def __init__(self,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
full_context_alignment=False,
alignment_heads=None):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=attention_dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
self.full_context_alignment = full_context_alignment
self.alignment_heads = alignment_heads
def __init__(self,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
tgt_concept_words_type=-1):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=attention_dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
self.tgt_concept_words_type = tgt_concept_words_type
if tgt_concept_words_type in [2]:
self.tgt_concept_mlp = nn.Linear(d_model * 2, d_model)
def __init__(self,
opt,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
dict_size=None,
label_emb=None):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=dropout,
max_relative_positions=max_relative_positions,
dict_size=dict_size,
label_emb=label_emb,
opt=opt)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=attention_dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
def __init__(self,
d_model,
heads,
d_ff,
dropout,
self_attn_type="scaled-dot",
max_relative_positions=0):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model, dropout=dropout)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
def __init__(self,
d_model,
heads,
d_ff,
dropout,
self_attn_type="scaled-dot"):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(heads,
d_model,
dropout=dropout)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model, dropout=dropout)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
mask = self._get_attn_subsequent_mask(MAX_SIZE)
# Register self.mask as a buffer in TransformerDecoderLayer, so
# it gets TransformerDecoderLayer's cuda behavior automatically.
self.register_buffer('mask', mask)
def __init__(self,
d_model,
heads,
d_ff,
dropout,
attn_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
ctx_weight_param=False):
super(TransformerGPTDecoderLayerCtxattn, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=attn_dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model, dropout=attn_dropout)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=dropout)
self.feed_forward = MLP(d_model, d_model * 4, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-5)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-5)
self.context_layer_norm = nn.LayerNorm(d_model, eps=1e-5)
self.drop = nn.Dropout(dropout)
if ctx_weight_param:
print('using ctx_weight_param')
self.ctx_weight = Parameter(torch.zeros(1))
self.ctx_weight_param = ctx_weight_param
def __init__(self,
d_model,
heads,
d_ff,
dropout,
attn_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0):
super(TransformerGPTUnconditionalDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=attn_dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model, dropout=attn_dropout)
self.feed_forward = MLP(d_model, d_model * 4, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-5)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-5)
self.drop = nn.Dropout(dropout)
class TransformerDecoderLayer(nn.Module):
"""
Args:
d_model (int): the dimension of keys/values/queries in
:class:`MultiHeadedAttention`, also the input size of
the first-layer of the :class:`PositionwiseFeedForward`.
heads (int): the number of heads for MultiHeadedAttention.
d_ff (int): the second-layer of the :class:`PositionwiseFeedForward`.
dropout (float): dropout probability.
self_attn_type (string): type of self-attention scaled-dot, average
"""
def __init__(self, d_model, heads, d_ff, dropout,
self_attn_type="scaled-dot", max_relative_positions=0):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads, d_model, dropout=dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model, dropout=dropout)
self.context_attn = MultiHeadedAttention(
heads, d_model, dropout=dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
def forward(self, inputs, memory_bank, src_pad_mask, tgt_pad_mask,
layer_cache=None, step=None):
"""
Args:
inputs (FloatTensor): ``(batch_size, 1, model_dim)``
memory_bank (FloatTensor): ``(batch_size, src_len, model_dim)``
src_pad_mask (LongTensor): ``(batch_size, 1, src_len)``
tgt_pad_mask (LongTensor): ``(batch_size, 1, 1)``
Returns:
(FloatTensor, FloatTensor):
* output ``(batch_size, 1, model_dim)``
* attn ``(batch_size, 1, src_len)``
"""
dec_mask = None
if step is None:
tgt_len = tgt_pad_mask.size(-1)
future_mask = torch.ones(
[tgt_len, tgt_len],
device=tgt_pad_mask.device,
dtype=torch.uint8)
future_mask = future_mask.triu_(1).view(1, tgt_len, tgt_len)
dec_mask = torch.gt(tgt_pad_mask + future_mask, 0)
input_norm = self.layer_norm_1(inputs)
if isinstance(self.self_attn, MultiHeadedAttention):
query, attn = self.self_attn(input_norm, input_norm, input_norm,
mask=dec_mask,
layer_cache=layer_cache,
attn_type="self")
elif isinstance(self.self_attn, AverageAttention):
query, attn = self.self_attn(input_norm, mask=dec_mask,
layer_cache=layer_cache, step=step)
query = self.drop(query) + inputs
query_norm = self.layer_norm_2(query)
context, attn = self.context_attn(memory_bank, memory_bank, query_norm,
mask=src_pad_mask,
layer_cache=layer_cache,
attn_type="context")
output = self.feed_forward(self.drop(context) + query)
return output, attn, context
def update_dropout(self, dropout):
self.self_attn.update_dropout(dropout)
self.context_attn.update_dropout(dropout)
self.feed_forward.update_dropout(dropout)
self.drop.p = dropout
class TransformerDecoderLayer(nn.Module):
"""
Args:
d_model (int): the dimension of keys/values/queries in
:class:`MultiHeadedAttention`, also the input size of
the first-layer of the :class:`PositionwiseFeedForward`.
