import math
from collections.abc import Callable
from typing import Any
import torch
from einops import rearrange
from torch import Tensor, nn
from torch.nn import functional as F
from torch.nn.common_types import _size_1_t, _size_2_t, _size_3_t
from .functional.packed import packed_linear, packed_multi_head_attention_forward
def check_packed_parameters_consistency(alpha: float, gamma: int, num_estimators: int) -> None:
"""Check the consistency of the parameters of the Packed-Ensembles layers.
Args:
alpha (int): The width multiplier of the layer.
gamma (int): The number of groups in the ensemble.
num_estimators (int): The number of estimators in the ensemble.
"""
if alpha is None:
raise ValueError("You must specify the value of the arg. `alpha`")
if alpha <= 0:
raise ValueError(f"Attribute `alpha` should be > 0, not {alpha}")
if not isinstance(gamma, int):
raise TypeError(f"Attribute `gamma` should be an int, not {type(gamma)}")
if gamma <= 0:
raise ValueError(f"Attribute `gamma` should be >= 1, not {gamma}")
if num_estimators is None:
raise ValueError("You must specify the value of the arg. `num_estimators`")
if not isinstance(num_estimators, int):
raise TypeError(
"Attribute `num_estimators` should be an int, not " f"{type(num_estimators)}"
)
if num_estimators <= 0:
raise ValueError("Attribute `num_estimators` should be >= 1, not " f"{num_estimators}")
[docs]class PackedLinear(nn.Module):
def __init__(
self,
in_features: int,
out_features: int,
alpha: float,
num_estimators: int,
gamma: int = 1,
bias: bool = True,
first: bool = False,
last: bool = False,
implementation: str = "legacy",
rearrange: bool = True,
device=None,
dtype=None,
) -> None:
r"""Packed-Ensembles-style Linear layer.
This layer computes fully-connected operation for a given number of
estimators (:attr:`num_estimators`).
Args:
in_features (int): Number of input features of the linear layer.
out_features (int): Number of channels produced by the linear layer.
alpha (float): The width multiplier of the linear layer.
num_estimators (int): The number of estimators grouped in the layer.
gamma (int, optional): Defaults to ``1``.
bias (bool, optional): It ``True``, adds a learnable bias to the
output. Defaults to ``True``.
first (bool, optional): Whether this is the first layer of the
network. Defaults to ``False``.
last (bool, optional): Whether this is the last layer of the network.
Defaults to ``False``.
implementation (str, optional): The implementation to use. Defaults
to ``"legacy"``.
rearrange (bool, optional): Rearrange the input and outputs for
compatibility with previous and later layers. Defaults to ``True``.
device (torch.device, optional): The device to use for the layer's
parameters. Defaults to ``None``.
dtype (torch.dtype, optional): The dtype to use for the layer's
parameters. Defaults to ``None``.
Explanation Note:
Increasing :attr:`alpha` will increase the number of channels of the
ensemble, increasing its representation capacity. Increasing
:attr:`gamma` will increase the number of groups in the network and
therefore reduce the number of parameters.
Note:
Each ensemble member will only see
:math:`\frac{\text{in_features}}{\text{num_estimators}}` features,
so when using :attr:`gamma` you should make sure that
:attr:`in_features` and :attr:`out_features` are both divisible by
:attr:`n_estimators` :math:`\times`:attr:`gamma`. However, the
number of input and output features will be changed to comply with
this constraint.
Note:
The input should be of shape (`batch_size`, :attr:`in_features`, 1,
1). The (often) necessary rearrange operation is executed by
default.
