-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathtransModel.py
More file actions
153 lines (123 loc) · 5.58 KB
/
transModel.py
File metadata and controls
153 lines (123 loc) · 5.58 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
141
142
143
144
145
146
147
148
149
150
151
152
153
from torch_geometric.nn import TransformerConv, LayerNorm, GATConv, GCNConv
import torch.nn.functional as F
import torch
import math
class TranOne(torch.nn.Module):
def __init__(self, hidden_dims, use_component='gene'):
super().__init__()
self.use_component = use_component
[in_dim, img_dim, num_hidden, out_dim] = hidden_dims
self.conv1 = TransformerConv(in_dim, num_hidden)
self.conv2 = TransformerConv(num_hidden, out_dim)
self.conv3 = TransformerConv(out_dim, num_hidden)
self.conv4 = TransformerConv(num_hidden, in_dim)
self.imgconv1 = TransformerConv(img_dim, num_hidden)
self.imgconv2 = TransformerConv(num_hidden, out_dim)
self.imgconv3 = TransformerConv(out_dim, num_hidden)
self.imgconv4 = TransformerConv(num_hidden, img_dim)
# layernorm
self.norm1 = LayerNorm(num_hidden)
self.norm2 = LayerNorm(out_dim)
# relu
self.activate = F.elu
def forward(self, feat, edge):
if self.use_component == 'gene':
h1 = self.activate(self.conv1(feat, edge))
h2 = self.conv2(h1, edge)
h3 = self.activate(self.conv3(h2, edge))
h4 = self.conv4(h3, edge)
else:
h1 = self.activate(self.imgconv1(feat, edge))
h2 = self.imgconv2(feat, edge)
h3 = self.activate(self.imgconv3(feat, edge))
h4 = self.imgconv4(feat, edge)
return h2, h4
class DataContrast(torch.nn.Module):
def __init__(self, hidden_dims, ncluster, nspots):
super().__init__()
[in_dim, img_dim, num_hidden, out_dim] = hidden_dims
self.conv1 = GCNConv(in_dim, 2048)
self.conv2 = GCNConv(2048, 4096)
self.emb = GCNConv(4096, out_dim)
self.conv3 = GCNConv(out_dim, num_hidden)
self.conv4 = GCNConv(num_hidden, in_dim)
mask = torch.Tensor(nspots, in_dim)
torch.nn.init.uniform_(mask)
mask = (mask > 0.5).float()
self.mask = torch.nn.Parameter(mask)
self.proj = GCNConv(4096, ncluster)
self.activate = F.elu
self.softmax = torch.nn.Softmax(dim=1)
def forward(self, xi, edge_index):
# print(xi.shape)
hi1 = self.activate(self.conv1(xi, edge_index))
hi2 = self.activate(self.conv2(hi1, edge_index))
emb = self.activate(self.emb(hi2, edge_index))
# combine1 = torch.concat([emb, hi2], dim=1)
up1 = self.activate(self.conv3(emb, edge_index))
# combine2 = torch.concat([up1, hi1], dim=1)
up2 = self.conv4(up1, edge_index)
ci = self.softmax(self.proj(hi2, edge_index))
# print(xi.shape, self.mask.shape)
xj = xi * self.mask
hj1 = self.activate(self.conv1(xj, edge_index))
hj2 = self.activate(self.conv2(hj1, edge_index))
cj = self.softmax(self.proj(hj2, edge_index))
return hi2, hj2, ci, cj, up2
class ImgContrast(torch.nn.Module):
def __init__(self, hidden_dims, ncluster):
super().__init__()
[in_dim, img_dim, num_hidden, out_dim] = hidden_dims
self.imgconv1 = TransformerConv(img_dim, num_hidden)
self.imgconv2 = TransformerConv(num_hidden, out_dim)
self.proj = TransformerConv(out_dim, ncluster)
self.activate = F.elu
def forward(self, xi, xj, edge_index):
hi1 = self.activate(self.imgconv1(xi, edge_index))
hi2 = self.imgconv2(hi1, edge_index)
ci = self.proj(hi1, edge_index)
hj1 = self.activate(self.imgconv1(xj, edge_index))
hj2 = self.imgconv(hj1, edge_index)
cj = self.proj(hj2, edge_index)
return hi2, hj2, ci, cj
class TransImg(torch.nn.Module):
def __init__(self, hidden_dims, use_img_loss=False):
super().__init__()
[in_dim, img_dim, num_hidden, out_dim] = hidden_dims
# [in_dim, emb_dim, img_dim, num_hidden, out_dim] = hidden_dims
self.conv1 = TransformerConv(in_dim, num_hidden)
self.conv2 = TransformerConv(num_hidden, out_dim)
self.conv3 = TransformerConv(out_dim, num_hidden)
self.conv4 = TransformerConv(num_hidden, in_dim)
self.imgconv1 = TransformerConv(img_dim, num_hidden)
self.imgconv2 = TransformerConv(num_hidden, out_dim)
self.imgconv3 = TransformerConv(out_dim, num_hidden)
if use_img_loss:
self.imgconv4 = TransformerConv(num_hidden, img_dim)
else:
self.imgconv4 = TransformerConv(num_hidden, in_dim)
self.neck = TransformerConv(out_dim * 2, out_dim)
self.neck2 = TransformerConv(out_dim, out_dim)
self.c3 = TransformerConv(out_dim, num_hidden)
self.c4 = TransformerConv(num_hidden, in_dim)
# layernorm
self.norm1 = LayerNorm(num_hidden)
self.norm2 = LayerNorm(out_dim)
# relu
self.activate = F.elu
def forward(self, features, img_feat, edge_index):
h1 = self.activate(self.conv1(features, edge_index))
h2 = self.conv2(h1, edge_index)
h3 = self.activate(self.conv3(h2, edge_index))
h4 = self.conv4(h3, edge_index)
img1 = self.activate(self.imgconv1(img_feat, edge_index))
img2 = self.imgconv2(img1, edge_index)
img3 = self.activate(self.imgconv3(img2, edge_index))
img4 = self.imgconv4(img3, edge_index)
concat = torch.cat([h2, img2], dim=1)
combine = self.activate(self.neck(concat, edge_index))
c2 = self.neck2(combine, edge_index)
c3 = self.activate(self.c3(c2, edge_index))
c4 = self.c4(c3, edge_index)
# print(h4.shape, img4.shape, c4.shape)
return h2, img2, c2, h4, img4, c4