/
bddpm.py
224 lines (184 loc) · 8.82 KB
/
bddpm.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
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
import torch
import torch.nn as nn
import torch.nn.functional as F
class SelfAttention(nn.Module):
def __init__(self, in_channels, grid_size):
super(SelfAttention, self).__init__()
self.in_channels = in_channels
self.grid_size = grid_size
self.multiatt = nn.MultiheadAttention(in_channels, 4)
self.layernorm = nn.LayerNorm([in_channels])
self.ff_self = nn.Sequential(
nn.LayerNorm([in_channels]),
nn.Linear(in_channels, in_channels),
nn.GELU(),
nn.Linear(in_channels, in_channels),
)
def forward(self, x):
x = x.view(-1, self.in_channels,
self.grid_size[0] * self.grid_size[1]).swapaxes(1, 2)
x_layernorm = self.layernorm(x)
x_layernorm = x_layernorm.transpose(0, 1)
attention_value, _ = self.multiatt(x_layernorm, x_layernorm,
x_layernorm)
attention_value = attention_value.transpose(0, 1)
attention_value = attention_value + x
attention_value = self.ff_self(attention_value) + attention_value
return attention_value.swapaxes(2, 1).view(-1, self.in_channels,
self.grid_size[0],
self.grid_size[1])
class DoubleConv(nn.Module):
def __init__(self,
in_channels,
out_channels,
mid_channels=None,
residual=False):
super(DoubleConv, self).__init__()
self.residual = residual
if not mid_channels:
mid_channels = out_channels
self.double_conv = nn.Sequential(
nn.Conv2d(in_channels,
mid_channels,
kernel_size=3,
padding=1,
bias=False),
nn.GroupNorm(1, mid_channels),
nn.GELU(),
nn.Conv2d(mid_channels,
out_channels,
kernel_size=3,
padding=1,
bias=False),
nn.GroupNorm(1, out_channels),
)
def forward(self, x):
if self.residual:
return F.gelu(x + self.double_conv(x))
else:
return self.double_conv(x)
class Down(nn.Module):
def __init__(self, in_channels, out_channels, hid_channels=256):
super(Down, self).__init__()
self.maxpool_conv = nn.Sequential(
nn.MaxPool2d(2),
DoubleConv(in_channels, in_channels, residual=True),
DoubleConv(in_channels, out_channels),
)
self.emb_layer = nn.Sequential(
nn.SiLU(),
nn.Linear(hid_channels, out_channels),
)
def forward(self, x, t):
x = self.maxpool_conv(x)
emb = self.emb_layer(t)[:, :, None,
None].repeat(1, 1, x.shape[-2], x.shape[-1])
return x + emb
class Up(nn.Module):
def __init__(self,
in_channels,
out_channels,
upsample_size,
hid_channels=256):
super(Up, self).__init__()
self.up = nn.Upsample(size=upsample_size,
mode="bilinear",
align_corners=True)
self.conv = nn.Sequential(
DoubleConv(in_channels, in_channels, residual=True),
DoubleConv(in_channels, out_channels, in_channels // 2),
)
self.emb_layer = nn.Sequential(
nn.SiLU(),
nn.Linear(hid_channels, out_channels),
)
def forward(self, x, skip_x, t):
x = self.up(x)
x = torch.cat([skip_x, x], dim=1)
x = self.conv(x)
emb = self.emb_layer(t)[:, :, None,
None].repeat(1, 1, x.shape[-2], x.shape[-1])
return x + emb
class BUNet(nn.Module):
r'''
The diffusion model consists of two processes, the forward process, and the backward process. The forward process is to gradually add Gaussian noise to an image until it becomes random noise, while the backward process is the de-noising process. We train an attention-based UNet network at the backward process to start with random noise and gradually de-noise it until an image is generated and use the UNet to generate a simulated image from random noises.
It is worth noting that this model is not designed for EEG analysis, but shows good performance and can serve as a good research start.
- Paper: Ho J, Jain A, Abbeel P. Denoising diffusion probabilistic models[J]. Advances in Neural Information Processing Systems, 2020, 33: 6840-6851.