"""
check_packed_parameters_consistency(alpha, gamma, num_estimators)
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.first = first
self.num_estimators = num_estimators
self.rearrange = rearrange
self.implementation = implementation
# Define the number of features of the underlying convolution
extended_in_features = int(in_features * (1 if first else alpha))
extended_out_features = int(out_features * (num_estimators if last else alpha))
# Define the number of groups of the underlying convolution
actual_groups = num_estimators * gamma if not first else 1
# fix if not divisible by groups
if extended_in_features % actual_groups:
extended_in_features += num_estimators - extended_in_features % (actual_groups)
if extended_out_features % actual_groups:
extended_out_features += num_estimators - extended_out_features % (actual_groups)
# FIXME: This is a temporary check
assert implementation in [
"legacy",
"sparse",
"full",
"einsum",
], f"Unknown implementation: {implementation} for PackedLinear"
if self.implementation == "legacy":
self.weight = nn.Parameter(
torch.empty(
(
extended_out_features,
extended_in_features // actual_groups,
1,
),
**factory_kwargs,
)
)
else:
self.weight = nn.Parameter(
torch.empty(
(
actual_groups,
extended_out_features // actual_groups,
extended_in_features // actual_groups,
),
**factory_kwargs,
)
)
self.in_features = extended_in_features // actual_groups
self.out_features = extended_out_features // actual_groups
self.groups = actual_groups
if bias:
self.bias = nn.Parameter(torch.empty(extended_out_features, **factory_kwargs))
else:
self.register_parameter("bias", None)
self.reset_parameters()
def reset_parameters(self) -> None:
if self.implementation == "legacy":
nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5))
else:
for n in range(self.groups):
nn.init.kaiming_uniform_(self.weight[n], a=math.sqrt(5))
if self.bias is not None:
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight[0])
bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
nn.init.uniform_(self.bias, -bound, bound)
if self.implementation == "sparse":
self.weight = nn.Parameter(torch.block_diag(*self.weight).to_sparse())
def _rearrange_forward(self, x: Tensor) -> Tensor:
x = x.unsqueeze(-1)
if not self.first:
x = rearrange(x, "(m e) c h -> e (m c) h", m=self.num_estimators)
x = F.conv1d(x, self.weight, self.bias, 1, 0, 1, self.groups)
x = rearrange(x, "e (m c) h -> (m e) c h", m=self.num_estimators)
return x.squeeze(-1)
def forward(self, inputs: Tensor) -> Tensor:
if self.implementation == "legacy":
if self.rearrange:
return self._rearrange_forward(inputs)
return F.conv1d(inputs, self.weight, self.bias, 1, 0, 1, self.groups)
return packed_linear(
inputs=inputs,
weight=self.weight,
num_groups=self.groups,
implementation=self.implementation,
bias=self.bias,
)
class PackedConv1d(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: _size_1_t,
alpha: int,
num_estimators: int,
gamma: int = 1,
stride: _size_1_t = 1,
padding: str | _size_1_t = 0,
dilation: _size_1_t = 1,
groups: int = 1,
minimum_channels_per_group: int = 64,
bias: bool = True,
padding_mode: str = "zeros",
first: bool = False,
last: bool = False,
device=None,
dtype=None,
) -> None:
r"""Packed-Ensembles-style Conv1d layer.
Args:
in_channels (int): Number of channels in the input image.
out_channels (int): Number of channels produced by the convolution.
kernel_size (int or tuple): Size of the convolving kernel.
alpha (int): The channel multiplier of the convolutional layer.
num_estimators (int): Number of estimators in the ensemble.
gamma (int, optional): Defaults to ``1``.
stride (int or tuple, optional): Stride of the convolution.
Defaults to ``1``.
padding (int, tuple or str, optional): Padding added to both sides of
the input. Defaults to ``0``.
dilation (int or tuple, optional): Spacing between kernel elements.
Defaults to ``1``.
groups (int, optional): Number of blocked connexions from input
channels to output channels for each estimator. Defaults to ``1``.
minimum_channels_per_group (int, optional): Smallest possible number of
channels per group.
bias (bool, optional): If ``True``, adds a learnable bias to the
output. Defaults to ``True``.
padding_mode (str, optional): ``'zeros'``, ``'reflect'``,
``'replicate'`` or ``'circular'``. Defaults to ``'zeros'``.
first (bool, optional): Whether this is the first layer of the
network. Defaults to ``False``.
last (bool, optional): Whether this is the last layer of the network.
