File size: 3,409 Bytes
719d0db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
import torch
import torch.nn as nn

class AttentionEdgeEncoder(nn.Module):
    def __init__(self, state_dim, emb_dim, dropout):
        super().__init__()
        self.state_dim = state_dim
        self.emb_dim   = emb_dim
        self.norm_factor = 1 / math.sqrt(emb_dim)

        # initial embedding for state
        if state_dim > 0:
            self.init_linear_state = nn.Linear(state_dim, emb_dim)
        
        # An attention layer
        self.w_q = nn.Parameter(torch.FloatTensor((2 + int(state_dim > 0)) * emb_dim, emb_dim))
        self.w_k = nn.Parameter(torch.FloatTensor(2 * emb_dim, emb_dim))
        self.w_v = nn.Parameter(torch.FloatTensor(2 * emb_dim, emb_dim))

        # out linear layer
        self.out_linear = nn.Linear(emb_dim, emb_dim)

        # Dropout
        self.dropout = nn.Dropout(dropout)

        self.reset_parameters()

    def reset_parameters(self):
        for param in self.parameters():
            stdv = 1. / math.sqrt(param.size(-1))
            param.data.uniform_(-stdv, stdv)
    
    def forward(self, inputs, node_emb):
        """
        Paramters
        ---------
        inputs: dict
            curr_node_id: torch.LongTensor [batch_size]
            next_node_id: torch.LongTensor [batch_size]
            mask: torch.LongTensor [batch_size x num_nodes]
            state: torch.FloatTensor [batch_size x state_dim]
        node_emb: torch.tensor [batch_size x num_nodes x emb_dim]
            node embeddings obtained from the node encoder

        Returns
        -------
        h: torch.tensor [batch_size x emb_dim]
            edge embeddings 
        """
        curr_node_id = inputs["curr_node_id"]
        next_node_id = inputs["next_node_id"]
        mask = inputs["mask"]
        state = inputs["state"]
        batch_size = curr_node_id.size(0)

        #--------------------------------
        # generate queries, keys, values
        #--------------------------------
        node_emb = self.dropout(node_emb)
        curr_emb = node_emb.gather(1, curr_node_id[:, None, None].expand(batch_size, 1, self.emb_dim))
        next_emb = node_emb.gather(1, next_node_id[:, None, None].expand(batch_size, 1, self.emb_dim))
        if state is not None and self.state_dim > 0:
            state_emb = self.init_linear_state(state) # [batch_size x emb_dim]
            input_q  = torch.cat((curr_emb, next_emb, state_emb[:, None, :]), -1) # [batch_size x 1 x (3*emb_dim)]
        else:
            input_q  = torch.cat((curr_emb, next_emb), -1) # [batch_size x 1 x (2*emb_dim)]
        input_kv = torch.cat((curr_emb.expand_as(node_emb), node_emb), -1) # [batch_size x num_nodes x (2*emb_dim)]

        #--------------------
        # An attention layer
        #--------------------
        q = torch.matmul(input_q,  self.w_q) # [batch_size x 1 x emb_dim]
        k = torch.matmul(input_kv, self.w_k) # [batch_size x num_nodes x emb_dim]
        v = torch.matmul(input_kv, self.w_v) # [batch_size x num_nodes x emb_dim]
        compatibility = self.norm_factor * torch.matmul(q, k.transpose(-2, -1)) # [batch_size x 1 x num_nodes]
        compatibility[(~mask).unsqueeze(1).expand_as(compatibility)] = -math.inf
        attn = torch.softmax(compatibility, dim=-1)
        h = torch.matmul(attn, v) # [batch_size x 1 x emb_dim]
        h = h.squeeze(1) # [batch_size x emb_dim]
        return self.out_linear(h) + q.squeeze(1)