- URL: https://arxiv.org/abs/2006.11239
- Related Project: https://github.com/dome272/Diffusion-Models-pytorch
Below is a recommended suite for use in EEG generation:
.. code-block:: python
noise = torch.randn(1, 4, 9, 9)
t = torch.randint(low=1, high=1000, size=(1, ))
unet = BUNet()
fake_X = unet(noise, t)
Args:
in_channels (int): The feature dimension of each electrode. (default: :obj:`4`)
hid_channels (int): The basic hidden channels in the network blocks. (default: :obj:`64`)
grid_size (tuple): Spatial dimensions of grid-like EEG representation. (default: :obj:`(9, 9)`)
beta_timesteps (int): The variance schedule controlling step sizes. (default: :obj:`256`)
'''
def __init__(self,
in_channels=4,
hid_channels=64,
grid_size=(9, 9),
beta_timesteps=256):
super(BUNet, self).__init__()
self.in_channels = in_channels
self.hid_channels = hid_channels
self.beta_timesteps = beta_timesteps
self.inc = DoubleConv(in_channels, hid_channels)
self.down1 = Down(hid_channels, hid_channels * 2)
att1_grid_size = (grid_size[0] // 2, grid_size[1] // 2)
self.att1 = SelfAttention(hid_channels * 2, att1_grid_size)
self.down2 = Down(hid_channels * 2, hid_channels * 4)
att2_grid_size = (att1_grid_size[0] // 2, att1_grid_size[1] // 2)
self.att2 = SelfAttention(hid_channels * 4, att2_grid_size)
self.down3 = Down(hid_channels * 4, hid_channels * 4)
att3_grid_size = (att2_grid_size[0] // 2, att2_grid_size[1] // 2)
self.att3 = SelfAttention(hid_channels * 4, att3_grid_size)
self.bot1 = DoubleConv(hid_channels * 4, hid_channels * 8)
self.bot2 = DoubleConv(hid_channels * 8, hid_channels * 8)
self.bot3 = DoubleConv(hid_channels * 8, hid_channels * 4)
self.up1 = Up(hid_channels * 8, hid_channels * 2, att2_grid_size)
self.att4 = SelfAttention(hid_channels * 2, att2_grid_size)
self.up2 = Up(hid_channels * 4, hid_channels, att1_grid_size)
self.att5 = SelfAttention(hid_channels, att1_grid_size)
self.up3 = Up(hid_channels * 2, hid_channels, grid_size)
self.att6 = SelfAttention(hid_channels, grid_size)
self.outc = nn.Conv2d(hid_channels, in_channels, kernel_size=1)
def position_encoding(self, t, channels):
inv_freq = 1.0 / (10000**(
torch.arange(0, channels, 2, device=t.device).float() / channels))
pos_enc_a = torch.sin(t.repeat(1, channels // 2) * inv_freq)
pos_enc_b = torch.cos(t.repeat(1, channels // 2) * inv_freq)
pos_enc = torch.cat([pos_enc_a, pos_enc_b], dim=-1)
return pos_enc
def forward(self, x: torch.Tensor, t: torch.Tensor):
r'''
Args:
x (torch.Tensor): The random noise to be denoised, which should have the same shape as the simulated EEG expected to be generated, i.e., :obj:`[n, 4, 9, 9]`. Here, :obj:`n` corresponds to the batch size, :obj:`4` corresponds to :obj:`in_channels`, and :obj:`(9, 9)` corresponds to :obj:`grid_size`.
t (torch.Tensor): The randomly sampled time steps (int) for denoising a batch of samples. The shape should be :obj:`[n,]`. Here, :obj:`n` corresponds to the batch size.
Returns:
torch.Tensor[n, 4, 9, 9]: the denoised results, which should have the same shape as the input noise, i.e., :obj:`[n, 4, 9, 9]`. Here, :obj:`n` corresponds to the batch size, :obj:`4` corresponds to :obj:`in_channels`, and :obj:`(9, 9)` corresponds to :obj:`grid_size`.
'''
t = t.unsqueeze(-1).float()
t = self.position_encoding(t, self.beta_timesteps)
x1 = self.inc(x)
x2 = self.down1(x1, t)
x2 = self.att1(x2)
x3 = self.down2(x2, t)
x3 = self.att2(x3)
x4 = self.down3(x3, t)
x4 = self.att3(x4)
x4 = self.bot1(x4)
x4 = self.bot2(x4)
x4 = self.bot3(x4)
x = self.up1(x4, x3, t)
x = self.att4(x)
x = self.up2(x, x2, t)
x = self.att5(x)
x = self.up3(x, x1, t)
x = self.att6(x)
output = self.outc(x)
return output