Spaces:
Sleeping
Sleeping
File size: 8,481 Bytes
e75a247 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import dgl.function as fn
"""
ResGatedGCN: Residual Gated Graph ConvNets
An Experimental Study of Neural Networks for Variable Graphs (Xavier Bresson and Thomas Laurent, ICLR 2018)
https://arxiv.org/pdf/1711.07553v2.pdf
"""
class GatedGCNLayer(nn.Module):
"""
Param: []
"""
def __init__(self, input_dim, output_dim, dropout, batch_norm, residual=False):
super().__init__()
self.in_channels = input_dim
self.out_channels = output_dim
self.dropout = dropout
self.batch_norm = batch_norm
self.residual = residual
if input_dim != output_dim:
self.residual = False
self.A = nn.Linear(input_dim, output_dim, bias=True)
self.B = nn.Linear(input_dim, output_dim, bias=True)
self.C = nn.Linear(input_dim, output_dim, bias=True)
self.D = nn.Linear(input_dim, output_dim, bias=True)
self.E = nn.Linear(input_dim, output_dim, bias=True)
self.bn_node_h = nn.BatchNorm1d(output_dim)
self.bn_node_e = nn.BatchNorm1d(output_dim)
def message_func(self, edges):
Bh_j = edges.src['Bh']
e_ij = edges.data['Ce'] + edges.src['Dh'] + edges.dst['Eh'] # e_ij = Ce_ij + Dhi + Ehj
edges.data['e'] = e_ij
return {'Bh_j' : Bh_j, 'e_ij' : e_ij}
def reduce_func(self, nodes):
Ah_i = nodes.data['Ah']
Bh_j = nodes.mailbox['Bh_j']
e = nodes.mailbox['e_ij']
sigma_ij = torch.sigmoid(e) # sigma_ij = sigmoid(e_ij)
#h = Ah_i + torch.mean( sigma_ij * Bh_j, dim=1 ) # hi = Ahi + mean_j alpha_ij * Bhj
h = Ah_i + torch.sum( sigma_ij * Bh_j, dim=1 ) / ( torch.sum( sigma_ij, dim=1 ) + 1e-6 ) # hi = Ahi + sum_j eta_ij/sum_j' eta_ij' * Bhj <= dense attention
return {'h' : h}
def forward(self, g, h, e):
h_in = h # for residual connection
e_in = e # for residual connection
g.ndata['h'] = h
g.ndata['Ah'] = self.A(h)
g.ndata['Bh'] = self.B(h)
g.ndata['Dh'] = self.D(h)
g.ndata['Eh'] = self.E(h)
g.edata['e'] = e
g.edata['Ce'] = self.C(e)
g.apply_edges(fn.u_add_v('Dh', 'Eh', 'DEh'))
g.edata['e'] = g.edata['DEh'] + g.edata['Ce']
g.edata['sigma'] = torch.sigmoid(g.edata['e'])
g.update_all(fn.u_mul_e('Bh', 'sigma', 'm'), fn.sum('m', 'sum_sigma_h'))
g.update_all(fn.copy_e('sigma', 'm'), fn.sum('m', 'sum_sigma'))
g.ndata['h'] = g.ndata['Ah'] + g.ndata['sum_sigma_h'] / (g.ndata['sum_sigma'] + 1e-6)
#g.update_all(self.message_func,self.reduce_func)
h = g.ndata['h'] # result of graph convolution
e = g.edata['e'] # result of graph convolution
if self.batch_norm:
h = self.bn_node_h(h) # batch normalization
e = self.bn_node_e(e) # batch normalization
h = F.relu(h) # non-linear activation
e = F.relu(e) # non-linear activation
if self.residual:
h = h_in + h # residual connection
e = e_in + e # residual connection
h = F.dropout(h, self.dropout, training=self.training)
e = F.dropout(e, self.dropout, training=self.training)
return h, e
def __repr__(self):
return '{}(in_channels={}, out_channels={})'.format(self.__class__.__name__,
self.in_channels,
self.out_channels)
##############################################################
#
# Additional layers for edge feature/representation analysis
#
##############################################################
class GatedGCNLayerEdgeFeatOnly(nn.Module):
"""
Param: []
"""
def __init__(self, input_dim, output_dim, dropout, batch_norm, residual=False):
super().__init__()
self.in_channels = input_dim
self.out_channels = output_dim
