99 lines
4.8 KiB
Python
99 lines
4.8 KiB
Python
|
#coding=utf8
|
||
|
|
import copy, math
|
||
|
|
import torch, dgl
|
||
|
|
import torch.nn as nn
|
||
|
|
import dgl.function as fn
|
||
|
|
from model.model_utils import Registrable, FFN
|
||
|
|
from model.encoder.rgatsql import RGATLayer, MultiViewRGATLayer
|
||
|
|
from model.encoder.functions import *
|
||
|
|
|
||
|
|
@Registrable.register('lgesql')
|
||
|
|
class LGESQL(nn.Module):
|
||
|
|
""" Compared with RGAT, we utilize a line graph to explicitly model the propagation among edges:
|
||
|
|
1. aggregate info from in-edges
|
||
|
|
2. aggregate info from src nodes
|
||
|
|
"""
|
||
|
|
def __init__(self, args):
|
||
|
|
super(LGESQL, self).__init__()
|
||
|
|
self.num_layers, self.num_heads = args.gnn_num_layers, args.num_heads
|
||
|
|
self.relation_share_heads = args.relation_share_heads
|
||
|
|
self.graph_view = args.local_and_nonlocal
|
||
|
|
self.ndim = args.gnn_hidden_size # node feature dim
|
||
|
|
self.edim = self.ndim // self.num_heads if self.relation_share_heads else \
|
||
|
|
self.ndim // 2 if self.graph_view == 'mmc' else self.ndim
|
||
|
|
self.relation_num = args.relation_num
|
||
|
|
self.relation_embed = nn.Embedding(self.relation_num, self.edim)
|
||
|
|
self.gnn_layers = nn.ModuleList([
|
||
|
|
DualRGATLayer(self.ndim, self.edim, num_heads=args.num_heads, feat_drop=args.dropout, graph_view=self.graph_view)
|
||
|
|
for _ in range(self.num_layers)])
|
||
|
|
|
||
|
|
def forward(self, x, batch):
|
||
|
|
global_lgx = self.relation_embed(batch.graph.global_edges)
|
||
|
|
mask = batch.graph.local_mask
|
||
|
|
local_lgx = global_lgx[mask]
|
||
|
|
local_g, global_g, lg = batch.graph.local_g, batch.graph.global_g, batch.graph.lg
|
||
|
|
src_ids, dst_ids = batch.graph.src_ids, batch.graph.dst_ids
|
||
|
|
for i in range(self.num_layers):
|
||
|
|
x, local_lgx = self.gnn_layers[i](x, local_lgx, global_lgx, local_g, global_g, lg, src_ids, dst_ids)
|
||
|
|
if self.graph_view == 'msde':
|
||
|
|
# update local edge embeddings in the global edge embeddings matrix
|
||
|
|
global_lgx = global_lgx.masked_scatter_(mask.unsqueeze(-1), local_lgx)
|
||
|
|
return x
|
||
|
|
|
||
|
|
class DualRGATLayer(nn.Module):
|
||
|
|
|
||
|
|
def __init__(self, ndim, edim, num_heads=8, feat_drop=0.2, graph_view='mmc'):
|
||
|
|
super(DualRGATLayer, self).__init__()
|
||
|
|
self.ndim, self.edim = ndim, edim
|
||
|
|
self.num_heads, self.graph_view = num_heads, graph_view
|
||
|
|
NodeRGATLayer = MultiViewRGATLayer if self.graph_view == 'mmc' else RGATLayer
|
||
|
|
self.node_update = NodeRGATLayer(self.ndim, self.edim, self.num_heads, feat_drop=feat_drop)
|
||
|
|
self.edge_update = EdgeRGATLayer(self.edim, self.ndim, self.num_heads, feat_drop=feat_drop)
|
||
|
|
|
||
|
|
def forward(self, x, local_lgx, global_lgx, local_g, global_g, lg, src_ids, dst_ids):
|
||
|
|
if self.graph_view == 'mmc':
|
||
|
|
out_x, _ = self.node_update(x, local_lgx, global_lgx, local_g, global_g)
|
||
|
|
elif self.graph_view == 'msde':
|
||
|
|
out_x, _ = self.node_update(x, global_lgx, global_g)
|
||
|
|
else:
|
||
|
|
out_x, _ = self.node_update(x, local_lgx, local_g)
|
||
|
|
src_x = torch.index_select(x, dim=0, index=src_ids)
|
||
|
|
dst_x = torch.index_select(x, dim=0, index=dst_ids)
|
||
|
|
out_local_lgx, _ = self.edge_update(local_lgx, src_x, dst_x, lg)
|
||
|
|
return out_x, out_local_lgx
|
||
|
|
|
||
|
|
class EdgeRGATLayer(nn.Module):
|
||
|
|
|
||
|
|
def __init__(self, edim, ndim, num_heads=8, feat_drop=0.2):
|
||
|
|
super(EdgeRGATLayer, self).__init__()
|
||
|
|
self.edim, self.ndim = edim, ndim
|
||
|
|
self.num_heads = num_heads
|
||
|
|
self.d_k = self.ndim // self.num_heads
|
||
|
|
self.affine_q, self.affine_k, self.affine_v = nn.Linear(self.edim, self.ndim), \
|
||
|
|
nn.Linear(self.edim, self.ndim, bias=False), nn.Linear(self.edim, self.ndim, bias=False)
|
||
|
|
self.affine_o = nn.Linear(self.ndim, self.edim)
|
||
|
|
self.layernorm = nn.LayerNorm(self.edim)
|
||
|
|
self.feat_dropout = nn.Dropout(p=feat_drop)
|
||
|
|
self.ffn = FFN(self.edim)
|
||
|
|
|
||
|
|
def forward(self, x, src_x, dst_x, g):
|
||
|
|
q, k, v = self.affine_q(self.feat_dropout(x)), self.affine_k(self.feat_dropout(x)), self.affine_v(self.feat_dropout(x))
|
||
|
|
with g.local_scope():
|
||
|
|
g.ndata['q'] = (q + src_x).view(-1, self.num_heads, self.d_k)
|
||
|
|
g.ndata['k'] = k.view(-1, self.num_heads, self.d_k)
|
||
|
|
g.ndata['v'] = (v + dst_x).view(-1, self.num_heads, self.d_k)
|
||
|
|
out_x = self.propagate_attention(g)
|
||
|
|
out_x = self.layernorm(x + self.affine_o(out_x.view(-1, self.num_heads * self.d_k)))
|
||
|
|
out_x = self.ffn(out_x)
|
||
|
|
return out_x, (src_x, dst_x)
|
||
|
|
|
||
|
|
def propagate_attention(self, g):
|
||
|
|
# Compute attention score
|
||
|
|
g.apply_edges(src_dot_dst('k', 'q', 'score'))
|
||
|
|
g.apply_edges(scaled_exp('score', math.sqrt(self.d_k)))
|
||
|
|
# Update node state
|
||
|
|
g.update_all(fn.src_mul_edge('v', 'score', 'v'), fn.sum('v', 'wv'))
|
||
|
|
g.update_all(fn.copy_edge('score', 'score'), fn.sum('score', 'z'), div_by_z('wv', 'z', 'o'))
|
||
|
|
out_x = g.ndata['o']
|
||
|
|
return out_x
|