##########################################################################
# basf2 (Belle II Analysis Software Framework) #
# Author: The Belle II Collaboration #
# #
# See git log for contributors and copyright holders. #
# This file is licensed under LGPL-3.0, see LICENSE.md. #
##########################################################################
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_scatter import scatter
def _init_weights(layer, normalize):
"""
Initializes the weights and biases.
"""
for m in layer.modules():
if isinstance(m, nn.Linear):
nn.init.xavier_normal_(m.weight.data)
if not normalize:
m.bias.data.fill_(0.1)
elif isinstance(m, nn.BatchNorm1d) or isinstance(m, nn.LayerNorm):
m.weight.data.fill_(1)
m.bias.data.zero_()
[docs]class EdgeLayer(nn.Module):
"""
Updates edge features in MetaLayer:
.. math::
e_{ij}^{'} = \\phi^{e}(e_{ij}, v_{i}, v_{j}, u),
where :math:`\\phi^{e}` is a neural network of the form
.. figure:: figs/MLP_structure.png
:width: 42em
:align: center
Args:
nfeat_in_dim (int): Node features input dimension (number of node features in input).
efeat_in_dim (int): Edge features input dimension (number of edge features in input).
gfeat_in_dim (int): Gloabl features input dimension (number of global features in input).
efeat_hid_dim (int): Edge features dimension in hidden layers.
efeat_out_dim (int): Edge features output dimension.
num_hid_layers (int): Number of hidden layers.
dropout (float): Dropout rate :math:`r \\in [0,1]`.
normalize (str): Type of normalization (batch/layer).
:return: Updated edge features tensor.
:rtype: `Tensor <https://pytorch.org/docs/stable/tensors.html#torch.Tensor>`_
"""
def __init__(
self,
nfeat_in_dim,
efeat_in_dim,
gfeat_in_dim,
efeat_hid_dim,
efeat_out_dim,
num_hid_layers,
dropout,
normalize=True,
):
"""
Initialization.
"""
super(EdgeLayer, self).__init__()
#: Non-linear activation
self.nonlin_function = F.elu
#: Number of hidden layers
self.num_hid_layers = num_hid_layers
#: Dropout probability
self.dropout_prob = dropout
#: Normalization
self.normalize = normalize
#: Linear input layer
self.lin_in = nn.Linear(
efeat_in_dim + 2 * nfeat_in_dim + gfeat_in_dim, efeat_hid_dim
)
#: Intermediate linear layers
self.lins_hid = nn.ModuleList(
[
nn.Linear(efeat_hid_dim, efeat_hid_dim)
for _ in range(self.num_hid_layers)
]
)
#: Output linear layer
self.lin_out = nn.Linear(efeat_hid_dim, efeat_out_dim, bias=not normalize)
if normalize:
#: Batch normalization
self.norm = nn.BatchNorm1d(efeat_out_dim)
_init_weights(self, normalize)
def forward(self, src, dest, edge_attr, u, batch):
"""
Called internally by PyTorch to propagate the input through the network.
- src, dest: [E, F_x], where E is the number of edges.
- edge_attr: [E, F_e]
- u: [B, F_u], where B is the number of graphs.
- batch: [E] with max entry B - 1.
"""
out = (
torch.cat([edge_attr, src, dest, u[batch]], dim=1)
if u.shape != torch.Size([0])
else torch.cat([edge_attr, src, dest], dim=1)
)
out = self.nonlin_function(self.lin_in(out))
out = F.dropout(out, self.dropout_prob, training=self.training)
out_skip = out
for lin_hid in self.lins_hid:
out = self.nonlin_function(lin_hid(out))
out = F.dropout(out, self.dropout_prob, training=self.training)
if self.num_hid_layers > 1:
out += out_skip
if self.normalize:
out = self.nonlin_function(self.norm(self.lin_out(out)))
else:
out = self.nonlin_function(self.lin_out(out))
return out
[docs]class NodeLayer(nn.Module):
"""
Updates node features in MetaLayer:
.. math::
v_{i}^{'} = \\phi^{v}(v_{i}, \\rho^{e \\to v}(v_{i}), u)
with
.. math::
\\rho^{e \\to v}(v_{i}) = \\frac{\\sum_{j=1,\\ j \\neq i}^{N} (e_{ji} + e _{ij})}{2 \\cdot (N-1)},
where :math:`\\phi^{v}` is a neural network of the form
.. figure:: figs/MLP_structure.png
:width: 42em
:align: center
Args:
nfeat_in_dim (int): Node features input dimension (number of node features in input).
efeat_in_dim (int): Edge features input dimension (number of edge features in input).
gfeat_in_dim (int): Gloabl features input dimension (number of global features in input).
nfeat_hid_dim (int): Node features dimension in hidden layers.
nfeat_out_dim (int): Node features output dimension.
num_hid_layers (int): Number of hidden layers.
dropout (float): Dropout rate :math:`r \\in [0,1]`.
normalize (str): Type of normalization (batch/layer).
:return: Updated node features tensor.
:rtype: `Tensor <https://pytorch.org/docs/stable/tensors.html#torch.Tensor>`_
"""
def __init__(
self,
nfeat_in_dim,
efeat_in_dim,
gfeat_in_dim,
nfeat_hid_dim,
nfeat_out_dim,
num_hid_layers,
dropout,
normalize=True,
):
"""
Initialization.
