Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,361 Bytes
7968cb0 |
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 |
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
# from torch_scatter import scatter_sum, scatter_softmax, scatter_mean
from torch_geometric.utils import scatter
#################################### node modules ###############################
class NeighborAttention(nn.Module):
def __init__(self, num_hidden, num_in, num_heads=4, edge_drop=0.0, output_mlp=True):
super(NeighborAttention, self).__init__()
self.num_heads = num_heads
self.num_hidden = num_hidden
self.edge_drop = edge_drop
self.output_mlp = output_mlp
self.W_V = nn.Sequential(nn.Linear(num_in, num_hidden),
nn.GELU(),
nn.Linear(num_hidden, num_hidden),
nn.GELU(),
nn.Linear(num_hidden, num_hidden)
)
self.Bias = nn.Sequential(
nn.Linear(num_hidden*3, num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_heads)
)
self.W_O = nn.Linear(num_hidden, num_hidden, bias=False)
def forward(self, h_V, h_E, center_id, batch_id, dst_idx=None):
N = h_V.shape[0]
E = h_E.shape[0]
n_heads = self.num_heads
d = int(self.num_hidden / n_heads)
w = self.Bias(torch.cat([h_V[center_id], h_E],dim=-1)).view(E, n_heads, 1)
attend_logits = w/np.sqrt(d)
V = self.W_V(h_E).view(-1, n_heads, d)
attend = scatter.scatter_softmax(attend_logits, index=center_id, dim=0)
h_V = scatter.scatter_sum(attend*V, center_id, dim=0).view([-1, self.num_hidden])
if self.output_mlp:
h_V_update = self.W_O(h_V)
else:
h_V_update = h_V
return h_V_update
#################################### edge modules ###############################
class EdgeMLP(nn.Module):
def __init__(self, num_hidden, num_in, dropout=0.1, num_heads=None, scale=30):
super(EdgeMLP, self).__init__()
self.num_hidden = num_hidden
self.num_in = num_in
self.scale = scale
self.dropout = nn.Dropout(dropout)
self.norm = nn.BatchNorm1d(num_hidden)
self.W11 = nn.Linear(num_hidden + num_in, num_hidden, bias=True)
self.W12 = nn.Linear(num_hidden, num_hidden, bias=True)
self.W13 = nn.Linear(num_hidden, num_hidden, bias=True)
self.act = torch.nn.GELU()
def forward(self, h_V, h_E, edge_idx, batch_id):
src_idx = edge_idx[0]
dst_idx = edge_idx[1]
h_EV = torch.cat([h_V[src_idx], h_E, h_V[dst_idx]], dim=-1)
h_message = self.W13(self.act(self.W12(self.act(self.W11(h_EV)))))
h_E = self.norm(h_E + self.dropout(h_message))
return h_E
#################################### context modules ###############################
class Context(nn.Module):
def __init__(self, num_hidden, num_in, dropout=0.1, num_heads=None, scale=30, node_context = False, edge_context = False):
super(Context, self).__init__()
self.num_hidden = num_hidden
self.num_in = num_in
self.scale = scale
self.node_context = node_context
self.edge_context = edge_context
self.V_MLP = nn.Sequential(
nn.Linear(num_hidden, num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden),
)
self.V_MLP_g = nn.Sequential(
nn.Linear(num_hidden, num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden),
nn.Sigmoid()
)
self.E_MLP = nn.Sequential(
nn.Linear(num_hidden, num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden)
)
self.E_MLP_g = nn.Sequential(
nn.Linear(num_hidden, num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden),
nn.ReLU(),
nn.Linear(num_hidden,num_hidden),
nn.Sigmoid()
)
def forward(self, h_V, h_E, edge_idx, batch_id):
if self.node_context:
c_V = scatter.scatter_mean(h_V, batch_id, dim=0)
h_V = h_V * self.V_MLP_g(c_V[batch_id])
if self.edge_context:
c_V = scatter.scatter_mean(h_V, batch_id, dim=0)
h_E = h_E * self.E_MLP_g(c_V[batch_id[edge_idx[0]]])
return h_V, h_E
class GeneralGNN(nn.Module):
def __init__(self, num_hidden, num_in, dropout=0.1, num_heads=None, scale=30, node_context = 0, edge_context = 0):
super(GeneralGNN, self).__init__()
self.num_hidden = num_hidden
self.num_in = num_in
self.scale = scale
self.dropout = nn.Dropout(dropout)
self.norm = nn.ModuleList([nn.BatchNorm1d(num_hidden) for _ in range(3)])
self.attention = NeighborAttention(num_hidden, num_in, num_heads=4)
self.edge_update = EdgeMLP(num_hidden, num_in, num_heads=4)
self.context = Context(num_hidden, num_in, num_heads=4, node_context=node_context, edge_context=edge_context)
self.dense = nn.Sequential(
nn.Linear(num_hidden, num_hidden*4),
nn.ReLU(),
nn.Linear(num_hidden*4, num_hidden)
)
self.W11 = nn.Linear(num_hidden + num_in, num_hidden, bias=True)
self.W12 = nn.Linear(num_hidden, num_hidden, bias=True)
self.W13 = nn.Linear(num_hidden, num_hidden, bias=True)
self.act = torch.nn.GELU()
def forward(self, h_V, h_E, edge_idx, batch_id):
src_idx = edge_idx[0]
dst_idx = edge_idx[1]
dh = self.attention(h_V, torch.cat([h_E, h_V[dst_idx]], dim=-1), src_idx, batch_id, dst_idx)
h_V = self.norm[0](h_V + self.dropout(dh))
dh = self.dense(h_V)
h_V = self.norm[1](h_V + self.dropout(dh))
h_E = self.edge_update( h_V, h_E, edge_idx, batch_id )
h_V, h_E = self.context(h_V, h_E, edge_idx, batch_id)
return h_V, h_E
class StructureEncoder(nn.Module):
def __init__(self, hidden_dim, num_encoder_layers=3, dropout=0, updating_edges=0, node_context = False, edge_context = False):
super(StructureEncoder, self).__init__()
encoder_layers = []
self.updating_edges = updating_edges
module = GeneralGNN
for i in range(num_encoder_layers):
encoder_layers.append(
module(hidden_dim, hidden_dim*2, dropout=dropout, node_context = node_context, edge_context = edge_context),
)
self.encoder_layers = nn.Sequential(*encoder_layers)
def forward(self, h_V, h_P, P_idx, batch_id):
for layer in self.encoder_layers:
if self.updating_edges == 0:
h_V = layer(h_V, torch.cat([h_P, h_V[P_idx[1]]], dim=1), P_idx, batch_id)
else:
h_V, h_P = layer(h_V, h_P, P_idx, batch_id)
return h_V, h_P
class MLPDecoder(nn.Module):
def __init__(self, hidden_dim, vocab=20):
super().__init__()
self.readout = nn.Linear(hidden_dim, vocab)
def forward(self, h_V):
logits = self.readout(h_V)
log_probs = F.log_softmax(logits, dim=-1)
return log_probs, logits |