"""Torch module for GraphSAGE."""
import torch
import torch.nn as nn
import torch.nn.functional as F
from grb.utils.normalize import SAGEAdjNorm
[docs]class GraphSAGE(nn.Module):
r"""
Description
-----------
Inductive Representation Learning on Large Graphs (`GraphSAGE <https://arxiv.org/abs/1706.02216>`__)
Parameters
----------
in_features : int
Dimension of input features.
out_features : int
Dimension of output features.
n_layers : int
Number of layers.
hidden_features : int or list of int
Dimension of hidden features. List if multi-layer.
layer_norm : bool, optional
Whether to use layer normalization. Default: ``False``.
activation : func of torch.nn.functional, optional
Activation function. Default: ``torch.nn.functional.relu``.
feat_norm : str, optional
Type of features normalization, choose from ["arctan", "tanh", None]. Default: ``None``.
adj_norm_func : func of utils.normalize, optional
Function that normalizes adjacency matrix. Default: ``SAGEAdjNorm``.
mu : float, optional
Hyper-parameter, refer to original paper. Default: ``2.0``.
dropout : float, optional
Rate of dropout. Default: ``0.0``.
"""
def __init__(self,
in_features,
out_features,
hidden_features,
n_layers,
activation=F.relu,
layer_norm=False,
feat_norm=None,
adj_norm_func=SAGEAdjNorm,
mu=2.0,
dropout=0.0):
super(GraphSAGE, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.feat_norm = feat_norm
self.adj_norm_func = adj_norm_func
if type(hidden_features) is int:
hidden_features = [hidden_features] * (n_layers - 1)
elif type(hidden_features) is list or type(hidden_features) is tuple:
assert len(hidden_features) == (n_layers - 1), "Incompatible sizes between hidden_features and n_layers."
n_features = [in_features] + hidden_features + [out_features]
self.layers = nn.ModuleList()
for i in range(n_layers):
if layer_norm:
self.layers.append(nn.LayerNorm(n_features[i]))
self.layers.append(SAGEConv(in_features=n_features[i],
pool_features=n_features[i],
out_features=n_features[i + 1],
activation=activation if i != n_layers - 1 else None,
mu=mu,
dropout=dropout if i != n_layers - 1 else 0.0))
self.reset_parameters()
@property
def model_type(self):
"""Indicate type of implementation."""
return "torch"
@property
def model_name(self):
return "graphsage"
[docs] def reset_parameters(self):
"""Reset parameters."""
for layer in self.layers:
layer.reset_parameters()
[docs] def forward(self, x, adj):
r"""
Parameters
----------
x : torch.Tensor
Tensor of input features.
adj : torch.SparseTensor
Sparse tensor of adjacency matrix.
Returns
-------
x : torch.Tensor
Output of model (logits without activation).
"""
for layer in self.layers:
if isinstance(layer, nn.LayerNorm):
x = layer(x)
else:
x = F.normalize(x, dim=1)
x = layer(x, adj)
return x
[docs]class SAGEConv(nn.Module):
r"""
Description
-----------
SAGE convolutional layer.
Parameters
----------
in_features : int
Dimension of input features.
pool_features : int
Dimension of pooling features.
out_features : int
Dimension of output features.
activation : func of torch.nn.functional, optional
Activation function. Default: ``None``.
dropout : float, optional
Rate of dropout. Default: ``0.0``.
mu : float, optional
Hyper-parameter, refer to original paper. Default: ``2.0``.
"""
def __init__(self,
in_features,
pool_features,
out_features,
activation=None,
mu=2.0,
dropout=0.0):
super(SAGEConv, self).__init__()
self.pool_layer = nn.Linear(in_features, pool_features)
self.linear1 = nn.Linear(pool_features, out_features)
self.linear2 = nn.Linear(pool_features, out_features)
self.activation = activation
self.mu = mu
if dropout > 0.0:
self.dropout = nn.Dropout(dropout)
else:
self.dropout = None
self.reset_parameters()
[docs] def reset_parameters(self):
"""Reset parameters."""
if self.activation == F.leaky_relu:
gain = nn.init.calculate_gain('leaky_relu')
else:
gain = nn.init.calculate_gain('relu')
nn.init.xavier_normal_(self.linear1.weight, gain=gain)
nn.init.xavier_normal_(self.linear2.weight, gain=gain)
nn.init.xavier_normal_(self.pool_layer.weight, gain=gain)
[docs] def forward(self, x, adj):
r"""
Parameters
----------
x : torch.Tensor
Tensor of input features.
adj : torch.SparseTensor
Sparse tensor of adjacency matrix.
Returns
-------
x : torch.Tensor
Output of layer.
"""
x = F.relu(self.pool_layer(x))
x_ = x ** self.mu
x_ = torch.spmm(adj, x_) ** (1 / self.mu)
# In original model this is actually max-pool, but **10/**0.1 result in gradient explosion.
# However we can still achieve similar performance using 2-norm.
x = self.linear1(x)
x_ = self.linear2(x_)
x = x + x_
if self.activation is not None:
x = self.activation(x)
if self.dropout is not None:
x = self.dropout(x)
return x