Defaults to ``False``.
device (torch.device, optional): The device to use for the layer's
parameters. Defaults to ``None``.
dtype (torch.dtype, optional): The dtype to use for the layer's
parameters. Defaults to ``None``.
Explanation Note:
Increasing :attr:`alpha` will increase the number of channels of the
ensemble, increasing its representation capacity. Increasing
:attr:`gamma` will increase the number of groups in the network and
therefore reduce the number of parameters.
Note:
Each ensemble member will only see
:math:`\frac{\text{in_channels}}{\text{num_estimators}}` channels,
so when using :attr:`groups` you should make sure that
:attr:`in_channels` and :attr:`out_channels` are both divisible by
:attr:`num_estimators` :math:`\times`:attr:`gamma` :math:`\times`
:attr:`groups`. However, the number of input and output channels will
be changed to comply with this constraint.
"""
check_packed_parameters_consistency(alpha, gamma, num_estimators)
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.num_estimators = num_estimators
# Define the number of channels of the underlying convolution
extended_in_channels = int(in_channels * (1 if first else alpha))
extended_out_channels = int(out_channels * (num_estimators if last else alpha))
# Define the number of groups of the underlying convolution
actual_groups = 1 if first else gamma * groups * num_estimators
while (
extended_in_channels % actual_groups != 0
or extended_in_channels // actual_groups < minimum_channels_per_group
) and actual_groups // (groups * num_estimators) > 1:
gamma -= 1
actual_groups = gamma * groups * num_estimators
# fix if not divisible by groups
if extended_in_channels % actual_groups:
extended_in_channels += num_estimators - extended_in_channels % actual_groups
if extended_out_channels % actual_groups:
extended_out_channels += num_estimators - extended_out_channels % actual_groups
self.conv = nn.Conv1d(
in_channels=extended_in_channels,
out_channels=extended_out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
groups=actual_groups,
bias=bias,
padding_mode=padding_mode,
**factory_kwargs,
)
def forward(self, inputs: Tensor) -> Tensor:
return self.conv(inputs)
@property
def weight(self) -> Tensor:
r"""The weight of the underlying convolutional layer."""
return self.conv.weight
@property
def bias(self) -> Tensor | None:
r"""The bias of the underlying convolutional layer."""
return self.conv.bias
[docs]class PackedConv2d(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: _size_2_t,
alpha: int,
num_estimators: int,
gamma: int = 1,
stride: _size_2_t = 1,
padding: str | _size_2_t = 0,
dilation: _size_2_t = 1,
groups: int = 1,
minimum_channels_per_group: int = 64,
bias: bool = True,
padding_mode: str = "zeros",
first: bool = False,
last: bool = False,
device: Any | None = None,
dtype: Any | None = None,
) -> None:
r"""Packed-Ensembles-style Conv2d layer.
Args:
in_channels (int): Number of channels in the input image.
out_channels (int): Number of channels produced by the convolution.
kernel_size (int or tuple): Size of the convolving kernel.
alpha (int): The channel multiplier of the convolutional layer.
num_estimators (int): Number of estimators in the ensemble.
gamma (int, optional): Defaults to ``1``.
stride (int or tuple, optional): Stride of the convolution.
Defaults to ``1``.
padding (int, tuple or str, optional): Padding added to all four sides
of the input. Defaults to ``0``.
dilation (int or tuple, optional): Spacing between kernel elements.
Defaults to ``1``.
groups (int, optional): Number of blocked connexions from input
channels to output channels for each estimator. Defaults to ``1``.
minimum_channels_per_group (int, optional): Smallest possible number of
channels per group.
bias (bool, optional): If ``True``, adds a learnable bias to the
output. Defaults to ``True``.
padding_mode (str, optional): ``'zeros'``, ``'reflect'``,
``'replicate'`` or ``'circular'``. Defaults to ``'zeros'``.
first (bool, optional): Whether this is the first layer of the
network. Defaults to ``False``.
last (bool, optional): Whether this is the last layer of the network.
