Graph kernel Network 代码理解

最近仔细阅读并研究了Graph kernel Network 这篇论文，论文中的公式还是很好理解的，作者同时提供了源代码，但是代码中有些符号和自定义函数没有加以注释，本人根据论文中的公式溯源代码，给出了部分注释，供参考。首先来看理论部分

公式(7)用来变换初始数据的维度，公式(8)是核积分迭代，（9）输出变换。（8）利用如下（10）实现。

再看代码部分，主要为如下几个函数。

width = 64
ker_width = 1024
depth = 6
edge_features = 6
node_features = 6

model = KernelNN(width, ker_width, depth, edge_features, node_features).cuda()

optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate, weight_decay=5e-4)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=scheduler_step, gamma=scheduler_gamma)

myloss = LpLoss(size_average=False)
u_normalizer.cuda()

model.train()

# in_width 是最初的数据点的维度，如论文中的d=6
# width 是变换后的维度，如变换后的 d'=64
# ker_in 是进入 kernel-NN 的边特征的维度，d=6
# ker_width=1024
# depth 是执行 kernel-NN的次数
class KernelNN(torch.nn.Module):
    def __init__(self, width, ker_width, depth, ker_in, in_width=1, out_width=1):
        super(KernelNN, self).__init__()
        self.depth = depth
		
		# fc1 用于将点的初始维度变换到高的维度
        self.fc1 = torch.nn.Linear(in_width, width)

        # 构建kernel网络 此处列表[ker_in, ker_width, ker_width, width**2]传入 DenseNet中的layers, 也就是说网络为4个隐藏层的DNN网络
        # ker_in 是输入kernel的维度，为 6
        # torch.nn.ReLU 传入 nonlinearity, 对每一层做非线性变换
        kernel = DenseNet([ker_in, ker_width, ker_width, width**2], torch.nn.ReLU)
        self.conv1 = NNConv_old(width, width, kernel, aggr='mean')

		# fc2 用于将点的最终的特征维度变换到维度为1
        self.fc2 = torch.nn.Linear(width, 1)

    def forward(self, data):
        x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr
        x = self.fc1(x)
        for k in range(self.depth):
            x = F.relu(self.conv1(x, edge_index, edge_attr))

        x = self.fc2(x)
        return x

下面看实现Kernel-NN需要的模块。

self.fc1 = torch.nn.Linear(in_width, width)
Applies a linear transformation to input data: y = xA^T + b
in_features: size of each input sample
out_features: size of each output sample
这里是没有激活函数的

class DenseNet(torch.nn.Module):
    def __init__(self, layers, nonlinearity, out_nonlinearity=None, normalize=False):
        super(DenseNet, self).__init__()

        self.n_layers = len(layers) - 1

        assert self.n_layers >= 1

        self.layers = nn.ModuleList()

        for j in range(self.n_layers):
            self.layers.append(nn.Linear(layers[j], layers[j+1]))

            if j != self.n_layers - 1:
                if normalize:
                    self.layers.append(nn.BatchNorm1d(layers[j+1]))

                self.layers.append(nonlinearity())

        if out_nonlinearity is not None:
            self.layers.append(out_nonlinearity())

    def forward(self, x):
        for _, l in enumerate(self.layers):
            x = l(x)

        return x

如下为pytorch geometric 中的代码，初始化该模块时的参变量为input_dim， out_dim 和 NN（一个神经网络）

class NNConv_old(MessagePassing):
    r"""The continuous kernel-based convolutional operator from the
    `"Neural Message Passing for Quantum Chemistry"
    <https://arxiv.org/abs/1704.01212>`_ paper.
    This convolution is also known as the edge-conditioned convolution from the
    `"Dynamic Edge-Conditioned Filters in Convolutional Neural Networks on
    Graphs" <https://arxiv.org/abs/1704.02901>`_ paper (see
    :class:`torch_geometric.nn.conv.ECConv` for an alias):

    .. math::
        \mathbf{x}^{\prime}_i = \mathbf{\Theta} \mathbf{x}_i +
        \sum_{j \in \mathcal{N}(i)} \mathbf{x}_j \cdot
        h_{\mathbf{\Theta}}(\mathbf{e}_{i,j}),

    where :math:`h_{\mathbf{\Theta}}` denotes a neural network, *.i.e.*
    a MLP.

    Args:
        in_channels (int): Size of each input sample.
        out_channels (int): Size of each output sample.
        nn (torch.nn.Module): A neural network :math:`h_{\mathbf{\Theta}}` that
            maps edge features :obj:`edge_attr` of shape :obj:`[-1,
            num_edge_features]` to shape
            :obj:`[-1, in_channels * out_channels]`, *e.g.*, defined by
            :class:`torch.nn.Sequential`.
        aggr (string, optional): The aggregation scheme to use
            (:obj:`"add"`, :obj:`"mean"`, :obj:`"max"`).
            (default: :obj:`"add"`)
        root_weight (bool, optional): If set to :obj:`False`, the layer will
            not add the transformed root node features to the output.
            (default: :obj:`True`)
        bias (bool, optional): If set to :obj:`False`, the layer will not learn
            an additive bias. (default: :obj:`True`)
        **kwargs (optional): Additional arguments of
            :class:`torch_geometric.nn.conv.MessagePassing`.
    """

    def __init__(self,
                 in_channels,
                 out_channels,
                 nn,
                 aggr='add',
                 root_weight=True,
                 bias=True,
                 **kwargs):
        super(NNConv_old, self).__init__(aggr=aggr, **kwargs)

        self.in_channels = in_channels
        self.out_channels = out_channels
        self.nn = nn
        self.aggr = aggr

        if root_weight:
            self.root = Parameter(torch.Tensor(in_channels, out_channels))
        else:
            self.register_parameter('root', None)

        if bias:
            self.bias = Parameter(torch.Tensor(out_channels))
        else:
            self.register_parameter('bias', None)

        self.reset_parameters()

    def reset_parameters(self):
        reset(self.nn)
        size = self.in_channels
        uniform(size, self.root)
        uniform(size, self.bias)

    def forward(self, x, edge_index, edge_attr):
        """"""
        x = x.unsqueeze(-1) if x.dim() == 1 else x
        pseudo = edge_attr.unsqueeze(-1) if edge_attr.dim() == 1 else edge_attr
        return self.propagate(edge_index, x=x, pseudo=pseudo)

    def message(self, x_j, pseudo):
        weight = self.nn(pseudo).view(-1, self.in_channels, self.out_channels)
        return torch.matmul(x_j.unsqueeze(1), weight).squeeze(1)

    def update(self, aggr_out, x):
        if self.root is not None:
            aggr_out = aggr_out + torch.mm(x, self.root)
        if self.bias is not None:
            aggr_out = aggr_out + self.bias
        return aggr_out

    def __repr__(self):
        return '{}({}, {})'.format(self.__class__.__name__, self.in_channels,
                                   self.out_channels)

代码中的其余部分为处理数据，感觉很乱，研究中。。。。