forked from thuml/Autoformer
-
Notifications
You must be signed in to change notification settings - Fork 0
/
AutoCorrelation.py
167 lines (150 loc) · 6.63 KB
/
AutoCorrelation.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
167
import torch
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
import numpy as np
import math
from math import sqrt
import os
class AutoCorrelation(nn.Module):
"""
AutoCorrelation Mechanism with the following two phases:
(1) period-based dependencies discovery
(2) time delay aggregation
This block can replace the self-attention family mechanism seamlessly.
"""
def __init__(self, mask_flag=True, factor=1, scale=None, attention_dropout=0.1, output_attention=False):
super(AutoCorrelation, self).__init__()
self.factor = factor
self.scale = scale
self.mask_flag = mask_flag
self.output_attention = output_attention
self.dropout = nn.Dropout(attention_dropout)
def time_delay_agg_training(self, values, corr):
"""
SpeedUp version of Autocorrelation (a batch-normalization style design)
This is for the training phase.
"""
head = values.shape[1]
channel = values.shape[2]
length = values.shape[3]
# find top k
top_k = int(self.factor * math.log(length))
mean_value = torch.mean(torch.mean(corr, dim=1), dim=1)
index = torch.topk(torch.mean(mean_value, dim=0), top_k, dim=-1)[1]
weights = torch.stack([mean_value[:, index[i]] for i in range(top_k)], dim=-1)
# update corr
tmp_corr = torch.softmax(weights, dim=-1)
# aggregation
tmp_values = values
delays_agg = torch.zeros_like(values).float()
for i in range(top_k):
pattern = torch.roll(tmp_values, -int(index[i]), -1)
delays_agg = delays_agg + pattern * \
(tmp_corr[:, i].unsqueeze(1).unsqueeze(1).unsqueeze(1).repeat(1, head, channel, length))
return delays_agg
def time_delay_agg_inference(self, values, corr):
"""
SpeedUp version of Autocorrelation (a batch-normalization style design)
This is for the inference phase.
"""
batch = values.shape[0]
head = values.shape[1]
channel = values.shape[2]
length = values.shape[3]
# index init
device = torch.device('cuda:0')
# init_index = torch.arange(length).unsqueeze(0).unsqueeze(0).unsqueeze(0).repeat(batch, head, channel, 1).cuda()
init_index = torch.arange(length).unsqueeze(0).unsqueeze(0).unsqueeze(0).repeat(batch, head, channel, 1).to(device)
# self.a
# find top k
top_k = int(self.factor * math.log(length))
mean_value = torch.mean(torch.mean(corr, dim=1), dim=1)
weights, delay = torch.topk(mean_value, top_k, dim=-1)
# update corr
tmp_corr = torch.softmax(weights, dim=-1)
# aggregation
tmp_values = values.repeat(1, 1, 1, 2)
delays_agg = torch.zeros_like(values).float()
for i in range(top_k):
tmp_delay = init_index + delay[:, i].unsqueeze(1).unsqueeze(1).unsqueeze(1).repeat(1, head, channel, length)
pattern = torch.gather(tmp_values, dim=-1, index=tmp_delay)
delays_agg = delays_agg + pattern * \
(tmp_corr[:, i].unsqueeze(1).unsqueeze(1).unsqueeze(1).repeat(1, head, channel, length))
return delays_agg
def time_delay_agg_full(self, values, corr):
"""
Standard version of Autocorrelation
"""
batch = values.shape[0]
head = values.shape[1]
channel = values.shape[2]
length = values.shape[3]
# index init
device = torch.device('cuda:0')
# init_index = torch.arange(length).unsqueeze(0).unsqueeze(0).unsqueeze(0).repeat(batch, head, channel, 1).cuda()
init_index = torch.arange(length).unsqueeze(0).unsqueeze(0).unsqueeze(0).repeat(batch, head, channel, 1).to(device)
# find top k
top_k = int(self.factor * math.log(length))
weights, delay = torch.topk(corr, top_k, dim=-1)
# update corr
tmp_corr = torch.softmax(weights, dim=-1)
# aggregation
tmp_values = values.repeat(1, 1, 1, 2)
delays_agg = torch.zeros_like(values).float()
for i in range(top_k):
tmp_delay = init_index + delay[..., i].unsqueeze(-1)
pattern = torch.gather(tmp_values, dim=-1, index=tmp_delay)
delays_agg = delays_agg + pattern * (tmp_corr[..., i].unsqueeze(-1))
return delays_agg
def forward(self, queries, keys, values, attn_mask):
B, L, H, E = queries.shape
_, S, _, D = values.shape
if L > S:
zeros = torch.zeros_like(queries[:, :(L - S), :]).float()
values = torch.cat([values, zeros], dim=1)
keys = torch.cat([keys, zeros], dim=1)
else:
values = values[:, :L, :, :]
keys = keys[:, :L, :, :]
# period-based dependencies
q_fft = torch.fft.rfft(queries.permute(0, 2, 3, 1).contiguous(), dim=-1)
k_fft = torch.fft.rfft(keys.permute(0, 2, 3, 1).contiguous(), dim=-1)
res = q_fft * torch.conj(k_fft)
corr = torch.fft.irfft(res, dim=-1)
# time delay agg
if self.training:
V = self.time_delay_agg_training(values.permute(0, 2, 3, 1).contiguous(), corr).permute(0, 3, 1, 2)
else:
V = self.time_delay_agg_inference(values.permute(0, 2, 3, 1).contiguous(), corr).permute(0, 3, 1, 2)
if self.output_attention:
return (V.contiguous(), corr.permute(0, 3, 1, 2))
else:
return (V.contiguous(), None)
class AutoCorrelationLayer(nn.Module):
def __init__(self, correlation, d_model, n_heads, d_keys=None,
d_values=None):
super(AutoCorrelationLayer, self).__init__()
d_keys = d_keys or (d_model // n_heads)
d_values = d_values or (d_model // n_heads)
self.inner_correlation = correlation
self.query_projection = nn.Linear(d_model, d_keys * n_heads)
self.key_projection = nn.Linear(d_model, d_keys * n_heads)
self.value_projection = nn.Linear(d_model, d_values * n_heads)
self.out_projection = nn.Linear(d_values * n_heads, d_model)
self.n_heads = n_heads
def forward(self, queries, keys, values, attn_mask):
B, L, _ = queries.shape
_, S, _ = keys.shape
H = self.n_heads
queries = self.query_projection(queries).view(B, L, H, -1)
keys = self.key_projection(keys).view(B, S, H, -1)
values = self.value_projection(values).view(B, S, H, -1)
out, attn = self.inner_correlation(
queries,
keys,
values,
attn_mask
)
out = out.view(B, L, -1)
return self.out_projection(out), attn