Defaults to ``False``.
device (torch.device, optional): The device to use for the layer's
parameters. Defaults to ``None``.
dtype (torch.dtype, optional): The dtype to use for the layer's
parameters. Defaults to ``None``.
Explanation Note:
Increasing :attr:`alpha` will increase the number of channels of the
ensemble, increasing its representation capacity. Increasing
:attr:`gamma` will increase the number of groups in the network and
therefore reduce the number of parameters.
Note:
Each ensemble member will only see
:math:`\frac{\text{in_channels}}{\text{num_estimators}}` channels,
so when using :attr:`groups` you should make sure that
:attr:`in_channels` and :attr:`out_channels` are both divisible by
:attr:`num_estimators` :math:`\times`:attr:`gamma` :math:`\times`
:attr:`groups`. However, the number of input and output channels will
be changed to comply with this constraint.
"""
check_packed_parameters_consistency(alpha, gamma, num_estimators)
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.num_estimators = num_estimators
# Define the number of channels of the underlying convolution
extended_in_channels = int(in_channels * (1 if first else alpha))
extended_out_channels = int(out_channels * (num_estimators if last else alpha))
# Define the number of groups of the underlying convolution
actual_groups = 1 if first else gamma * groups * num_estimators
while (
extended_in_channels % actual_groups != 0
or extended_in_channels // actual_groups < minimum_channels_per_group
) and actual_groups // (groups * num_estimators) > 1:
gamma -= 1
actual_groups = gamma * groups * num_estimators
# fix if not divisible by groups
if extended_in_channels % actual_groups:
extended_in_channels += num_estimators - extended_in_channels % actual_groups
if extended_out_channels % actual_groups:
extended_out_channels += num_estimators - extended_out_channels % actual_groups
self.conv = nn.Conv2d(
in_channels=extended_in_channels,
out_channels=extended_out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
groups=actual_groups,
bias=bias,
padding_mode=padding_mode,
**factory_kwargs,
)
def forward(self, inputs: Tensor) -> Tensor:
return self.conv(inputs)
@property
def weight(self) -> Tensor:
r"""The weight of the underlying convolutional layer."""
return self.conv.weight
@property
def bias(self) -> Tensor | None:
r"""The bias of the underlying convolutional layer."""
return self.conv.bias
class PackedConv3d(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: _size_3_t,
alpha: int,
num_estimators: int,
gamma: int = 1,
stride: _size_3_t = 1,
padding: str | _size_3_t = 0,
dilation: _size_3_t = 1,
groups: int = 1,
minimum_channels_per_group: int = 64,
bias: bool = True,
padding_mode: str = "zeros",
first: bool = False,
last: bool = False,
device: Any | None = None,
dtype: Any | None = None,
) -> None:
r"""Packed-Ensembles-style Conv3d layer.
Args:
in_channels (int): Number of channels in the input image.
out_channels (int): Number of channels produced by the convolution.
kernel_size (int or tuple): Size of the convolving kernel.
alpha (int): The channel multiplier of the convolutional layer.
num_estimators (int): Number of estimators in the ensemble.
gamma (int, optional): Defaults to ``1``.
stride (int or tuple, optional): Stride of the convolution.
Defaults to ``1``.
padding (int, tuple or str, optional): Padding added to all six sides
of the input. Defaults to ``0``.
dilation (int or tuple, optional): Spacing between kernel elements.
Defaults to ``1``.
groups (int, optional): Number of blocked connexions from input
channels to output channels for each estimator. Defaults to ``1``.
minimum_channels_per_group (int, optional): Smallest possible number of
channels per group.
bias (bool, optional): If ``True``, adds a learnable bias to the
output. Defaults to ``True``.
padding_mode (str, optional): ``'zeros'``, ``'reflect'``,
``'replicate'`` or ``'circular'``. Defaults to ``'zeros'``.
first (bool, optional): Whether this is the first layer of the
network. Defaults to ``False``.
last (bool, optional): Whether this is the last layer of the network.
