-
Notifications
You must be signed in to change notification settings - Fork 1
/
MHA.py
166 lines (114 loc) · 5.51 KB
/
MHA.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
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
# Associative Transformer Is A Sparse Representation Learner
# See https://arxiv.org/abs/2309.12862
#
# Author: Yuwei Sun
#
# The code is based on the
# https://github.com/lucidrains/vit-pytorch/blob/main/vit_pytorch/vit.py
import torch
import torch.nn as nn
import torch.nn.functional as F
class ScaledDotProductAttention(nn.Module):
''' Scaled Dot-Product Attention '''
def __init__(self, temperature, attn_dropout=0.1):
super().__init__()
self.temperature = temperature
self.dropout = nn.Dropout(attn_dropout)
def cv_squared(self, x):
eps = 1e-10
# if only num_experts = 1
if x.shape[0] == 1:
return torch.tensor([0], device=x.device, dtype=x.dtype)
return x.float().var() / (x.float().mean()**2 + eps)
def forward(self, q, k, v, bottleneck, coe = 1e-2):
loss = 0
output = []
gates = []
#for each head
for i in range(q.size(1)):
attn = torch.matmul(q / self.temperature, k.transpose(2, 3))[0][i]
# calculate topk + 1 that will be needed for the noisy gates
top_logits, top_indices = attn.topk(min(bottleneck + 1, attn.size(1)), dim=-1)
top_k_logits = top_logits[:, :bottleneck]
top_k_indices = top_indices[:, :bottleneck]
top_k_gates = F.softmax(top_k_logits, dim = -1)
zeros = torch.zeros_like(attn, requires_grad=True)
gate = zeros.scatter(1, top_k_indices, top_k_gates)
output.append(torch.matmul(gate, v[0,i,:,:]))
gates.append(gate)
# calculate importance loss and load loss for balanced expert usage
importance = gate.sum(0)
loads = (gate > 0).sum(0)
loss += (self.cv_squared(importance) + self.cv_squared(loads)) * coe
#stack results of all heads
output = torch.stack(output, dim = 1)
gates = torch.stack(gates, dim = 1)
return output, gates, loss
class MultiHeadAttention(nn.Module):
''' Multi-Head Attention module '''
def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
super().__init__()
self.n_head = n_head
self.d_k = d_k
self.d_v = d_v
self.w_qs = nn.Linear(d_k, n_head * d_k, bias=False)
self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False)
self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False)
self.fc = nn.Linear(n_head * d_v, d_v, bias=False)
self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5)
self.dropout = nn.Dropout(dropout)
self.layer_norm = nn.LayerNorm(self.d_v, eps=1e-6)
def forward(self, q, k, v, bottleneck, mask=None):
d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1)
residual = q
# Pass through the pre-attention projection: b x lq x (n*dv)
# Separate different heads: b x lq x n x dv
q = self.w_qs(q).view(sz_b, len_q, n_head, d_k)
k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)
# Transpose for attention dot product: b x n x lq x dv
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
if mask is not None:
mask = mask.unsqueeze(0).expand(n_head,*(mask.size(0),mask.size(1))) # For head axis broadcasting.
q, gates, loss = self.attention(q, k, v, bottleneck)
# Transpose to move the head dimension back: b x lq x n x dv
# Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv)
q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1)
q = self.dropout(self.fc(q))
q += residual
q = self.layer_norm(q)
return q, gates, loss
class HopfieldReplaceMHA(nn.Module):
''' Multi-Head Attention module '''
def __init__(self, n_head, d_model, d_k, d_v, dropout=0.1):
super().__init__()
self.n_head = n_head
self.d_k = d_k
self.d_v = d_v
self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False)
self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False)
self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False)
self.fc = nn.Linear(n_head * d_v, d_model, bias=False)
self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5)
self.dropout = nn.Dropout(dropout)
self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
def forward(self, q, k, v):
d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1)
residual = q
# Pass through the pre-attention projection: b x lq x (n*dv)
# Separate different heads: b x lq x n x dv
q = self.w_qs(q).view(sz_b, len_q, n_head, d_k)
k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)
# Transpose for attention dot product: b x n x lq x dv
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
q, attn, _ = self.attention(q, k, v, v.size(1))
# Transpose to move the head dimension back: b x lq x n x dv
# Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv)
q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1)
q = self.dropout(self.fc(q))
q += residual
q = self.layer_norm(q)
return q, attn