"""
super(NodeLayer, self).__init__()
#: Non-linear activation
self.nonlin_function = F.elu
#: Number of hidden layers
self.num_hid_layers = num_hid_layers
#: Dropout probability
self.dropout_prob = dropout
#: Normalize
self.normalize = normalize
#: Input linear layer
self.lin_in = nn.Linear(
gfeat_in_dim + nfeat_in_dim + efeat_in_dim, nfeat_hid_dim
)
#: Intermediate linear layers
self.lins_hid = nn.ModuleList(
[
nn.Linear(nfeat_hid_dim, nfeat_hid_dim)
for _ in range(self.num_hid_layers)
]
)
#: Output linear layer
self.lin_out = nn.Linear(nfeat_hid_dim, nfeat_out_dim, bias=not normalize)
if normalize:
#: Batch normalization
self.norm = nn.BatchNorm1d(nfeat_out_dim)
_init_weights(self, normalize)
def forward(self, x, edge_index, edge_attr, u, batch):
"""
Called internally by PyTorch to propagate the input through the network.
- x: [N, F_x], where N is the number of nodes.
- edge_index: [2, E] with max entry N - 1.
- edge_attr: [E, F_e]
- u: [B, F_u]
- batch: [N] with max entry B - 1.
Edge labels are averaged (dim_size = N: number of nodes in the graph)
"""
out = scatter(
edge_attr, edge_index[1], dim=0, dim_size=batch.size(0), reduce="mean"
)
out = (
torch.cat([x, out, u[batch]], dim=1)
if u.shape != torch.Size([0])
else torch.cat([x, out], dim=1)
)
out = self.nonlin_function(self.lin_in(out))
out = F.dropout(out, self.dropout_prob, training=self.training)
out_skip = out
for lin_hid in self.lins_hid:
out = self.nonlin_function(lin_hid(out))
out = F.dropout(out, self.dropout_prob, training=self.training)
if self.num_hid_layers > 1:
out += out_skip
if self.normalize:
out = self.nonlin_function(self.norm(self.lin_out(out)))
else:
out = self.nonlin_function(self.lin_out(out))
return out
[docs]class GlobalLayer(nn.Module):
"""
Updates node features in MetaLayer:
.. math::
u_{i}^{'} = \\phi^{u}(\\rho^{e \\to u}(e), \\rho^{v \\to u}(v), u)
with
.. math::
\\rho^{e \\to u}(e) = \\frac{\\sum_{i,j=1,\\ i \\neq j}^{N} e_{ij}}{N \\cdot (N-1)},\\\\
\\rho^{v \\to u}(e) = \\frac{\\sum_{i=1}^{N} v_{i}}{N},
where :math:`\\phi^{u}` is a neural network of the form
.. figure:: figs/MLP_structure.png
:width: 42em
:align: center
Args:
nfeat_in_dim (int): Node features input dimension (number of node features in input).
efeat_in_dim (int): Edge features input dimension (number of edge features in input).
gfeat_in_dim (int): Gloabl features input dimension (number of global features in input).
nfeat_hid_dim (int): Global features dimension in hidden layers.
nfeat_out_dim (int): Global features output dimension.
num_hid_layers (int): Number of hidden layers.
dropout (float): Dropout rate :math:`r \\in [0,1]`.
normalize (str): Type of normalization (batch/layer).
:return: Updated global features tensor.
:rtype: `Tensor <https://pytorch.org/docs/stable/tensors.html#torch.Tensor>`_
"""
def __init__(
self,
nfeat_in_dim,
efeat_in_dim,
gfeat_in_dim,
gfeat_hid_dim,
gfeat_out_dim,
num_hid_layers,
dropout,
normalize=True,
):
"""
Initialization.
"""
super(GlobalLayer, self).__init__()
#: Non-linear activation
self.nonlin_function = F.elu
#: Number of hidden layers
self.num_hid_layers = num_hid_layers
#: Dropout probability
self.dropout_prob = dropout
#: Normalization
self.normalize = normalize
#: Input linear layer
self.lin_in = nn.Linear(
nfeat_in_dim + efeat_in_dim + gfeat_in_dim, gfeat_hid_dim
)
#: Intermediate linear layers
self.lins_hid = nn.ModuleList(
[
nn.Linear(gfeat_hid_dim, gfeat_hid_dim)
for _ in range(self.num_hid_layers)
]
)
#: Output linear layer
self.lin_out = nn.Linear(gfeat_hid_dim, gfeat_out_dim, bias=not normalize)
if normalize:
#: Batch normalization
self.norm = nn.BatchNorm1d(gfeat_out_dim)
_init_weights(self, normalize)
def forward(self, x, edge_index, edge_attr, u, batch):
"""
Called internally by Pytorch to propagate the input through the network.
- x: [N, F_x], where N is the number of nodes.
- edge_index: [2, E] with max entry N - 1.
- edge_attr: [E, F_e]
- u: [B, F_u]
- batch: [N] with max entry B - 1.
Nodes are averaged over graph
"""
node_mean = scatter(
x, batch, dim=0, reduce="mean"
)
# Edges are averaged over nodes
edge_mean = scatter(
edge_attr, edge_index[1], dim=0, reduce="mean"
)
# Edges are averaged over graph
edge_mean = scatter(
edge_mean, batch, dim=0, reduce="mean"
)
out = (
torch.cat([u, node_mean, edge_mean], dim=1)
if u.shape != torch.Size([0])
else torch.cat([node_mean, edge_mean], dim=1)
)
out = self.nonlin_function(self.lin_in(out))
out = F.dropout(out, self.dropout_prob, training=self.training)
out_skip = out
for lin_hid in self.lins_hid:
out = self.nonlin_function(lin_hid(out))
out = F.dropout(out, self.dropout_prob, training=self.training)
if self.num_hid_layers > 1:
out += out_skip
if self.normalize:
out = self.nonlin_function(self.norm(self.lin_out(out)))
else:
out = self.lin_out(out)
return out