Defaults to ``False``.
device (torch.device, optional): The device to use for the layer's
parameters. Defaults to ``None``.
dtype (torch.dtype, optional): The dtype to use for the layer's
parameters. Defaults to ``None``.
Explanation Note:
Increasing :attr:`alpha` will increase the number of channels of the
ensemble, increasing its representation capacity. Increasing
:attr:`gamma` will increase the number of groups in the network and
therefore reduce the number of parameters.
Note:
Each ensemble member will only see
:math:`\frac{\text{in_channels}}{\text{num_estimators}}` channels,
so when using :attr:`groups` you should make sure that
:attr:`in_channels` and :attr:`out_channels` are both divisible by
:attr:`num_estimators` :math:`\times`:attr:`gamma` :math:`\times`
:attr:`groups`. However, the number of input and output channels will
be changed to comply with this constraint.
"""
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
check_packed_parameters_consistency(alpha, gamma, num_estimators)
self.num_estimators = num_estimators
# Define the number of channels of the underlying convolution
extended_in_channels = int(in_channels * (1 if first else alpha))
extended_out_channels = int(out_channels * (num_estimators if last else alpha))
# Define the number of groups of the underlying convolution
actual_groups = 1 if first else gamma * groups * num_estimators
while (
extended_in_channels % actual_groups != 0
or extended_in_channels // actual_groups < minimum_channels_per_group
) and actual_groups // (groups * num_estimators) > 1:
gamma -= 1
actual_groups = gamma * groups * num_estimators
# fix if not divisible by groups
if extended_in_channels % actual_groups:
extended_in_channels += num_estimators - extended_in_channels % actual_groups
if extended_out_channels % actual_groups:
extended_out_channels += num_estimators - extended_out_channels % actual_groups
self.conv = nn.Conv3d(
in_channels=extended_in_channels,
out_channels=extended_out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
groups=actual_groups,
bias=bias,
padding_mode=padding_mode,
**factory_kwargs,
)
def forward(self, inputs: Tensor) -> Tensor:
return self.conv(inputs)
@property
def weight(self) -> Tensor:
r"""The weight of the underlying convolutional layer."""
return self.conv.weight
@property
def bias(self) -> Tensor | None:
r"""The bias of the underlying convolutional layer."""
return self.conv.bias
[docs]class PackedLayerNorm(nn.GroupNorm):
def __init__(
self,
embed_dim: int,
num_estimators: int,
alpha: float,
eps: float = 1e-5,
affine: bool = True,
device=None,
dtype=None,
) -> None:
r"""Packed-Ensembles-style LayerNorm layer.
Args:
embed_dim (int): the number of features in the input tensor.
num_estimators (int): the number of estimators in the ensemble.
alpha (float): the width multiplier of the layer.
eps (float, optional): a value added to the denominator for numerical stability. Defaults
to 1e-5.
affine (bool, optional): a boolean value that when set to ``True``, this module has
learnable per_channel affine parameters initialized to ones (for weights) and zeros
(for biases). Defaults to ``True``.
device (torch.device, optional): the device to use for the layer's parameters. Defaults
to ``None``.
dtype (torch.dtype, optional): the dtype to use for the layer's parameters. Defaults to
``None``.
Shape:
- Input: :math:`(B, *)` where :math:`*` means any number of additional dimensions.
