-
Notifications
You must be signed in to change notification settings - Fork 84
/
diff_models.py
180 lines (147 loc) · 6.57 KB
/
diff_models.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from linear_attention_transformer import LinearAttentionTransformer
def get_torch_trans(heads=8, layers=1, channels=64):
encoder_layer = nn.TransformerEncoderLayer(
d_model=channels, nhead=heads, dim_feedforward=64, activation="gelu"
)
return nn.TransformerEncoder(encoder_layer, num_layers=layers)
def get_linear_trans(heads=8,layers=1,channels=64,localheads=0,localwindow=0):
return LinearAttentionTransformer(
dim = channels,
depth = layers,
heads = heads,
max_seq_len = 256,
n_local_attn_heads = 0,
local_attn_window_size = 0,
)
def Conv1d_with_init(in_channels, out_channels, kernel_size):
layer = nn.Conv1d(in_channels, out_channels, kernel_size)
nn.init.kaiming_normal_(layer.weight)
return layer
class DiffusionEmbedding(nn.Module):
def __init__(self, num_steps, embedding_dim=128, projection_dim=None):
super().__init__()
if projection_dim is None:
projection_dim = embedding_dim
self.register_buffer(
"embedding",
self._build_embedding(num_steps, embedding_dim / 2),
persistent=False,
)
self.projection1 = nn.Linear(embedding_dim, projection_dim)
self.projection2 = nn.Linear(projection_dim, projection_dim)
def forward(self, diffusion_step):
x = self.embedding[diffusion_step]
x = self.projection1(x)
x = F.silu(x)
x = self.projection2(x)
x = F.silu(x)
return x
def _build_embedding(self, num_steps, dim=64):
steps = torch.arange(num_steps).unsqueeze(1) # (T,1)
frequencies = 10.0 ** (torch.arange(dim) / (dim - 1) * 4.0).unsqueeze(0) # (1,dim)
table = steps * frequencies # (T,dim)
table = torch.cat([torch.sin(table), torch.cos(table)], dim=1) # (T,dim*2)
return table
class diff_CSDI(nn.Module):
def __init__(self, config, inputdim=2):
super().__init__()
self.channels = config["channels"]
self.diffusion_embedding = DiffusionEmbedding(
num_steps=config["num_steps"],
embedding_dim=config["diffusion_embedding_dim"],
)
self.input_projection = Conv1d_with_init(inputdim, self.channels, 1)
self.output_projection1 = Conv1d_with_init(self.channels, self.channels, 1)
self.output_projection2 = Conv1d_with_init(self.channels, 1, 1)
nn.init.zeros_(self.output_projection2.weight)
self.residual_layers = nn.ModuleList(
[
ResidualBlock(
side_dim=config["side_dim"],
channels=self.channels,
diffusion_embedding_dim=config["diffusion_embedding_dim"],
nheads=config["nheads"],
is_linear=config["is_linear"],
)
for _ in range(config["layers"])
]
)
def forward(self, x, cond_info, diffusion_step):
B, inputdim, K, L = x.shape
x = x.reshape(B, inputdim, K * L)
x = self.input_projection(x)
x = F.relu(x)
x = x.reshape(B, self.channels, K, L)
diffusion_emb = self.diffusion_embedding(diffusion_step)
skip = []
for layer in self.residual_layers:
x, skip_connection = layer(x, cond_info, diffusion_emb)
skip.append(skip_connection)
x = torch.sum(torch.stack(skip), dim=0) / math.sqrt(len(self.residual_layers))
x = x.reshape(B, self.channels, K * L)
x = self.output_projection1(x) # (B,channel,K*L)
x = F.relu(x)
x = self.output_projection2(x) # (B,1,K*L)
x = x.reshape(B, K, L)
return x
class ResidualBlock(nn.Module):
def __init__(self, side_dim, channels, diffusion_embedding_dim, nheads, is_linear=False):
super().__init__()
self.diffusion_projection = nn.Linear(diffusion_embedding_dim, channels)
self.cond_projection = Conv1d_with_init(side_dim, 2 * channels, 1)
self.mid_projection = Conv1d_with_init(channels, 2 * channels, 1)
self.output_projection = Conv1d_with_init(channels, 2 * channels, 1)
self.is_linear = is_linear
if is_linear:
self.time_layer = get_linear_trans(heads=nheads,layers=1,channels=channels)
self.feature_layer = get_linear_trans(heads=nheads,layers=1,channels=channels)
else:
self.time_layer = get_torch_trans(heads=nheads, layers=1, channels=channels)
self.feature_layer = get_torch_trans(heads=nheads, layers=1, channels=channels)
def forward_time(self, y, base_shape):
B, channel, K, L = base_shape
if L == 1:
return y
y = y.reshape(B, channel, K, L).permute(0, 2, 1, 3).reshape(B * K, channel, L)
if self.is_linear:
y = self.time_layer(y.permute(0, 2, 1)).permute(0, 2, 1)
else:
y = self.time_layer(y.permute(2, 0, 1)).permute(1, 2, 0)
y = y.reshape(B, K, channel, L).permute(0, 2, 1, 3).reshape(B, channel, K * L)
return y
def forward_feature(self, y, base_shape):
B, channel, K, L = base_shape
if K == 1:
return y
y = y.reshape(B, channel, K, L).permute(0, 3, 1, 2).reshape(B * L, channel, K)
if self.is_linear:
y = self.feature_layer(y.permute(0, 2, 1)).permute(0, 2, 1)
else:
y = self.feature_layer(y.permute(2, 0, 1)).permute(1, 2, 0)
y = y.reshape(B, L, channel, K).permute(0, 2, 3, 1).reshape(B, channel, K * L)
return y
def forward(self, x, cond_info, diffusion_emb):
B, channel, K, L = x.shape
base_shape = x.shape
x = x.reshape(B, channel, K * L)
diffusion_emb = self.diffusion_projection(diffusion_emb).unsqueeze(-1) # (B,channel,1)
y = x + diffusion_emb
y = self.forward_time(y, base_shape)
y = self.forward_feature(y, base_shape) # (B,channel,K*L)
y = self.mid_projection(y) # (B,2*channel,K*L)
_, cond_dim, _, _ = cond_info.shape
cond_info = cond_info.reshape(B, cond_dim, K * L)
cond_info = self.cond_projection(cond_info) # (B,2*channel,K*L)
y = y + cond_info
gate, filter = torch.chunk(y, 2, dim=1)
y = torch.sigmoid(gate) * torch.tanh(filter) # (B,channel,K*L)
y = self.output_projection(y)
residual, skip = torch.chunk(y, 2, dim=1)
x = x.reshape(base_shape)
residual = residual.reshape(base_shape)
skip = skip.reshape(base_shape)
return (x + residual) / math.sqrt(2.0), skip