heads (int): the number of heads for MultiHeadedAttention.
d_ff (int): the second-layer of the :class:`PositionwiseFeedForward`.
dropout (float): dropout probability.
self_attn_type (string): type of self-attention scaled-dot, average
"""
def __init__(self,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
full_context_alignment=False,
alignment_heads=None):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=attention_dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
self.full_context_alignment = full_context_alignment
self.alignment_heads = alignment_heads
def forward(self, *args, **kwargs):
""" Extend _forward for (possibly) multiple decoder pass:
1. Always a default (future masked) decoder forward pass,
2. Possibly a second future aware decoder pass for joint learn
full context alignement.
Args:
* All arguments of _forward.
with_align (bool): whether return alignment attention.
Returns:
(FloatTensor, FloatTensor, FloatTensor or None):
* output ``(batch_size, 1, model_dim)``
* top_attn ``(batch_size, 1, src_len)``
* attn_align ``(batch_size, 1, src_len)`` or None
"""
with_align = kwargs.pop('with_align', False)
output, attns = self._forward(*args, **kwargs)
top_attn = attns[:, 0, :, :].contiguous()
attn_align = None
if with_align:
if self.full_context_alignment:
# return _, (B, Q_len, K_len)
_, attns = self._forward(*args, **kwargs, future=True)
if self.alignment_heads is not None:
attns = attns[:, :self.alignment_heads, :, :].contiguous()
# layer average attention across heads, get ``(B, Q, K)``
# Case 1: no full_context, no align heads -> layer avg baseline
# Case 2: no full_context, 1 align heads -> guided align
# Case 3: full_context, 1 align heads -> full cte guided align
attn_align = attns.mean(dim=1)
return output, top_attn, attn_align
def _forward(self,
inputs,
memory_bank,
src_pad_mask,
tgt_pad_mask,
layer_cache=None,
step=None,
future=False):
""" A naive forward pass for transformer decoder.