- Output: :math:`(B, *)` (same shape as input)
"""
super().__init__(
num_groups=num_estimators,
num_channels=int(embed_dim * alpha),
eps=eps,
affine=affine,
device=device,
dtype=dtype,
)
def forward(self, inputs: Tensor) -> Tensor:
x = rearrange(inputs, "b ... h -> b h ...")
x = F.group_norm(
x,
self.num_groups,
self.weight,
self.bias,
self.eps,
)
return rearrange(x, "b h ... -> b ... h")
[docs]class PackedMultiheadAttention(nn.Module):
__constants__ = ["batch_first"]
bias_k: Tensor | None
bias_v: Tensor | None
def __init__(
self,
embed_dim: int,
num_heads: int,
alpha: float,
num_estimators: int,
gamma: int = 1,
dropout=0.0,
bias=True,
add_bias_kv=False,
add_zero_attn=False,
kdim: int | None = None,
vdim: int | None = None,
batch_first=False,
first=False,
last=False,
device=None,
dtype=None,
) -> None:
r"""Packed-Ensembles-style MultiheadAttention layer.
Args:
embed_dim (int): Size of the embedding dimension.
num_heads (int): Number of parallel attention heads.
alpha (float): The width multiplier of the embedding dimension.
num_estimators (int): The number of estimators packed in the layer.
gamma (int, optional): Defaults to ``1``.
dropout (float, optional): Dropout probability on ``attn_output_weights``. Defaults to ``0.0``
(no dropout).
bias (bool, optional): Ì specified, adds bias to input / output projection layers.
Defaults to ``True``.
add_bias_kv (bool, optional): If specified, adds bias to the key and value sequences at
``dim=0``. Defaults to ``False``.
add_zero_attn (bool, optional): If specified, adds a new batch of zeros to the key and
value sequences at ``dim=1``. Defaults to ``False``.
kdim (int | None, optional): Total number of features for keys. Defaults to ``None``
(uses ``kdim=embed_dim``).
vdim (int | None, optional): Total number of features for values. Defaults to ``None``
(uses ``vdim=embed_dim``).
batch_first (bool, optional): If ``True``, then the input and output tensors are provided
as (batch, seq, feature). Defaults to ``False`` (seq, batch, feature).
first (bool, optional): Whether this is the first layer of the network. Defaults to
``False``.
last (bool, optional): Whether this is the last layer of the network. Defaults to
``False``.
device (torch.device, optional): The device to use for the layer's parameters. Defaults
to ``None``.
dtype (torch.dtype, optional): The dtype to use for the layer's parameters. Defaults to
``None``.
Reference:
- `Attention Is All You Need <https://arxiv.org/abs/1706.03762>`_: Original Multihead Attention formulation.
- `Hierarchical Light Tranformer Ensembles for Multimodal Trajectory Forecasting <https://arxiv.org/abs/2403.17678>`_
: Packed-Ensembles-style Multihead Attention formulation.
"""
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.kdim = kdim if kdim is not None else embed_dim
self.vdim = vdim if vdim is not None else embed_dim
self._qkv_same_embed_dim = self.kdim == embed_dim and self.vdim == embed_dim
self.embed_dim = int(embed_dim * alpha)
augmentation = 1 if first else alpha
in_embed_dim = int(embed_dim * augmentation)
self.kdim = int(self.kdim * augmentation)
self.vdim = int(self.vdim * augmentation)
self.num_groups = 1 if first else num_estimators * gamma
self.num_heads = num_heads * self.num_groups
self.dropout = dropout
self.batch_first = batch_first
self.head_dim = self.embed_dim // self.num_heads
assert (
self.head_dim * self.num_heads == self.embed_dim
), "embed_dim must be divisible by num_heads"
self.num_estimators = num_estimators
self.alpha = alpha
self.gamma = gamma
if not self._qkv_same_embed_dim:
self.q_proj_weight = nn.Parameter(
torch.empty(
(
self.num_groups,
self.embed_dim // self.num_groups,
in_embed_dim // self.num_groups,
),
**factory_kwargs,
)
)
self.k_proj_weight = nn.Parameter(
