HetSANN.py

import dgl
import torch
import torch.nn as nn
import dgl.function as Fn
import torch.nn.functional as F

from dgl.ops import edge_softmax
from dgl.nn.pytorch import TypedLinear
from ..utils import to_hetero_feat
from . import BaseModel, register_model

@register_model('HetSANN')
class HetSANN(BaseModel):
    r"""
    This is a model HetSANN from `An Attention-Based Graph Neural Network for Heterogeneous Structural Learning
    <https://arxiv.org/abs/1912.10832>`__

    It contains the following part:

    Apply a linear transformation:
    
    .. math::
       h^{(l+1, m)}_{\phi(j),i} = W^{(l+1, m)}_{\phi(j),\phi(i)} h^{(l)}_i \quad (1)
    
    And return the new embeddings.
    
    You may refer to the paper HetSANN-Section 2.1-Type-aware Attention Layer-(1)

    Aggregation of Neighborhood:
    
    Computing the attention coefficient:
    
    .. math::
       o^{(l+1,m)}_e = \sigma(f^{(l+1,m)}_r(h^{(l+1, m)}_{\phi(j),j}, h^{(l+1, m)}_{\phi(j),i})) \quad (2)
        
    .. math::
       f^{(l+1,m)}_r(e) = [h^{(l+1, m)^T}_{\phi(j),j}||h^{(l+1, m)^T}_{\phi(j),i}]a^{(l+1, m)}_r ] \quad (3)
    
    .. math::
       \alpha^{(l+1,m)}_e = exp(o^{(l+1,m)}_e) / \sum_{k\in \varepsilon_j} exp(o^{(l+1,m)}_k) \quad (4)
    
    Getting new embeddings with multi-head and residual
    
    .. math::
       h^{(l + 1, m)}_j = \sigma(\sum_{e = (i,j,r)\in \varepsilon_j} \alpha^{(l+1,m)}_e h^{(l+1, m)}_{\phi(j),i}) \quad (5)
    
    Multi-heads:
    
    .. math::
       h^{(l+1)}_j = \parallel^M_{m = 1}h^{(l + 1, m)}_j \quad (6)
    
    Residual:
    
    .. math::
       h^{(l+1)}_j = h^{(l)}_j + \parallel^M_{m = 1}h^{(l + 1, m)}_j \quad (7)
    
    Parameters
    ----------
    num_heads: int
        the number of heads in the attention computing
    num_layers: int
        the number of layers we used in the computing
    in_dim: int
        the input dimension
    num_classes: int
        the number of the output classes
    num_etypes: int
        the number of the edge types
    dropout: float
        the dropout rate
    negative_slope: float
        the negative slope used in the LeakyReLU
    residual: boolean
        if we need the residual operation
    ntype: list
        the list of node type
    """
    @classmethod
    def build_model_from_args(cls, args, hg):
        return cls(
            args.num_heads,
            args.num_layers,
            args.hidden_dim,
            args.out_dim,
            hg.ntypes,
            len(hg.etypes),
            args.dropout,
            args.slope,
            args.residual,
            )
    
    def __init__(self, num_heads, num_layers, in_dim, num_classes,
                ntypes, num_etypes, dropout, negative_slope, residual):
        super(HetSANN, self).__init__()
        self.num_layers = num_layers
        self.ntypes = ntypes
        # self.dropout = nn.Dropout(dropout)
        self.residual = residual
        self.activation = F.elu
        
        self.het_layers = nn.ModuleList()
        
        # input projection
        self.het_layers.append(
            HetSANNConv(
                num_heads,
                in_dim,
                in_dim // num_heads,
                num_etypes,
                dropout,
                negative_slope,
                False,
                self.activation,
            )
        )

        # hidden layer
        for i in range(1, num_layers - 1):
            self.het_layers.append(
                HetSANNConv(
                    num_heads,
                    in_dim,
                    in_dim // num_heads,
                    num_etypes,
                    dropout,
                    negative_slope,
                    residual,
                    self.activation
                )
            )

        # output projection
        self.het_layers.append(
            HetSANNConv(
                1,
                in_dim,
                num_classes,
                num_etypes,
                dropout,
                negative_slope,
                residual,
                None,
            )
        )

    def forward(self, hg, h_dict):
        """
        The forward part of the HetSANN.
        