self.dropout = dropout
self.batch_norm = batch_norm
self.residual = residual
if input_dim != output_dim:
self.residual = False
self.A = nn.Linear(input_dim, output_dim, bias=True)
self.B = nn.Linear(input_dim, output_dim, bias=True)
self.D = nn.Linear(input_dim, output_dim, bias=True)
self.E = nn.Linear(input_dim, output_dim, bias=True)
self.bn_node_h = nn.BatchNorm1d(output_dim)
def message_func(self, edges):
Bh_j = edges.src['Bh']
e_ij = edges.src['Dh'] + edges.dst['Eh'] # e_ij = Dhi + Ehj
edges.data['e'] = e_ij
return {'Bh_j' : Bh_j, 'e_ij' : e_ij}
def reduce_func(self, nodes):
Ah_i = nodes.data['Ah']
Bh_j = nodes.mailbox['Bh_j']
e = nodes.mailbox['e_ij']
sigma_ij = torch.sigmoid(e) # sigma_ij = sigmoid(e_ij)
h = Ah_i + torch.sum( sigma_ij * Bh_j, dim=1 ) / ( torch.sum( sigma_ij, dim=1 ) + 1e-6 ) # hi = Ahi + sum_j eta_ij/sum_j' eta_ij' * Bhj <= dense attention
return {'h' : h}
def forward(self, g, h, e):
h_in = h # for residual connection
g.ndata['h'] = h
g.ndata['Ah'] = self.A(h)
g.ndata['Bh'] = self.B(h)
g.ndata['Dh'] = self.D(h)
g.ndata['Eh'] = self.E(h)
#g.update_all(self.message_func,self.reduce_func)
g.apply_edges(fn.u_add_v('Dh', 'Eh', 'e'))
g.edata['sigma'] = torch.sigmoid(g.edata['e'])
g.update_all(fn.u_mul_e('Bh', 'sigma', 'm'), fn.sum('m', 'sum_sigma_h'))
g.update_all(fn.copy_e('sigma', 'm'), fn.sum('m', 'sum_sigma'))
g.ndata['h'] = g.ndata['Ah'] + g.ndata['sum_sigma_h'] / (g.ndata['sum_sigma'] + 1e-6)
h = g.ndata['h'] # result of graph convolution
if self.batch_norm:
h = self.bn_node_h(h) # batch normalization
h = F.relu(h) # non-linear activation
if self.residual:
h = h_in + h # residual connection
h = F.dropout(h, self.dropout, training=self.training)
return h, e
def __repr__(self):
return '{}(in_channels={}, out_channels={})'.format(self.__class__.__name__,
self.in_channels,
self.out_channels)
##############################################################
class GatedGCNLayerIsotropic(nn.Module):
"""
Param: []
"""
def __init__(self, input_dim, output_dim, dropout, batch_norm, residual=False):
super().__init__()
self.in_channels = input_dim
self.out_channels = output_dim
self.dropout = dropout
self.batch_norm = batch_norm
self.residual = residual
if input_dim != output_dim:
self.residual = False
self.A = nn.Linear(input_dim, output_dim, bias=True)
self.B = nn.Linear(input_dim, output_dim, bias=True)
self.bn_node_h = nn.BatchNorm1d(output_dim)
def message_func(self, edges):
Bh_j = edges.src['Bh']
return {'Bh_j' : Bh_j}
def reduce_func(self, nodes):
Ah_i = nodes.data['Ah']
Bh_j = nodes.mailbox['Bh_j']
h = Ah_i + torch.sum( Bh_j, dim=1 ) # hi = Ahi + sum_j Bhj
return {'h' : h}
def forward(self, g, h, e):
h_in = h # for residual connection
g.ndata['h'] = h
g.ndata['Ah'] = self.A(h)
g.ndata['Bh'] = self.B(h)
#g.update_all(self.message_func,self.reduce_func)
g.update_all(fn.copy_u('Bh', 'm'), fn.sum('m', 'sum_h'))
g.ndata['h'] = g.ndata['Ah'] + g.ndata['sum_h']
h = g.ndata['h'] # result of graph convolution
if self.batch_norm:
h = self.bn_node_h(h) # batch normalization
h = F.relu(h) # non-linear activation
if self.residual:
h = h_in + h # residual connection
h = F.dropout(h, self.dropout, training=self.training)
return h, e
def __repr__(self):
return '{}(in_channels={}, out_channels={})'.format(self.__class__.__name__,
self.in_channels,
self.out_channels) |