# TODO: change 1 to T as T could be 1 or tgt_len
Args:
inputs (FloatTensor): ``(batch_size, 1, model_dim)``
memory_bank (FloatTensor): ``(batch_size, src_len, model_dim)``
src_pad_mask (LongTensor): ``(batch_size, 1, src_len)``
tgt_pad_mask (LongTensor): ``(batch_size, 1, 1)``
Returns:
(FloatTensor, FloatTensor):
* output ``(batch_size, 1, model_dim)``
* attns ``(batch_size, head, 1, src_len)``
"""
dec_mask = None
if step is None:
tgt_len = tgt_pad_mask.size(-1)
if not future: # apply future_mask, result mask in (B, T, T)
future_mask = torch.ones([tgt_len, tgt_len],
device=tgt_pad_mask.device,
dtype=torch.uint8)
future_mask = future_mask.triu_(1).view(1, tgt_len, tgt_len)
# BoolTensor was introduced in pytorch 1.2
try:
future_mask = future_mask.bool()
except AttributeError:
pass
dec_mask = torch.gt(tgt_pad_mask + future_mask, 0)
else: # only mask padding, result mask in (B, 1, T)
dec_mask = tgt_pad_mask
input_norm = self.layer_norm_1(inputs)
if isinstance(self.self_attn, MultiHeadedAttention):
query, _ = self.self_attn(input_norm,
input_norm,
input_norm,
mask=dec_mask,
layer_cache=layer_cache,
attn_type="self")
elif isinstance(self.self_attn, AverageAttention):
query, _ = self.self_attn(input_norm,
mask=dec_mask,
layer_cache=layer_cache,
step=step)
elif isinstance(self.self_attn, MultiHeadedCausalAttention):
query, _ = self.self_attn(input_norm,
input_norm,
input_norm,
mask=dec_mask,
layer_cache=layer_cache,
attn_type="self",
decoder=True)
query = self.drop(query) + inputs
query_norm = self.layer_norm_2(query)
mid, attns = self.context_attn(memory_bank,
memory_bank,
query_norm,
mask=src_pad_mask,
layer_cache=layer_cache,
attn_type="context")
output = self.feed_forward(self.drop(mid) + query)
return output, attns
def update_dropout(self, dropout, attention_dropout):
self.self_attn.update_dropout(attention_dropout)
self.context_attn.update_dropout(attention_dropout)
self.feed_forward.update_dropout(dropout)
self.drop.p = dropout
def __init__(
self,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
full_context_alignment=False,
alignment_heads=0,
pos_ffn_activation_fn=ActivationFunction.relu,
):
"""
Args:
d_model (int): the dimension of keys/values/queries in
:class:`MultiHeadedAttention`, also the input size of
the first-layer of the :class:`PositionwiseFeedForward`.
heads (int): the number of heads for MultiHeadedAttention.
d_ff (int): the second-layer of the
:class:`PositionwiseFeedForward`.
dropout (float): dropout in residual, self-attn(dot) and
feed-forward
attention_dropout (float): dropout in context_attn (and
self-attn(avg))
self_attn_type (string): type of self-attention scaled-dot,
average
max_relative_positions (int):
Max distance between inputs in relative positions
representations
aan_useffn (bool): Turn on the FFN layer in the AAN decoder
full_context_alignment (bool):
whether enable an extra full context decoder forward for
alignment
alignment_heads (int):
N. of cross attention heads to use for alignment guiding
pos_ffn_activation_fn (ActivationFunction):
activation function choice for PositionwiseFeedForward layer
"""
super(TransformerDecoderLayerBase, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=attention_dropout,
max_relative_positions=max_relative_positions,
)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout,
pos_ffn_activation_fn)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
self.full_context_alignment = full_context_alignment
self.alignment_heads = alignment_heads
class TransformerDecoderLayerBase(nn.Module):
def __init__(
self,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
full_context_alignment=False,
alignment_heads=0,
pos_ffn_activation_fn=ActivationFunction.relu,
):
"""
Args:
d_model (int): the dimension of keys/values/queries in
:class:`MultiHeadedAttention`, also the input size of
the first-layer of the :class:`PositionwiseFeedForward`.
heads (int): the number of heads for MultiHeadedAttention.
d_ff (int): the second-layer of the
:class:`PositionwiseFeedForward`.
dropout (float): dropout in residual, self-attn(dot) and
feed-forward
attention_dropout (float): dropout in context_attn (and
self-attn(avg))
self_attn_type (string): type of self-attention scaled-dot,
average
max_relative_positions (int):
Max distance between inputs in relative positions
representations
aan_useffn (bool): Turn on the FFN layer in the AAN decoder
full_context_alignment (bool):
whether enable an extra full context decoder forward for
alignment
alignment_heads (int):
N. of cross attention heads to use for alignment guiding
pos_ffn_activation_fn (ActivationFunction):
activation function choice for PositionwiseFeedForward layer
"""
super(TransformerDecoderLayerBase, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=attention_dropout,
max_relative_positions=max_relative_positions,
)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout,
pos_ffn_activation_fn)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
self.full_context_alignment = full_context_alignment
self.alignment_heads = alignment_heads
def forward(self, *args, **kwargs):
"""Extend `_forward` for (possibly) multiple decoder pass:
Always a default (future masked) decoder forward pass,
Possibly a second future aware decoder pass for joint learn
full context alignement, :cite:`garg2019jointly`.