torch.empty(
(
self.num_groups,
self.embed_dim // self.num_groups,
self.kdim // self.num_groups,
),
**factory_kwargs,
)
)
self.v_proj_weight = nn.Parameter(
torch.empty(
(
self.num_groups,
self.embed_dim // self.num_groups,
self.vdim // self.num_groups,
),
**factory_kwargs,
)
)
self.register_parameter("in_proj_weight", None)
else:
self.in_proj_weight = nn.Parameter(
torch.empty(
(
self.num_groups,
3 * self.embed_dim // self.num_groups,
in_embed_dim // self.num_groups,
),
**factory_kwargs,
)
)
self.register_parameter("q_proj_weight", None)
self.register_parameter("k_proj_weight", None)
self.register_parameter("v_proj_weight", None)
if bias:
self.in_proj_bias = nn.Parameter(torch.empty(3 * self.embed_dim, **factory_kwargs))
else:
self.register_parameter("in_proj_bias", None)
if add_bias_kv:
self.bias_k = nn.Parameter(torch.empty((1, 1, self.embed_dim), **factory_kwargs))
self.bias_v = nn.Parameter(torch.empty((1, 1, self.embed_dim), **factory_kwargs))
else:
self.bias_k = self.bias_v = None
self.out_proj = PackedLinear(
in_features=embed_dim,
out_features=embed_dim,
alpha=alpha,
num_estimators=num_estimators,
gamma=gamma,
implementation="einsum",
bias=bias,
first=False,
last=last,
**factory_kwargs,
)
self.add_zero_attn = add_zero_attn
self._reset_parameters()
def _reset_parameters(self):
if self._qkv_same_embed_dim:
for i in range(self.in_proj_weight.size(0)):
nn.init.xavier_uniform_(self.in_proj_weight[i])
else:
for i in range(self.q_proj_weight.size(0)):
nn.init.xavier_uniform_(self.q_proj_weight[i])
nn.init.xavier_uniform_(self.k_proj_weight[i])
nn.init.xavier_uniform_(self.v_proj_weight[i])
if self.in_proj_bias is not None:
nn.init.constant_(self.in_proj_bias, 0.0)
nn.init.constant_(self.out_proj.bias, 0.0)
[docs] def forward(
self,
query: Tensor,
key: Tensor,
value: Tensor,
key_padding_mask: Tensor | None = None,
need_weights: bool = False,
attn_mask: Tensor | None = None,
average_attn_weights: bool = True,
is_causal: bool = False,
) -> tuple[Tensor, None]:
r"""Computes attention outputs given query, key, and value tensors.
Args:
query (Tensor): Query embeddings of shape :math:`(L, E_q)` for unbatched input,
:math:`(L, B, E_q)` when ``batch_first=False`` or :math:`(B, L, E_q)` when
``batch_first=True``, where :math:`L` is the target sequence length, :math:`B` is
the batch size, and :math:`E_q` is the query embedding dimension ``embed_dim``.
key (Tensor): Key embeddingd of shape :math:`(S, E_k)` for unbatched input,
:math:`(S, B, E_k)` when ``batch_first=False`` or :math:`(B, S, E_k)` when
``batch_first=True``, where :math:`S` is the source sequence length, :math:`B` is
the batch size and :math:`E_k` is the key embedding dimension ``kdim``.
value (Tensor): Value embeddings of shape :math:`(S, E_v)` for unbatched input,
:math:`(S, B, E_v)` when ``batch_first=False`` or :math:`(B, S, E_v)` when
``batch_first=True``, where :math:`S` is the source sequence length, :math:`B` is
the batch size and :math:`E_v` is the value embedding dimension ``vdim``.
key_padding_mask (Tensor | None, optional): If specified, a mask of shape
:math:`(B, S)` indicating which elements within ``key`` to ignore for the purpose
of attention (i.e. treat as "padding"). For unbatched `query`, shape should be
:math:`(S)`. Binary and float masks are supported. For a binary mask, a ``True``
value indicates that the corresponding ``key`` value will be ignored for the
purpose of attention. For a float mask, it will be directly added to the
corresponding ``key`` value. Defaults to ``None``.
need_weights (bool, optional): If specified, returns ``attn_output_weights`` in
addition to ``attn_outputs``. Set ``need_weights=False`` to use the optimized
``scale_dot_product_attention`` and achieve the best performance for MHA.
