# Copyright 2022 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""EGConv Layer"""
from typing import List
import mindspore as ms
from mindspore.common.initializer import initializer
from mindspore.common.initializer import XavierUniform
from mindspore_gl import Graph
from .. import GNNCell
[文档]class EGConv(GNNCell):
r"""
Efficient Graph Convolution. From the paper `Adaptive Filters and Aggregator Fusion for Efficient Graph Convolutions
<https://arxiv.org/abs/2104.01481>`_ .
.. math::
h_i^{(l+1)} = {\LARGE ||}_{h=1}^{H} \sum_{\oplus \in \mathcal{A}} \sum_{b=1}^{B} w_{h,\oplus,b}^{(l)}
\bigoplus_{j \in \mathcal{N(i)}} W_{b}^{(l)} h_{j}^{(l)}
:math:`\mathcal{N}(i)` represents the neighbour node of :math:`i`,
:math:`W_{b}^{(l)}` represents a basis weight,
:math:`\oplus` represents an aggregator,
:math:`w_{h,\oplus,b}^{(l)}` represents per-vertex weighting coefficients across heads, aggregator and bases.
Args:
in_feat_size (int): Input node feature size.
out_feat_size (int): Output node feature size.
aggregators (str, optional): aggregators to be used. Supported aggregators are
`sum`, `mean`, `max`, `min`, `std`, `var`, `symnorm`. Default: 'symnorm'.
num_heads (int, optional): Number of heads :math:`H`. Default: 8.
Must have `out_feat_size % num_heads == 0`.
num_bases (int, optional): Number of basis weight :math:`B`. Default: 4.
bias (bool, optional): Whether the layer will learn an additive bias. Default: True.
Inputs:
- **x** (Tensor) - The input node features. The shape is :math:`(N, D_{in})`
where :math:`N` is the number of nodes,
and :math:`D_{in}` should be equal to `in_feat_size` in `Args`.
- **g** (Graph) - The input graph.
Outputs:
- Tensor, output node features with shape of :math:`(N, D_{out})`, where :math:`(D_{out})` should be the same as
`out_feat_size` in `Args`.
Raises:
TypeError: If `in_feat_size` or `out_feat_size` or `num_heads` is not a positive int.
ValueError: If `out_feat_size` is not divisible by 'num_heads'.
ValueError: If `aggregators` is not in ['sum', 'mean', 'max', 'min', 'symnorm', 'var', 'std'].
Supported Platforms:
``Ascend`` ``GPU``
Examples:
>>> import mindspore as ms
>>> from mindspore_gl.nn import EGConv
>>> from mindspore_gl import GraphField
>>> n_nodes = 4
>>> n_edges = 7
>>> feat_size = 4
>>> src_idx = ms.Tensor([0, 1, 1, 2, 2, 3, 3], ms.int32)
>>> dst_idx = ms.Tensor([0, 0, 2, 1, 3, 0, 1], ms.int32)
>>> ones = ms.ops.Ones()
>>> feat = ones((n_nodes, feat_size), ms.float32)
>>> graph_field = GraphField(src_idx, dst_idx, n_nodes, n_edges)
>>> conv = EGConv(in_feat_size=4, out_feat_size=6, aggregators=['sum'], num_heads=3, num_bases=3)
>>> res = conv(feat, *graph_field.get_graph())
>>> print(res.shape)
(4, 6)
"""
def __init__(self,
in_feat_size: int,
out_feat_size: int,
aggregators: List[str],
num_heads: int = 8,
num_bases: int = 4,
bias: bool = True):
super().__init__()
assert isinstance(in_feat_size, int) and in_feat_size > 0, "in_feat_size must be positive int"
assert isinstance(out_feat_size, int) and out_feat_size > 0, "out_feat_size must be positive int"
assert isinstance(num_heads, int) and num_heads > 0, "num_heads must be positive int"
self.in_feat_size = in_feat_size
self.out_feat_size = out_feat_size
self.num_heads = num_heads
if out_feat_size % num_heads != 0:
raise ValueError(f"For '{self.cls_name}', the 'out_feat_size' should be divisible by 'num_heads', "
f"but got out_feat_size: {out_feat_size}, num_heads: {num_heads}.")
self.num_bases = num_bases
for agg in aggregators:
if agg not in ['sum', 'mean', 'max', 'min', 'symnorm', 'var', 'std']:
raise ValueError(f"For '{self.cls_name}', the aggregator: '{agg}' is unsupported.")
self.agg_num = len(aggregators)
self.aggregators = aggregators
self.basis_fc = ms.nn.Dense(in_feat_size, (out_feat_size // num_heads) * num_bases,
weight_init=XavierUniform(), has_bias=False)
self.combine_fc = ms.nn.Dense(in_feat_size, num_heads * num_bases * self.agg_num,
weight_init=XavierUniform(), has_bias=True)
if bias:
self.bias = ms.Parameter(initializer('zero', (out_feat_size), ms.float32), name="bias")
else:
self.bias = None
self.reshape = ms.ops.Reshape()
self.matmul = ms.nn.MatMul()
self.sqrt = ms.ops.Sqrt()
self.relu = ms.ops.ReLU()
self.stack = ms.ops.Stack(axis=1)
self.eps = 1e-5
def _combine(self, weights, aggregated):
aggregated = aggregated.view(-1, self.agg_num * self.num_bases, self.out_feat_size // self.num_heads)
x = self.matmul(weights, aggregated)
x = x.view(-1, self.out_feat_size)
if self.bias is not None:
x = x + self.bias
return x
# pylint: disable=arguments-differ
def construct(self, x, g: Graph):
"""
Construct function for EGConv.
"""
bases = self.basis_fc(x)
weights = self.combine_fc(x)
weights = self.reshape(weights, (-1, self.num_heads, self.num_bases * self.agg_num))
outs = []
for agg in self.aggregators:
if agg == 'symnorm':
in_deg = self.sqrt(1.0 / g.in_degree())
g.set_vertex_attr({'x': bases, 'deg': in_deg})
for v in g.dst_vertex:
v.x = g.sum([u.deg * u.x for u in v.innbs])
out = [v.x for v in g.dst_vertex] * in_deg
elif agg in ['var', 'std']:
g.set_vertex_attr({'x': bases, 'x_square': bases * bases})
for v in g.dst_vertex:
v.x = g.avg([u.x for u in v.innbs])
v.x_square = g.avg([u.x_square for u in v.innbs])
out = [v.x_square - v.x * v.x for v in g.dst_vertex]
if agg == 'std':
out = self.sqrt(self.relu(out) + self.eps)
else:
g.set_vertex_attr({'x': bases})
for v in g.dst_vertex:
x_list = [u.x for u in v.innbs]
if agg == 'sum':
v.x = g.sum(x_list)
elif agg == 'mean':
v.x = g.avg(x_list)
elif agg == 'max':
v.x = g.max(x_list)
elif agg == 'min':
v.x = g.min(x_list)
out = [v.x for v in g.dst_vertex]
outs.append(out)
aggregated = self.stack(outs)
return self._combine(weights, aggregated)