Args:
* All arguments of _forward.
with_align (bool): whether return alignment attention.
Returns:
(FloatTensor, FloatTensor, FloatTensor or None):
* output ``(batch_size, T, model_dim)``
* top_attn ``(batch_size, T, src_len)``
* attn_align ``(batch_size, T, src_len)`` or None
"""
with_align = kwargs.pop("with_align", False)
output, attns = self._forward(*args, **kwargs)
top_attn = attns[:, 0, :, :].contiguous()
attn_align = None
if with_align:
if self.full_context_alignment:
# return _, (B, Q_len, K_len)
_, attns = self._forward(*args, **kwargs, future=True)
if self.alignment_heads > 0:
attns = attns[:, :self.alignment_heads, :, :].contiguous()
# layer average attention across heads, get ``(B, Q, K)``
# Case 1: no full_context, no align heads -> layer avg baseline
# Case 2: no full_context, 1 align heads -> guided align
# Case 3: full_context, 1 align heads -> full cte guided align
attn_align = attns.mean(dim=1)
return output, top_attn, attn_align
def update_dropout(self, dropout, attention_dropout):
self.self_attn.update_dropout(attention_dropout)
self.feed_forward.update_dropout(dropout)
self.drop.p = dropout
def _forward(self, *args, **kwargs):
raise NotImplementedError
def _compute_dec_mask(self, tgt_pad_mask, future):
tgt_len = tgt_pad_mask.size(-1)
if not future: # apply future_mask, result mask in (B, T, T)
future_mask = torch.ones(
[tgt_len, tgt_len],
device=tgt_pad_mask.device,
dtype=torch.uint8,
)
future_mask = future_mask.triu_(1).view(1, tgt_len, tgt_len)
# BoolTensor was introduced in pytorch 1.2
try:
future_mask = future_mask.bool()
except AttributeError:
pass
dec_mask = torch.gt(tgt_pad_mask + future_mask, 0)
else: # only mask padding, result mask in (B, 1, T)
dec_mask = tgt_pad_mask
return dec_mask
def _forward_self_attn(self, inputs_norm, dec_mask, layer_cache, step):
if isinstance(self.self_attn, MultiHeadedAttention):
return self.self_attn(
inputs_norm,
inputs_norm,
inputs_norm,
mask=dec_mask,
layer_cache=layer_cache,
attn_type="self",
)
elif isinstance(self.self_attn, AverageAttention):
return self.self_attn(inputs_norm,
mask=dec_mask,
layer_cache=layer_cache,
step=step)
else:
raise ValueError(
f"self attention {type(self.self_attn)} not supported")
class TransformerDecoderLayer(nn.Module):
"""Transformer Decoder layer block in Pre-Norm style.
Pre-Norm style is an improvement w.r.t. Original paper's Post-Norm style,
providing better converge speed and performance. This is also the actual
implementation in tensor2tensor and also avalable in fairseq.
See https://tunz.kr/post/4 and :cite:`DeeperTransformer`.
.. mermaid::
graph LR
%% "*SubLayer" can be self-attn, src-attn or feed forward block
A(input) --> B[Norm]
B --> C["*SubLayer"]
C --> D[Drop]
D --> E((+))
A --> E
E --> F(out)
Args:
d_model (int): the dimension of keys/values/queries in
:class:`MultiHeadedAttention`, also the input size of
the first-layer of the :class:`PositionwiseFeedForward`.
heads (int): the number of heads for MultiHeadedAttention.
d_ff (int): the second-layer of the :class:`PositionwiseFeedForward`.