        Parameters
        ----------
        hg : object
            the dgl heterogeneous graph
        h_dict: dict
            the feature dict of different node types
            
        Returns
        -------
        dict
            The embeddings after the output projection.
        """
        if hasattr(hg, 'ntypes'):
            with hg.local_scope():
                # input layer and hidden layers
                hg.ndata['h'] = h_dict
                g = dgl.to_homogeneous(hg, ndata = 'h')
                h = g.ndata['h']
                for i in range(self.num_layers - 1):
                    h = self.het_layers[i](g, h, g.ndata['_TYPE'], g.edata['_TYPE'], True)

                # output layer
                h = self.het_layers[-1](g, h, g.ndata['_TYPE'], g.edata['_TYPE'], True)

                h_dict = to_hetero_feat(h, g.ndata['_TYPE'], self.ntypes)
            
        else:
            # for minibatch training, input h_dict is a tensor
            h = h_dict
            for layer, block in zip(self.het_layers, hg):
                h = layer(block, h, block.ndata['_TYPE']['_N'], block.edata['_TYPE'], presorted=False)
            h_dict = to_hetero_feat(h, block.ndata['_TYPE']['_N'][:block.num_dst_nodes()], self.ntypes)

        return h_dict
    
    @property
    def to_homo_flag(self):
        return True

class HetSANNConv(nn.Module):
    """
    The HetSANN convolution layer.

    Parameters
    ----------
    num_heads: int
        the number of heads in the attention computing
    in_dim: int
        the input dimension of the features
    hidden_dim: int
        the hidden dimension of the features
    num_etypes: int
        the number of the edge types
    dropout: float
        the dropout rate
    negative_slope: float
        the negative slope used in the LeakyReLU
    residual: boolean
        if we need the residual operation
    activation: str
        the activation function
    """
    def __init__(self, num_heads, in_dim, hidden_dim, num_etypes,
                 dropout, negative_slope, residual, activation):
        super(HetSANNConv, self).__init__()
        self.num_heads = num_heads
        self.in_dim = in_dim
        self.hidden_dim = hidden_dim

        self.W = TypedLinear(self.in_dim, self.hidden_dim * self.num_heads, num_etypes)
        # self.W_out = TypedLinear(hidden_dim * num_heads, num_classes, num_etypes)

        # self.W_hidden = nn.ModuleDict()
        # self.W_out = nn.ModuleDict()
        
        # for etype in etypes:
        #     self.W_hidden[etype] = nn.Linear(in_dim, hidden_dim * num_heads)
        
        # for etype in etypes:
        #     self.W_out[etype] = nn.Linear(hidden_dim * num_heads, num_classes)
        
        self.a_l = TypedLinear(self.hidden_dim * self.num_heads, self.hidden_dim * self.num_heads, num_etypes)
        self.a_r = TypedLinear(self.hidden_dim * self.num_heads, self.hidden_dim * self.num_heads, num_etypes)
        
        self.dropout = nn.Dropout(dropout)
        self.leakyrelu = nn.LeakyReLU(negative_slope)
        if residual:
            self.residual = nn.Linear(in_dim, self.hidden_dim * num_heads)
        else:
            self.register_buffer("residual", None)
            
        self.activation = activation
        
        
    def forward(self, g, x, ntype, etype, presorted = False):
        """
        The forward part of the HetSANNConv.

        Parameters
        ----------
        g : object
            the dgl homogeneous graph
        x: tensor
            the original features of the graph
        ntype: tensor
            the node type of the graph
        etype: tensor
            the edge type of the graph
        presorted: boolean
            if the ntype and etype are preordered, default: ``False``
            
        Returns
        -------
        tensor
            The embeddings after aggregation.
        """
        # formula (1)
        g.srcdata['h'] = x
        g.apply_edges(Fn.copy_u('h', 'm'))
        h = g.edata['m']
        feat = self.W(h, etype, presorted)
        h = self.dropout(feat)
        g.edata['m'] = h
        h = h.view(-1, self.num_heads, self.hidden_dim)

        # formula (2) (3) (4)
        h_l = self.a_l(h.view(-1, self.num_heads * self.hidden_dim), etype, presorted) \
        .view(-1, self.num_heads, self.hidden_dim).sum(dim = -1)

        h_r = self.a_r(h.view(-1, self.num_heads * self.hidden_dim), etype, presorted) \
        .view(-1, self.num_heads, self.hidden_dim).sum(dim = -1)
        
        attention = self.leakyrelu(h_l + h_r)
        attention = edge_softmax(g, attention)
        
        # formula (5) (6)
        with g.local_scope():
            h = h.permute(0, 2, 1).contiguous()
            g.edata['alpha'] = h @ attention.reshape(-1, self.num_heads, 1)
            
            g.update_all(Fn.copy_e('m', 'w'),Fn.sum('w', 'emb'))
            h_output = g.dstdata['emb']
            # h_prime = []
            # h = h.permute(1, 0, 2).contiguous()
            # for i in range(self.num_heads):
            #     g.edata['alpha'] = attention[:, i]
            #     g.srcdata.update({'emb': h[i]})
            #     g.update_all(Fn.u_mul_e('emb', 'alpha', 'm'),
            #                  Fn.sum('m', 'emb'))
            #     h_prime.append(g.ndata['emb'])
            # h_output = torch.cat(h_prime, dim=1)

        # formula (7)
        if g.is_block:
            x = x[:g.num_dst_nodes()] 
        if self.residual:
            res = self.residual(x)
            h_output += res
        
        if self.activation is not None:
            h_output = self.activation(h_output)       
        
        return h_output