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
|
"""Implementes the attention module for the transformer."""
from typing import Optional, Tuple
import attr
from einops import rearrange
from einops.layers.torch import Rearrange
import torch
from torch import einsum
from torch import nn
from torch import Tensor
import torch.nn.functional as F
from text_recognizer.networks.transformer.positional_encodings.rotary_embedding import (
apply_rotary_pos_emb,
)
@attr.s
class Attention(nn.Module):
"""Standard attention."""
def __attrs_pre_init__(self) -> None:
super().__init__()
dim: int = attr.ib()
num_heads: int = attr.ib()
dim_head: int = attr.ib(default=64)
dropout_rate: float = attr.ib(default=0.0)
casual: bool = attr.ib(default=False)
scale: float = attr.ib(init=False)
dropout: nn.Dropout = attr.ib(init=False)
fc: nn.Linear = attr.ib(init=False)
qkv_fn: nn.Sequential = attr.ib(init=False)
attn_fn: F.softmax = attr.ib(init=False, default=F.softmax)
def __attrs_post_init__(self) -> None:
"""Post init configuration."""
self.scale = self.dim ** -0.5
inner_dim = self.dim * self.dim_head
# Attnetion
self.qkv_fn = nn.Sequential(
nn.Linear(self.dim, 3 * inner_dim, bias=False),
Rearrange("b n (qkv h d) -> qkv b h n d", qkv=3, h=self.num_heads),
)
self.dropout = nn.Dropout(p=self.dropout_rate)
# Feedforward
self.fc = nn.Linear(inner_dim, self.dim)
@staticmethod
def _apply_rotary_emb(
q: Tensor, k: Tensor, rotary_pos_emb: Tensor
) -> Tuple[Tensor, Tensor]:
l = rotary_pos_emb.shape[-1]
(ql, qr), (kl, kr) = map(lambda t: (t[..., :l], t[..., l:]), (q, k))
ql, kl = apply_rotary_pos_emb(ql, kl, rotary_pos_emb)
q = torch.cat((ql, qr), dim=-1)
k = torch.cat((kl, kr), dim=-1)
return q, k
@staticmethod
def _compute_input_mask(
b: int,
n: int,
k: Tensor,
mask: Optional[Tensor],
context: Optional[Tensor],
context_mask: Optional[Tensor],
device: str,
) -> Optional[Tensor]:
if any(x is not None for x in (mask, context_mask)):
q_mask = (
mask if mask is not None else torch.ones((b, n), device=device).bool()
)
k_mask = q_mask if context is not None else context_mask
k_mask = (
torch.ones((b, k.shape[-2]), device=device).bool()
if k_mask is None
else k_mask
)
q_mask = rearrange(q_mask, "b i -> b () i ()")
k_mask = rearrange(k_mask, "b i -> b () () j")
return q_mask * k_mask
return
@staticmethod
def _apply_causal_mask(
energy: Tensor, mask: Tensor, mask_value: Tensor, device: str
) -> Tensor:
i, j = energy.shape[-2:]
r = torch.arange(i, device=device)
mask = rearrange(r, "i -> () () i ()") < rearrange(r, "j -> () () () j")
mask = F.pad(mask, (j - i, 0), value=False)
energy.masked_fill_(mask, mask_value)
del mask
return energy
def forward(
self,
x: Tensor,
context: Optional[Tensor] = None,
mask: Optional[Tensor] = None,
context_mask: Optional[Tensor] = None,
rotary_pos_emb: Optional[Tensor] = None,
) -> Tuple[Tensor, Tensor]:
b, n, _, device = *x.shape, x.device
q, k, v = self.qkv_fn(x)
q, k = (
self._apply_rotary_emb(q, k, rotary_pos_emb)
if rotary_pos_emb is not None
else q,
k,
)
input_mask = self._compute_input_mask(
b, n, k, mask, context, context_mask, device
)
# Compute the attention
energy = einsum("b h i d, b h j d -> b h i j", q, k) * self.scale
mask_value = -torch.finfo(energy.dtype).max
# Apply input mask
if input_mask is not None:
energy = energy.masked_fill_(~input_mask, mask_value)
del input_mask
if self.causal:
energy = self._apply_causal_mask(energy, mask, mask_value, device)
attn = self.attn_fn(energy, dim=-1)
attn = self.dropout(attn)
out = einsum("b h i j, b h j d -> b h i d", attn, v)
out = rearrange(out, "b h n d -> b n (h d)")
out = self.fc(out)
return out, attn
|