dropout (float): dropout in residual, self-attn(dot) and feed-forward
attention_dropout (float): dropout in context_attn (and self-attn(avg))
self_attn_type (string): type of self-attention scaled-dot, average
max_relative_positions (int):
Max distance between inputs in relative positions representations
aan_useffn (bool): Turn on the FFN layer in the AAN decoder
full_context_alignment (bool):
whether enable an extra full context decoder forward for alignment
alignment_heads (int):
N. of cross attention heads to use for alignment guiding
"""
def __init__(self,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
full_context_alignment=False,
alignment_heads=0):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=attention_dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=attention_dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
self.full_context_alignment = full_context_alignment
self.alignment_heads = alignment_heads
def forward(self, *args, **kwargs):
""" Extend `_forward` for (possibly) multiple decoder pass:
Always a default (future masked) decoder forward pass,
Possibly a second future aware decoder pass for joint learn
full context alignement, :cite:`garg2019jointly`.
Args:
* All arguments of _forward.
with_align (bool): whether return alignment attention.
Returns:
(FloatTensor, FloatTensor, FloatTensor or None):
* output ``(batch_size, T, model_dim)``
* top_attn ``(batch_size, T, src_len)``
* attn_align ``(batch_size, T, src_len)`` or None
"""
with_align = kwargs.pop('with_align', False)
output, attns = self._forward(*args, **kwargs)
top_attn = attns[:, 0, :, :].contiguous()
attn_align = None
if with_align:
if self.full_context_alignment:
# return _, (B, Q_len, K_len)
_, attns = self._forward(*args, **kwargs, future=True)
if self.alignment_heads > 0:
attns = attns[:, :self.alignment_heads, :, :].contiguous()
# layer average attention across heads, get ``(B, Q, K)``
# Case 1: no full_context, no align heads -> layer avg baseline
# Case 2: no full_context, 1 align heads -> guided align
# Case 3: full_context, 1 align heads -> full cte guided align
attn_align = attns.mean(dim=1)
return output, top_attn, attn_align
def _forward(self,
inputs,
memory_bank,
src_pad_mask,
tgt_pad_mask,
layer_cache=None,
step=None,
future=False):
""" A naive forward pass for transformer decoder.
# T: could be 1 in the case of stepwise decoding or tgt_len
Args:
inputs (FloatTensor): ``(batch_size, T, model_dim)``
memory_bank (FloatTensor): ``(batch_size, src_len, model_dim)``
src_pad_mask (LongTensor): ``(batch_size, 1, src_len)``
tgt_pad_mask (LongTensor): ``(batch_size, 1, T)``
layer_cache (dict or None): cached layer info when stepwise decode
step (int or None): stepwise decoding counter
future (bool): If set True, do not apply future_mask.
Returns:
(FloatTensor, FloatTensor):
* output ``(batch_size, T, model_dim)``
* attns ``(batch_size, head, T, src_len)``
"""
dec_mask = None
if step is None:
tgt_len = tgt_pad_mask.size(-1)
if not future: # apply future_mask, result mask in (B, T, T)
future_mask = torch.ones([tgt_len, tgt_len],
device=tgt_pad_mask.device,
dtype=torch.uint8)
future_mask = future_mask.triu_(1).view(1, tgt_len, tgt_len)
# BoolTensor was introduced in pytorch 1.2
try:
future_mask = future_mask.bool()
except AttributeError:
pass
dec_mask = torch.gt(tgt_pad_mask + future_mask, 0)
else: # only mask padding, result mask in (B, 1, T)
dec_mask = tgt_pad_mask
input_norm = self.layer_norm_1(inputs)
if isinstance(self.self_attn, MultiHeadedAttention):
query, _ = self.self_attn(input_norm,
input_norm,
input_norm,
mask=dec_mask,
layer_cache=layer_cache,
attn_type="self")
elif isinstance(self.self_attn, AverageAttention):
query, _ = self.self_attn(input_norm,
mask=dec_mask,
layer_cache=layer_cache,
step=step)
query = self.drop(query) + inputs
query_norm = self.layer_norm_2(query)
mid, attns = self.context_attn(memory_bank,
memory_bank,
query_norm,
mask=src_pad_mask,
layer_cache=layer_cache,
attn_type="context")
output = self.feed_forward(self.drop(mid) + query)
return output, attns
def update_dropout(self, dropout, attention_dropout):
self.self_attn.update_dropout(attention_dropout)
self.context_attn.update_dropout(attention_dropout)
self.feed_forward.update_dropout(dropout)
self.drop.p = dropout
class TransformerDecoderLayer(nn.Module):
"""
Args:
d_model (int): the dimension of keys/values/queries in
:class:`MultiHeadedAttention`, also the input size of
the first-layer of the :class:`PositionwiseFeedForward`.
