attend to encoder outputs.
Default: ``False``
kernel_size: kernel size of the convolution
"""
def __init__(self, args, no_encoder_attn=False, kernel_size=0):
super().__init__()
self.embed_dim = args.decoder_embed_dim
self.conv_dim = args.decoder_conv_dim
if args.decoder_glu:
self.linear1 = Linear(self.embed_dim, 2 * self.conv_dim)
self.act = nn.GLU()
else:
self.linear1 = Linear(self.embed_dim, self.conv_dim)
self.act = None
if args.decoder_conv_type == "lightweight":
self.conv = LightweightConv(
self.conv_dim,
kernel_size,
padding_l=kernel_size - 1,
weight_softmax=args.weight_softmax,
num_heads=args.decoder_attention_heads,
weight_dropout=args.weight_dropout,
)
elif args.decoder_conv_type == "dynamic":
self.conv = DynamicConv(
self.conv_dim,
kernel_size,
padding_l=kernel_size - 1,
weight_softmax=args.weight_softmax,
num_heads=args.decoder_attention_heads,
weight_dropout=args.weight_dropout,
)
else:
raise NotImplementedError
self.linear2 = Linear(self.conv_dim, self.embed_dim)
self.dropout_module = FairseqDropout(
args.dropout, module_name=self.__class__.__name__)
self.relu_dropout_module = FairseqDropout(
args.relu_dropout, module_name=self.__class__.__name__)
self.input_dropout_module = FairseqDropout(
args.input_dropout, module_name=self.__class__.__name__)
self.normalize_before = args.decoder_normalize_before
self.conv_layer_norm = LayerNorm(self.embed_dim)
if no_encoder_attn:
self.encoder_attn = None
self.encoder_attn_layer_norm = None
else:
self.encoder_attn = MultiheadAttention(
self.embed_dim,
args.decoder_attention_heads,
dropout=args.attention_dropout,
encoder_decoder_attention=True,
)
self.encoder_attn_layer_norm = LayerNorm(self.embed_dim)
self.fc1 = Linear(self.embed_dim, args.decoder_ffn_embed_dim)
self.fc2 = Linear(args.decoder_ffn_embed_dim, self.embed_dim)
self.final_layer_norm = LayerNorm(self.embed_dim)
self.need_attn = True
def forward(
self,
x,
encoder_out,
encoder_padding_mask,
incremental_state,
prev_conv_state=None,
prev_attn_state=None,
conv_mask=None,
conv_padding_mask=None,
):
"""
Args:
x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)`
encoder_padding_mask (ByteTensor): binary ByteTensor of shape
`(batch, src_len)` where padding elements are indicated by ``1``.
Returns:
encoded output of shape `(batch, src_len, embed_dim)`
"""
residual = x
x = self.maybe_layer_norm(self.conv_layer_norm, x, before=True)
if prev_conv_state is not None:
if incremental_state is None:
incremental_state = {}
self.conv._set_input_buffer(incremental_state, prev_conv_state)
x = self.input_dropout_module(x)
x = self.linear1(x)
if self.act is not None:
x = self.act(x)
x = self.conv(x, incremental_state=incremental_state)
x = self.linear2(x)
x = self.dropout_module(x)
x = residual + x
x = self.maybe_layer_norm(self.conv_layer_norm, x, after=True)
attn = None
if self.encoder_attn is not None:
residual = x
x = self.maybe_layer_norm(self.encoder_attn_layer_norm,
x,
before=True)
if prev_attn_state is not None:
if incremental_state is None:
incremental_state = {}
prev_key, prev_value = prev_attn_state
saved_state = {"prev_key": prev_key, "prev_value": prev_value}
self.encoder_attn._set_input_buffer(incremental_state,
saved_state)
x, attn = self.encoder_attn(
query=x,
key=encoder_out,
value=encoder_out,
key_padding_mask=encoder_padding_mask,
incremental_state=incremental_state,
static_kv=True,
need_weights=(not self.training and self.need_attn),
)
x = self.dropout_module(x)
x = residual + x
x = self.maybe_layer_norm(self.encoder_attn_layer_norm,
x,
after=True)
residual = x
x = self.maybe_layer_norm(self.final_layer_norm, x, before=True)
x = F.relu(self.fc1(x))
x = self.relu_dropout_module(x)
x = self.fc2(x)
x = self.dropout_module(x)
x = residual + x
x = self.maybe_layer_norm(self.final_layer_norm, x, after=True)
return x, attn
def maybe_layer_norm(self, layer_norm, x, before=False, after=False):
assert before ^ after
if after ^ self.normalize_before:
return layer_norm(x)
else:
return x
def make_generation_fast_(self, need_attn=False, **kwargs):
self.need_attn = need_attn
def extra_repr(self):
return (
"dropout={}, relu_dropout={}, input_dropout={}, normalize_before={}"
.format(
self.dropout_module.p,
self.relu_dropout_module.p,
self.input_dropout_module.p,
self.normalize_before,
))