Defaults to ``False``.
attn_mask (Tensor | None, optional): If specified, a 2D or 3D mask preventing attention
to certain positions. Must be of shape :math:`(L,S)` or
:math:`(B \times \text{num_heads}, L, S)`, where :math:`B` is the batch size, :math:`L`
is the target sequence length, and :math:`S` is the source sequence length. A 2D mask
will be broadcasted across the batch while a 3D mask allows for a different mask for
each entry in the batch. Binary and float masks are supported. For a binary mask, a
``True`` value indicates that the corresponding position is not allowed to attend to.
For a float mask, the mask values will be added to the attention weight. If both
``attn_mask`` and ``key_padding_mask`` are provided, their types should match.
Defaults to ``None``.
average_attn_weights (bool, optional): If ``True``, indicates that the returned
``attn_weights`` should be averaged across heads. Otherwise, ``attn_weights`` are
provided separately per head. Note that this flag only has an effect when
``need_weights=True``. Defaults to ``True``.
is_causal (bool, optional): _description_. Defaults to ``False``.
Warning:
``need_weights=True`` and therefore ``average_attn_weights`` are not supported yet thus
have no effect.
Returns:
tuple[Tensor, None]:
- *attn_output* (Tensor): The output tensor of shape :math:`(L, E_q)`, :math:`(L, B, E_q)`
or :math:`(B, L, E_q)` where :math:`L` is the target sequence length, :math:`B` is
the batch size, and :math:`E_q` is the embedding dimension ``embed_dim``.
- *attn_output_weights* (None): Always ``None`` has we do not support
``need_weights=True`` yet.
"""
is_batched = query.dim() == 3
key_padding_mask = F._canonical_mask(
mask=key_padding_mask,
mask_name="key_padding_mask",
other_type=F._none_or_dtype(attn_mask),
other_name="attn_mask",
target_type=query.dtype,
)
attn_mask = F._canonical_mask(
mask=attn_mask,
mask_name="attn_mask",
other_type=None,
other_name="",
target_type=query.dtype,
check_other=False,
)
if self.batch_first and is_batched:
# make sure that the transpose op does not affect the "is" property
if key is value:
if query is key:
query = key = value = query.transpose(1, 0)
else:
query, key = (x.transpose(1, 0) for x in (query, key))
value = key
else:
query, key, value = (x.transpose(1, 0) for x in (query, key, value))
if not self._qkv_same_embed_dim:
(
attn_output,
_,
) = packed_multi_head_attention_forward(
query,
key,
value,
self.embed_dim,
self.num_heads,
self.num_groups,
self.in_proj_weight,
self.in_proj_bias,
self.bias_k,
self.bias_v,
self.add_zero_attn,
self.dropout,
self.out_proj.weight,
self.out_proj.bias,
training=self.training,
key_padding_mask=key_padding_mask,
need_weights=need_weights,
attn_mask=attn_mask,
use_separate_proj_weight=True,
q_proj_weight=self.q_proj_weight,
k_proj_weight=self.k_proj_weight,
v_proj_weight=self.v_proj_weight,
average_attn_weights=average_attn_weights,
is_causal=is_causal,
)
else:
(
attn_output,
_,
) = packed_multi_head_attention_forward(
query,
key,
value,
self.embed_dim,
self.num_heads,
self.num_groups,
self.in_proj_weight,
self.in_proj_bias,
self.bias_k,
self.bias_v,
self.add_zero_attn,
self.dropout,
self.out_proj.weight,
self.out_proj.bias,
training=self.training,
key_padding_mask=key_padding_mask,
need_weights=need_weights,
attn_mask=attn_mask,
average_attn_weights=average_attn_weights,
is_causal=is_causal,
)
if self.batch_first and is_batched:
return attn_output.transpose(1, 0), None
return attn_output, None
