-
Notifications
You must be signed in to change notification settings - Fork 0
/
td_layer.py
143 lines (127 loc) · 3.82 KB
/
td_layer.py
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
"""
This file contains an implementation of a TransformerEncoderLayer.
Explicit differences from nn.TransformerEncoderLayer:
- No need for {tgt/memory}_padding_mask with nested tensors :)
- Only supports batch_first=True: nested tensors do not support seq_len as the
first dimension
"""
import torch
import torch.nn as nn
from mha import MultiHeadAttention
from torch import Tensor
from typing import Optional
class TransformerDecoderLayer(nn.Module):
def __init__(
self,
d_model,
nhead,
dim_feedforward=2048,
dropout=0.1,
activation : nn.Module = torch.nn.functional.relu,
layer_norm_eps=1e-5,
norm_first = False,
bias=True,
device=None,
dtype=None,
):
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.self_attn = MultiHeadAttention(
d_model,
d_model,
d_model,
d_model,
nhead,
dropout=dropout,
bias=bias,
**factory_kwargs,
)
self.multihead_attn = MultiHeadAttention(
d_model,
d_model,
d_model,
d_model,
nhead,
dropout=dropout,
bias=bias,
**factory_kwargs,
)
self.linear1 = nn.Linear(d_model, dim_feedforward, bias=bias, **factory_kwargs)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model, bias=bias, **factory_kwargs)
self.norm_first = norm_first
self.norm1 = nn.LayerNorm(d_model, eps=layer_norm_eps, bias=bias, **factory_kwargs)
self.norm2 = nn.LayerNorm(d_model, eps=layer_norm_eps, bias=bias, **factory_kwargs)
self.norm3 = nn.LayerNorm(d_model, eps=layer_norm_eps, bias=bias, **factory_kwargs)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
self.activation = activation
# self-attention block
def _sa_block(
self,
x: Tensor,
attn_mask: Optional[Tensor],
is_causal: bool = False,
) -> Tensor:
x = self.self_attn(
x,
x,
x,
attn_mask=attn_mask,
is_causal=is_causal,
)
return self.dropout1(x)
# multihead attention block
def _mha_block(
self,
x: Tensor,
mem: Tensor,
attn_mask: Optional[Tensor],
is_causal: bool = False,
) -> Tensor:
x = self.multihead_attn(
x,
mem,
mem,
attn_mask=attn_mask,
is_causal=is_causal,
)
return self.dropout2(x)
# feed forward block
def _ff_block(self, x: Tensor) -> Tensor:
x = self.linear2(self.dropout(self.activation(self.linear1(x))))
return self.dropout3(x)
def forward(
self,
tgt: Tensor,
memory: Tensor,
tgt_mask: Optional[Tensor] = None,
memory_mask: Optional[Tensor] = None,
tgt_is_causal=False,
memory_is_causal=False,
):
x = tgt
if self.norm_first:
x = x + self._sa_block(
self.norm1(x), tgt_mask, tgt_is_causal
)
x = x + self._mha_block(
self.norm2(x),
memory,
memory_mask,
memory_is_causal,
)
x = x + self._ff_block(self.norm3(x))
else:
x = self.norm1(
x + self._sa_block(x, tgt_mask, tgt_is_causal)
)
x = self.norm2(
x
+ self._mha_block(
x, memory, memory_mask, memory_is_causal
)
)
x = self.norm3(x + self._ff_block(x))
return x