heads (int): the number of heads for MultiHeadedAttention.
d_ff (int): the second-layer of the :class:`PositionwiseFeedForward`.
dropout (float): dropout probability.
self_attn_type (string): type of self-attention scaled-dot, average
"""
def __init__(self,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
tgt_concept_words_type=-1):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=dropout,
max_relative_positions=max_relative_positions)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=attention_dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
self.tgt_concept_words_type = tgt_concept_words_type
if tgt_concept_words_type in [2]:
self.tgt_concept_mlp = nn.Linear(d_model * 2, d_model)
def forward(self,
inputs,
memory_bank,
src_pad_mask,
tgt_pad_mask,
layer_cache=None,
step=None,
tgt_concept_words_emb=None,
tgt_concept_words_type=-1):
"""
Args:
inputs (FloatTensor): ``(batch_size, 1, model_dim)``
memory_bank (FloatTensor): ``(batch_size, src_len, model_dim)``
src_pad_mask (LongTensor): ``(batch_size, 1, src_len)``
tgt_pad_mask (LongTensor): ``(batch_size, 1, 1)``
Returns:
(FloatTensor, FloatTensor):
* output ``(batch_size, 1, model_dim)``
* attn ``(batch_size, 1, src_len)``
"""
dec_mask = None
if step is None:
tgt_len = tgt_pad_mask.size(-1)
future_mask = torch.ones([tgt_len, tgt_len],
device=tgt_pad_mask.device,
dtype=torch.uint8)
future_mask = future_mask.triu_(1).view(1, tgt_len, tgt_len)
# BoolTensor was introduced in pytorch 1.2
try:
future_mask = future_mask.bool()
except AttributeError:
pass
dec_mask = torch.gt(tgt_pad_mask + future_mask, 0)
input_norm = self.layer_norm_1(inputs)
if isinstance(self.self_attn, MultiHeadedAttention):
query, attn = self.self_attn(input_norm,
input_norm,
input_norm,
mask=dec_mask,
layer_cache=layer_cache,
attn_type="self")
elif isinstance(self.self_attn, AverageAttention):
query, attn = self.self_attn(input_norm,
mask=dec_mask,
layer_cache=layer_cache,
step=step)
query = self.drop(query) + inputs
# ablation
if tgt_concept_words_emb is not None:
# print(query.shape, tgt_concept_words_emb.shape)
if self.tgt_concept_words_type == 2:
query = self.tgt_concept_mlp(
torch.cat([query, tgt_concept_words_emb], dim=2))
if self.tgt_concept_words_type == 3:
query = (query + tgt_concept_words_emb) / 2
query_norm = self.layer_norm_2(query)
mid, attn = self.context_attn(memory_bank,
memory_bank,
query_norm,
mask=src_pad_mask,
layer_cache=layer_cache,
attn_type="context")
output = self.feed_forward(self.drop(mid) + query)
return output, attn
def update_dropout(self, dropout, attention_dropout):
self.self_attn.update_dropout(attention_dropout)
self.context_attn.update_dropout(attention_dropout)
self.feed_forward.update_dropout(dropout)
self.drop.p = dropout
class TransformerDecoderLayer(nn.Module):
"""
Args:
d_model (int): the dimension of keys/values/queries in
:class:`MultiHeadedAttention`, also the input size of
the first-layer of the :class:`PositionwiseFeedForward`.
