-
Notifications
You must be signed in to change notification settings - Fork 23
Expand file tree
/
Copy pathdiff_models.py
More file actions
141 lines (117 loc) · 6.29 KB
/
diff_models.py
File metadata and controls
141 lines (117 loc) · 6.29 KB
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
from layers import *
class Guide_diff(nn.Module):
def __init__(self, config, inputdim=1, target_dim=36, is_itp=False):
super().__init__()
self.channels = config["channels"]
self.is_itp = is_itp
self.itp_channels = None
if self.is_itp:
self.itp_channels = config["channels"]
self.itp_projection = Conv1d_with_init(inputdim-1, self.itp_channels, 1)
self.itp_modeling = GuidanceConstruct(channels=self.itp_channels, nheads=config["nheads"], target_dim=target_dim,
order=2, include_self=True, device=config["device"], is_adp=config["is_adp"],
adj_file=config["adj_file"], proj_t=config["proj_t"])
self.cond_projection = Conv1d_with_init(config["side_dim"], self.itp_channels, 1)
self.itp_projection2 = Conv1d_with_init(self.itp_channels, 1, 1)
self.diffusion_embedding = DiffusionEmbedding(
num_steps=config["num_steps"],
embedding_dim=config["diffusion_embedding_dim"],
)
if config["adj_file"] == 'AQI36':
self.adj = get_adj_AQI36()
elif config["adj_file"] == 'metr-la':
self.adj = get_similarity_metrla(thr=0.1)
elif config["adj_file"] == 'pems-bay':
self.adj = get_similarity_pemsbay(thr=0.1)
self.device = config["device"]
self.support = compute_support_gwn(self.adj, device=config["device"])
self.is_adp = config["is_adp"]
if self.is_adp:
node_num = self.adj.shape[0]
self.nodevec1 = nn.Parameter(torch.randn(node_num, 10).to(self.device), requires_grad=True).to(self.device)
self.nodevec2 = nn.Parameter(torch.randn(10, node_num).to(self.device), requires_grad=True).to(self.device)
self.support.append([self.nodevec1, self.nodevec2])
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(
[
NoiseProject(
side_dim=config["side_dim"],
channels=self.channels,
diffusion_embedding_dim=config["diffusion_embedding_dim"],
nheads=config["nheads"],
target_dim=target_dim,
proj_t=config["proj_t"],
is_adp=config["is_adp"],
device=config["device"],
adj_file=config["adj_file"],
is_cross_t=config["is_cross_t"],
is_cross_s=config["is_cross_s"],
)
for _ in range(config["layers"])
]
)
def forward(self, x, side_info, diffusion_step, itp_x, cond_mask):
if self.is_itp:
x = torch.cat([x, itp_x], dim=1)
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)
if self.is_itp:
itp_x = itp_x.reshape(B, inputdim-1, K * L)
itp_x = self.itp_projection(itp_x)
itp_cond_info = side_info.reshape(B, -1, K * L)
itp_cond_info = self.cond_projection(itp_cond_info)
itp_x = itp_x + itp_cond_info
itp_x = self.itp_modeling(itp_x, [B, self.itp_channels, K, L], self.support)
itp_x = F.relu(itp_x)
itp_x = itp_x.reshape(B, self.itp_channels, K, L)
diffusion_emb = self.diffusion_embedding(diffusion_step)
skip = []
for i in range(len(self.residual_layers)):
x, skip_connection = self.residual_layers[i](x, side_info, diffusion_emb, itp_x, self.support)
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 NoiseProject(nn.Module):
def __init__(self, side_dim, channels, diffusion_embedding_dim, nheads, target_dim, proj_t, order=2, include_self=True,
device=None, is_adp=False, adj_file=None, is_cross_t=False, is_cross_s=True):
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.forward_time = TemporalLearning(channels=channels, nheads=nheads, is_cross=is_cross_t)
self.forward_feature = SpatialLearning(channels=channels, nheads=nheads, target_dim=target_dim,
order=order, include_self=include_self, device=device, is_adp=is_adp,
adj_file=adj_file, proj_t=proj_t, is_cross=is_cross_s)
def forward(self, x, side_info, diffusion_emb, itp_info, support):
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, itp_info)
y = self.forward_feature(y, base_shape, support, itp_info) # (B,channel,K*L)
y = self.mid_projection(y) # (B,2*channel,K*L)
_, side_dim, _, _ = side_info.shape
side_info = side_info.reshape(B, side_dim, K * L)
side_info = self.cond_projection(side_info) # (B,2*channel,K*L)
y = y + side_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