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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
|
import torch
import torch.nn as nn
import torch.nn.functional as F
"""
3WLGNN / ThreeWLGNN
Provably Powerful Graph Networks (Maron et al., 2019)
https://papers.nips.cc/paper/8488-provably-powerful-graph-networks.pdf
CODE adapted from https://github.com/hadarser/ProvablyPowerfulGraphNetworks_torch/
"""
class ThreeWLGNNNet(nn.Module):
def __init__(self, gnn_type, num_tasks, three_wl_in_dim, n_layers=3, depth_of_mlp=2, hidden_dim=128,
residual=False, **model_kwargs):
assert gnn_type == '3wlgnn'
super(ThreeWLGNNNet, self).__init__()
self.in_dim_node = three_wl_in_dim
self.residual = residual
block_features = [hidden_dim] * n_layers # L here is the block number
original_features_num = self.in_dim_node + 1 # Number of features of the input
# sequential mlp blocks
last_layer_features = original_features_num
self.reg_blocks = nn.ModuleList()
for layer, next_layer_features in enumerate(block_features):
mlp_block = RegularBlock(depth_of_mlp, last_layer_features, next_layer_features, self.residual)
self.reg_blocks.append(mlp_block)
last_layer_features = next_layer_features
self.fc_layers = nn.ModuleList()
for output_features in block_features:
# each block's output will be pooled (thus have 2*output_features), and pass through a fully connected
fc = FullyConnected(2 * output_features, num_tasks, activation_fn=None)
self.fc_layers.append(fc)
def forward(self, x):
scores = torch.tensor(0, device=x.device, dtype=x.dtype)
for i, block in enumerate(self.reg_blocks):
x = block(x)
scores = self.fc_layers[i](diag_offdiag_maxpool(x)) + scores
return scores
# class ThreeWLGNNEdgeNet(nn.Module):
# def __init__(self, gnn_type, num_tasks, feature_dim, n_layers=3, depth_of_mlp=2, hidden_dim=128,
# residual=False, **model_kwargs):
#
# assert gnn_type == '3wlgnn-edge'
# super(ThreeWLGNNEdgeNet, self).__init__()
#
# self.in_dim_node = feature_dim
# self.residual = residual
#
# block_features = [hidden_dim] * n_layers # L here is the block number
# original_features_num = self.in_dim_node + self.num_bond_type + 1 # Number of features of the input
#
# # sequential mlp blocks
# last_layer_features = original_features_num
# self.reg_blocks = nn.ModuleList()
# for layer, next_layer_features in enumerate(block_features):
# mlp_block = RegularBlock(depth_of_mlp, last_layer_features, next_layer_features, self.residual)
# self.reg_blocks.append(mlp_block)
# last_layer_features = next_layer_features
#
# self.fc_layers = nn.ModuleList()
# for output_features in block_features:
# # each block's output will be pooled (thus have 2*output_features), and pass through a fully connected
# fc = FullyConnected(2 * output_features, num_tasks, activation_fn=None)
# self.fc_layers.append(fc)
#
# def forward(self, x_no_edge_feat, x_with_edge_feat):
# x = x_with_edge_feat
#
# scores = torch.tensor(0, device=x.device, dtype=x.dtype)
# for i, block in enumerate(self.reg_blocks):
# x = block(x)
# scores = self.fc_layers[i](diag_offdiag_maxpool(x)) + scores
# return scores
class MLPReadout(nn.Module):
def __init__(self, input_dim, output_dim, L=2): # L=nb_hidden_layers
super().__init__()
list_FC_layers = [nn.Linear(input_dim // 2 ** l, input_dim // 2 ** (l + 1), bias=True) for l in range(L)]
list_FC_layers.append(nn.Linear(input_dim // 2 ** L, output_dim, bias=True))
self.FC_layers = nn.ModuleList(list_FC_layers)
self.L = L
def forward(self, x):
y = x
for l in range(self.L):
y = self.FC_layers[l](y)
y = F.relu(y)
y = self.FC_layers[self.L](y)
return y
class RegularBlock(nn.Module):
"""
Imputs: N x input_depth x m x m
Take the input through 2 parallel MLP routes, multiply the result, and add a skip-connection at the end.