heads (int): the number of heads for MultiHeadedAttention.
d_ff (int): the second-layer of the :class:`PositionwiseFeedForward`.
dropout (float): dropout probability.
self_attn_type (string): type of self-attention scaled-dot, average
"""
def __init__(self,
opt,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
dict_size=None,
label_emb=None):
super(TransformerDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=dropout,
max_relative_positions=max_relative_positions,
dict_size=dict_size,
label_emb=label_emb,
opt=opt)
elif self_attn_type == "average":
self.self_attn = AverageAttention(d_model,
dropout=attention_dropout,
aan_useffn=aan_useffn)
self.context_attn = MultiHeadedAttention(heads,
d_model,
dropout=attention_dropout)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
def forward(self,
inputs,
memory_bank,
src_pad_mask,
tgt_pad_mask,
layer_cache=None,
step=None,
gold_par_attn=None,
gold_ch_attn=None):
"""
Args:
inputs (FloatTensor): ``(batch_size, 1, model_dim)``
memory_bank (FloatTensor): ``(batch_size, src_len, model_dim)``
src_pad_mask (LongTensor): ``(batch_size, 1, src_len)``
tgt_pad_mask (LongTensor): ``(batch_size, 1, 1)``
Returns:
(FloatTensor, FloatTensor):
* output ``(batch_size, 1, model_dim)``
* attn ``(batch_size, 1, src_len)``
"""
dec_mask = None
if step is None:
tgt_len = tgt_pad_mask.size(-1)
future_mask = torch.ones([tgt_len, tgt_len],
device=tgt_pad_mask.device,
dtype=torch.uint8)
future_mask = future_mask.triu_(1).view(1, tgt_len, tgt_len)
#future_mask = future_mask.triu_(0).view(1, tgt_len, tgt_len)
#future_mask[0,0,0]=0
# BoolTensor was introduced in pytorch 1.2
try:
future_mask = future_mask.bool()
except AttributeError:
pass
dec_mask = torch.gt(tgt_pad_mask + future_mask, 0)
# elif step!=0 and synsa:
# self_mask = torch.zeros(
# [1,1, step+1],
# device=tgt_pad_mask.device,
# dtype=torch.uint8)
# self_mask[:,:,-1]=1
# try:
# self_mask = self_mask.bool()
# except AttributeError:
# pass
# dec_mask = torch.gt(self_mask, 0)
input_norm = self.layer_norm_1(inputs)
if isinstance(self.self_attn, MultiHeadedAttention):
query, tgt_attn, second_attn, ch_labels, par_labels = self.self_attn(
input_norm,
input_norm,
input_norm,
mask=dec_mask,
layer_cache=layer_cache,
attn_type="self",
gold_par_attn=gold_par_attn,
gold_ch_attn=gold_ch_attn)
elif isinstance(self.self_attn, AverageAttention):
query, attn = self.self_attn(input_norm,
mask=dec_mask,
layer_cache=layer_cache,
step=step)
query = self.drop(query) + inputs
query_norm = self.layer_norm_2(query)
mid, src_attn, _, _, _ = self.context_attn(memory_bank,
memory_bank,
query_norm,
mask=src_pad_mask,
layer_cache=layer_cache,
attn_type="context")
output = self.feed_forward(self.drop(mid) + query)
return output, src_attn, tgt_attn, second_attn, dec_mask, ch_labels, par_labels
def update_dropout(self, dropout, attention_dropout):
self.self_attn.update_dropout(attention_dropout)
self.context_attn.update_dropout(attention_dropout)
self.feed_forward.update_dropout(dropout)
self.drop.p = dropout
class TransformerLMDecoderLayer(TransformerDecoderLayerBase):
"""Transformer Decoder only layer block in GPT style.
