GraphPDE/gnn_module.py at main · ruio248/GraphPDE · GitHub

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
from torch_geometric.nn import MessagePassing
import modules
import torch
import utils
from torch import nn

class processor(MessagePassing):
    def __init__(self, in_channels, out_channels,layer_norm=False,nl='relu'):
        super(processor, self).__init__(aggr='add')  # "Add" aggregation.
        self.edge_encoder = modules.FCBlock(in_features=in_channels*3,
                               out_features=out_channels,
                               num_hidden_layers=2,
                               hidden_features=in_channels,
                               outermost_linear=True,
                               nonlinearity=nl,layer_norm=layer_norm)
        self.node_encoder = modules.FCBlock(in_features=in_channels*2,
                               out_features=out_channels,
                               num_hidden_layers=2,
                               hidden_features=in_channels,
                               outermost_linear=True,
                               nonlinearity=nl,layer_norm=layer_norm)
        self.latent_dim = out_channels

    def forward(self, graph):
        edge_index = graph.edge_index
        # cat features together (eij,vi,ei)
        x_receiver = torch.gather(graph.x,0,edge_index[0,:].unsqueeze(-1).repeat(1,graph.x.shape[1]))
        x_sender = torch.gather(graph.x,0,edge_index[1,:].unsqueeze(-1).repeat(1,graph.x.shape[1]))
        edge_features = torch.cat([x_receiver,x_sender,graph.edge_attr],dim=-1)
        # edge processor
        edge_features = self.edge_encoder(edge_features)

        # aggregate edge_features
        node_features = self.propagate(edge_index, x=graph.x, edge_attr=edge_features)
        # cat features for node processor (vi,\sum_eij)
        features = torch.cat([graph.x,node_features[:,self.latent_dim:]],dim=-1)
        # node processor and update graph
        graph.x = self.node_encoder(features) + graph.x
        graph.edge_attr = edge_features
        return graph

    def message(self, x_i, edge_attr):
        z = torch.cat([x_i,edge_attr],dim=-1)
        return z

class mesh_PDE(torch.nn.Module):
    def __init__(self, edge_dim, node_dim,latent_dim=32,num_steps=3,layer_norm=False,nl="relu",var=0,
                 batch_norm=False, normalize = False,encoder_nl='relu',diffMLP=False,checkpoints=None):
        super().__init__()
        self.encoder_edge = modules.FCBlock(in_features=edge_dim,
                               out_features=latent_dim,
                               num_hidden_layers=2,
                               hidden_features=latent_dim,
                               outermost_linear=True,
                               nonlinearity=nl,layer_norm=layer_norm)
        self.encoder_nodes = modules.FCBlock(in_features=node_dim,
                               out_features=latent_dim,
                               num_hidden_layers=2,
                               hidden_features=latent_dim,
                               outermost_linear=True,
                               nonlinearity=nl,layer_norm=layer_norm)


        self.num_steps = num_steps
        self.diffMLP = diffMLP
        if diffMLP:
            # message passing with different MLP for each steps
            self.processors = []
            for _ in range(num_steps):
                self.processors.append((processor(latent_dim, latent_dim,layer_norm=layer_norm)))
                if batch_norm:
                    self.processors.append(torch.nn.BatchNorm1d(latent_dim))
                    self.processors.append(torch.nn.BatchNorm1d(latent_dim))
            self.processors = torch.nn.Sequential(*self.processors)
        else:
            self.processors=((processor(latent_dim, latent_dim,layer_norm=layer_norm)))


        self.decoder_node = modules.FCBlock(in_features=latent_dim,
                                out_features=1,
                                num_hidden_layers=3,
                                hidden_features=latent_dim,
                                outermost_linear=True,
                                nonlinearity=encoder_nl,layer_norm=False)

        self.var = var

        self.batch_norm = batch_norm
        self.normalize = normalize
        if self.normalize:
            if checkpoints==None:
                self.normalizer_node_feature = normalizer(node_dim)
                self.normalizer_edge_feature = normalizer(edge_dim)
                self.normalizer_v_gt = normalizer(1)
            else:
                self.normalizer_node_feature = normalizer(node_dim,max_acc=0)
                self.normalizer_edge_feature = normalizer(edge_dim,max_acc=0)
                self.normalizer_v_gt = normalizer(1,max_acc=0)


    def encoder(self,graph):
        # add noise to input training sample
        noise = (torch.normal(0,1,size=(graph.x.shape[0],graph.x.shape[1]-1))*self.var).cuda()
        graph.x[:,:-1] = graph.x[:,:-1] + graph.x[:,[-1]]*noise
        graph.noise = graph.x[:,[-1]]*noise

        graph.x = self.encoder_nodes(graph.x)
        graph.edge_attr = self.encoder_edge(graph.edge_attr)
        return graph

    def decoder(self,graph):
        graph.x = self.decoder_node(graph.x)
        return graph

    def forward(self, graph, train=False):
        # normalize the dataset
        if self.normalize:
            graph.x = self.normalizer_node_feature.update(graph.x,train)
            graph.edge_attr = self.normalizer_edge_feature.update(graph.edge_attr,train)
            graph.v_gt = self.normalizer_v_gt.update(graph.v_gt,train)
        #ecode edges and nodes to latent dim
        graph_latent = self.encoder(graph.clone())

        if self.diffMLP:
            # message passing steps with different MLP each time
            for i in range(self.num_steps):
                if self.batch_norm:
                    graph_latent = self.processors[i*3](graph_latent)
                    graph_latent.x = self.processors[i*3+1](graph_latent.x)
                    graph_latent.edge_attr = self.processors[i*3+2](graph_latent.edge_attr)
                else:
                    graph_latent = self.processors[i](graph_latent)
        else:
            # message passing steps with same MLP each time
            for _ in range(self.num_steps):
                graph_latent = self.processors(graph_latent)


        # decoding
        graph_latent = self.decoder(graph_latent)
        graph_latent.x = graph_latent.x/10 #div 10 for 46, div 100 for 31
        if self.normalize and not(train):
            graph_latent.x = self.normalizer_v_gt.reverse(graph_latent.x)
        graph_latent.eval = graph_latent.x
        return graph_latent


class normalizer(nn.Module):
    def __init__(self, dim, mean=0, std=1e-8, max_acc = 60*600):
        super().__init__()
        self.acc_sum = nn.Parameter(torch.zeros(dim).cuda(),requires_grad=False)
        self.acc_sum_squared = nn.Parameter(torch.zeros(dim).cuda(),requires_grad=False)
        self.mean = nn.Parameter(torch.zeros(dim).cuda(),requires_grad=False)
        self.std = nn.Parameter(torch.ones(dim).cuda(),requires_grad=False)

        self.total_acc = 0
        self.max_acc = max_acc

    def update(self,value,train):
        if self.total_acc<self.max_acc*value.shape[0] and train:
            self.total_acc += value.shape[0]
            self.acc_sum += torch.sum(value,0).data
            self.acc_sum_squared += torch.sum(value**2,0).data
            safe_count = max(1,self.total_acc)
            self.mean = nn.Parameter(self.acc_sum/safe_count)
            self.std = nn.Parameter(torch.maximum(torch.sqrt(self.acc_sum_squared / safe_count - self.mean**2),torch.tensor(1e-5).cuda()))
        return (value-self.mean.data)/self.std.data

    def reverse(self,value):
        return value*self.std.data+self.mean.data