1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
|
"""Nyströmer encoder.
Efficient attention module that reduces the complexity of the attention module from
O(n**2) to O(n). The attention matrix is assumed low rank and thus the information
can be represented by a smaller matrix.
Stolen from:
https://github.com/lucidrains/nystrom-attention/blob/main/nystrom_attention/nystrom_attention.py
"""
from math import ceil
from typing import Optional, Tuple, Union
from einops import rearrange, reduce
import torch
from torch import einsum, nn, Tensor
from torch.nn import functional as F
def moore_penrose_inverse(x: Tensor, iters: int = 6) -> Tensor:
"""Moore-Penrose pseudoinverse."""
x_abs = torch.abs(x)
col = x_abs.sum(dim=-1)
row = x_abs.sum(dim=-2)
z = rearrange(x, "... i j -> ... j i") / (torch.max(col) * torch.max(row))
I = torch.eye(x.shape[-1], device=x.device)
I = rearrange(I, "i j -> () i j")
for _ in range(iters):
xz = x @ z
z = 0.25 * z @ (13 * I - (xz @ (15 * I - (xz @ (7 * I - xz)))))
return z
class NystromAttention(nn.Module):
def __init__(
self,
dim: int,
dim_head: int = 64,
num_heads: int = 8,
num_landmarks: int = 256,
inverse_iter: int = 6,
residual: bool = True,
residual_conv_kernel: int = 13,
eps: float = 1.0e-8,
dropout_rate: float = 0.0,
):
super().__init__()
self.residual = None
self.eps = eps
self.num_heads = num_heads
inner_dim = self.num_heads * dim_head
self.num_landmarks = num_landmarks
self.inverse_iter = inverse_iter
self.scale = dim_head ** -0.5
self.qkv_fn = nn.Linear(dim, 3 * inner_dim, bias=False)
self.fc_out = nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout_rate))
if residual:
self.residual = nn.Conv2d(
in_channels=num_heads,
out_channels=num_heads,
kernel_size=(residual_conv_kernel, 1),
padding=(residual_conv_kernel // 2, 0),
groups=num_heads,
bias=False,
)
@staticmethod
def _pad_sequence(
x: Tensor, mask: Optional[Tensor], n: int, m: int
) -> Tuple[Tensor, Tensor]:
"""Pad sequence."""
padding = m - (n % m)
x = F.pad(x, (0, 0, padding, 0), value=0)
mask = F.pad(mask, (padding, 0), value=False) if mask is not None else mask
return x, mask
def _compute_landmarks(
self, q: Tensor, k: Tensor, mask: Optional[Tensor], n: int, m: int
) -> Tuple[Tensor, Tensor, Optional[Tensor]]:
"""Compute landmarks of the attention matrix."""
divisor = ceil(n / m)
landmark_einops_eq = "... (n l) d -> ... n d"
q_landmarks = reduce(q, landmark_einops_eq, "sum", l=divisor)
k_landmarks = reduce(k, landmark_einops_eq, "sum", l=divisor)
mask_landmarks = None
if mask is not None:
mask_landmarks_sum = reduce(mask, "... (n l) -> ... n", "sum", l=divisor)
divisor = mask_landmarks_sum[..., None] + self.eps
mask_landmarks = mask_landmarks_sum > 0
q_landmarks /= divisor
k_landmarks /= divisor
return q_landmarks, k_landmarks, mask_landmarks
@staticmethod
def _compute_similarities(
q: Tensor,
k: Tensor,
q_landmarks: Tensor,
k_landmarks: Tensor,
mask: Optional[Tensor],
mask_landmarks: Optional[Tensor],
) -> Tuple[Tensor, Tensor, Tensor]:
einops_eq = "... i d, ... j d -> ... i j"
sim1 = einsum(einops_eq, q, k_landmarks)
sim2 = einsum(einops_eq, q_landmarks, k_landmarks)
sim3 = einsum(einops_eq, q_landmarks, k)
if mask is not None and mask_landmarks is not None:
mask_value = -torch.finfo(q.type).max
sim1.masked_fill_(
~(mask[..., None] * mask_landmarks[..., None, :]), mask_value
)
sim2.masked_fill_(
~(mask_landmarks[..., None] * mask_landmarks[..., None, :]), mask_value
)
sim3.masked_fill_(
~(mask_landmarks[..., None] * mask[..., None, :]), mask_value
)
return sim1, sim2, sim3
def _nystrom_attention(
self,
q: Tensor,
k: Tensor,
v: Tensor,
mask: Optional[Tensor],
n: int,
m: int,
return_attn: bool,
) -> Tuple[Tensor, Optional[Tensor]]:
q_landmarks, k_landmarks, mask_landmarks = self._compute_landmarks(
q, k, mask, n, m
)
sim1, sim2, sim3 = self._compute_similarities(
q, k, q_landmarks, k_landmarks, mask, mask_landmarks
)
# Compute attention
attn1, attn2, attn3 = map(lambda t: t.softmax(dim=-1), (sim1, sim2, sim3))
attn2_inv = moore_penrose_inverse(attn2, self.inverse_iter)
out = (attn1 @ attn2_inv) @ (attn3 @ v)
if return_attn:
return out, attn1 @ attn2_inv @ attn3
return out, None
def forward(
self, x: Tensor, mask: Optional[Tensor] = None, return_attn: bool = False
) -> Union[Tensor, Tuple[Tensor, Tensor]]:
"""Compute the Nystrom attention."""
_, n, _, h, m = x.shape, self.num_heads
if n % m != 0:
x, mask = self._pad_sequence(x, mask, n, m)
q, k, v = self.qkv_fn(x).chunk(3, dim=-1)
q, k, v = map(lambda t: rearrange(t, "b n (h d) -> b h n d", h=h), (q, k, v))
q *= self.scale
out, attn = self._nystrom_attention(q, k, v, mask, n, m, return_attn)
# Add depth-wise convolutional residual of values
if self.residual is not None:
out += self.residual(out)
out = rearrange(out, "b h n d -> b n (h d)", h=h)
out = self.fc_out(out)
out = out[:, -n:]
if return_attn:
return out, attn
return out
|