.. mermaid::
graph LR
%% "*SubLayer" can be self-attn, src-attn or feed forward block
A(input) --> B[Norm]
B --> C["*SubLayer"]
C --> D[Drop]
D --> E((+))
A --> E
E --> F(out)
Args:
d_model (int): the dimension of keys/values/queries in
:class:`MultiHeadedAttention`, also the input size of
the first-layer of the :class:`PositionwiseFeedForward`.
heads (int): the number of heads for MultiHeadedAttention.
d_ff (int): the second-layer of the :class:`PositionwiseFeedForward`.
dropout (float): dropout in residual, self-attn(dot) and feed-forward
attention_dropout (float): dropout in context_attn (and self-attn(avg))
self_attn_type (string): type of self-attention scaled-dot, average
max_relative_positions (int):
Max distance between inputs in relative positions representations
aan_useffn (bool): Turn on the FFN layer in the AAN decoder
full_context_alignment (bool):
whether enable an extra full context decoder forward for alignment
alignment_heads (int):
N. of cross attention heads to use for alignment guiding
"""
def __init__(
self,
d_model,
heads,
d_ff,
dropout,
attention_dropout,
self_attn_type="scaled-dot",
max_relative_positions=0,
aan_useffn=False,
full_context_alignment=False,
alignment_heads=0,
):
super(TransformerLMDecoderLayer, self).__init__()
if self_attn_type == "scaled-dot":
self.self_attn = MultiHeadedAttention(
heads,
d_model,
dropout=attention_dropout,
max_relative_positions=max_relative_positions,
)
elif self_attn_type == "average":
self.self_attn = AverageAttention(
d_model, dropout=attention_dropout, aan_useffn=aan_useffn
)
self.feed_forward = PositionwiseFeedForward(d_model, d_ff, dropout)
self.layer_norm_1 = nn.LayerNorm(d_model, eps=1e-6)
self.layer_norm_2 = nn.LayerNorm(d_model, eps=1e-6)
self.drop = nn.Dropout(dropout)
self.full_context_alignment = full_context_alignment
self.alignment_heads = alignment_heads
def _forward(
self, inputs, tgt_pad_mask, layer_cache=None, step=None, future=False
):
"""A naive forward pass for transformer decoder.
# T: could be 1 in the case of stepwise decoding or tgt_len
Args:
inputs (FloatTensor): ``(batch_size, T, model_dim)``
tgt_pad_mask (bool): ``(batch_size, 1, T)``
layer_cache (dict or None): cached layer info when stepwise decode
step (int or None): stepwise decoding counter
future (bool): If set True, do not apply future_mask.
Returns:
(FloatTensor, FloatTensor):
* output ``(batch_size, T, model_dim)``
* attns ``(batch_size, head, T, T)``
"""
dec_mask = None
if step is None:
tgt_len = tgt_pad_mask.size(-1)
if not future: # apply future_mask, result mask in (B, T, T)
future_mask = torch.ones(
[tgt_len, tgt_len],
device=tgt_pad_mask.device,
dtype=torch.uint8,
)
future_mask = future_mask.triu_(1).view(1, tgt_len, tgt_len)
# BoolTensor was introduced in pytorch 1.2
try:
future_mask = future_mask.bool()
except AttributeError:
pass
dec_mask = torch.gt(tgt_pad_mask + future_mask, 0)
else: # only mask padding, result mask in (B, 1, T)
dec_mask = tgt_pad_mask
inputs_norm = self.layer_norm_1(inputs)
if isinstance(self.self_attn, MultiHeadedAttention):
query, attns = self.self_attn(
inputs_norm,
inputs_norm,
inputs_norm,
mask=dec_mask,
layer_cache=layer_cache,
attn_type="self",
)
elif isinstance(self.self_attn, AverageAttention):
query, attns = self.self_attn(
inputs_norm, mask=dec_mask, layer_cache=layer_cache, step=step
)
output = self.drop(query) + inputs
output_feedforward = self.feed_forward(self.layer_norm_2(output))
output_norm = self.drop(output_feedforward) + output
return output_norm, attns
def update_dropout(self, dropout, attention_dropout):
self.self_attn.update_dropout(attention_dropout)
self.feed_forward.update_dropout(dropout)
self.drop.p = dropout