At the skip-connection, reduce the dimension back to output_depth
"""
def __init__(self, depth_of_mlp, in_features, out_features, residual=False):
super().__init__()
self.residual = residual
self.mlp1 = MlpBlock(in_features, out_features, depth_of_mlp)
self.mlp2 = MlpBlock(in_features, out_features, depth_of_mlp)
self.skip = SkipConnection(in_features + out_features, out_features)
if self.residual:
self.res_x = nn.Linear(in_features, out_features)
def forward(self, inputs):
mlp1 = self.mlp1(inputs)
mlp2 = self.mlp2(inputs)
mult = torch.matmul(mlp1, mlp2)
out = self.skip(in1=inputs, in2=mult)
if self.residual:
# Now, changing shapes from [1xdxnxn] to [nxnxd] for Linear() layer
inputs, out = inputs.permute(3, 2, 1, 0).squeeze(), out.permute(3, 2, 1, 0).squeeze()
residual_ = self.res_x(inputs)
out = residual_ + out # residual connection
# Returning output back to original shape
out = out.permute(2, 1, 0).unsqueeze(0)
return out
class MlpBlock(nn.Module):
"""
Block of MLP layers with activation function after each (1x1 conv layers).
"""
def __init__(self, in_features, out_features, depth_of_mlp, activation_fn=nn.functional.relu):
super().__init__()
self.activation = activation_fn
self.convs = nn.ModuleList()
for i in range(depth_of_mlp):
self.convs.append(nn.Conv2d(in_features, out_features, kernel_size=1, padding=0, bias=True))
_init_weights(self.convs[-1])
in_features = out_features
def forward(self, inputs):
out = inputs
for conv_layer in self.convs:
out = self.activation(conv_layer(out))
return out
class SkipConnection(nn.Module):
"""
Connects the two given inputs with concatenation
:param in1: earlier input tensor of shape N x d1 x m x m
:param in2: later input tensor of shape N x d2 x m x m
:param in_features: d1+d2
:param out_features: output num of features
:return: Tensor of shape N x output_depth x m x m
"""
def __init__(self, in_features, out_features):
super().__init__()
self.conv = nn.Conv2d(in_features, out_features, kernel_size=1, padding=0, bias=True)
_init_weights(self.conv)
def forward(self, in1, in2):
# in1: N x d1 x m x m
# in2: N x d2 x m x m
out = torch.cat((in1, in2), dim=1)
out = self.conv(out)
return out
class FullyConnected(nn.Module):
def __init__(self, in_features, out_features, activation_fn=nn.functional.relu):
super().__init__()
self.fc = nn.Linear(in_features, out_features)
_init_weights(self.fc)
self.activation = activation_fn
def forward(self, input):
out = self.fc(input)
if self.activation is not None:
out = self.activation(out)
return out
def diag_offdiag_maxpool(input):
N = input.shape[-1]
max_diag = torch.max(torch.diagonal(input, dim1=-2, dim2=-1), dim=2)[0] # BxS
# with torch.no_grad():
max_val = torch.max(max_diag)
min_val = torch.max(-1 * input)
val = torch.abs(torch.add(max_val, min_val))
min_mat = torch.mul(val, torch.eye(N, device=input.device)).view(1, 1, N, N)
max_offdiag = torch.max(torch.max(input - min_mat, dim=3)[0], dim=2)[0] # BxS
return torch.cat((max_diag, max_offdiag), dim=1) # output Bx2S
def _init_weights(layer):
"""
Init weights of the layer
:param layer:
:return:
"""
nn.init.xavier_uniform_(layer.weight)
# nn.init.xavier_normal_(layer.weight)
if layer.bias is not None:
nn.init.zeros_(layer.bias)
class LayerNorm(nn.Module):
def __init__(self, d):
super().__init__()
self.a = nn.Parameter(torch.ones(d).unsqueeze(0).unsqueeze(0)) # shape is 1 x 1 x d
self.b = nn.Parameter(torch.zeros(d).unsqueeze(0).unsqueeze(0)) # shape is 1 x 1 x d
def forward(self, x):
# x tensor of the shape n x n x d
mean = x.mean(dim=(0, 1), keepdim=True)
var = x.var(dim=(0, 1), keepdim=True, unbiased=False)
x = self.a * (x - mean) / torch.sqrt(var + 1e-6) + self.b # shape is n x n x d
return x
|