# -*- coding: utf-8 -*-
""" Graph Convolutional Network Autoencoder
"""
# Author: Kay Liu <zliu234@uic.edu>
# License: BSD 2 clause
import torch
import numpy as np
import torch.nn.functional as F
from torch_geometric.loader import NeighborLoader
from sklearn.utils.validation import check_is_fitted
from . import BaseDetector
from .basic_nn import GCN
from ..utils.utility import validate_device
from ..utils.metric import eval_roc_auc
[docs]class GCNAE(BaseDetector):
"""
Vanila Graph Convolutional Networks Autoencoder
See :cite:`yuan2021higher` for details.
Parameters
----------
hid_dim : int, optional
Hidden dimension of model. Default: ``0``.
num_layers : int, optional
Total number of layers in autoencoders. Default: ``4``.
dropout : float, optional
Dropout rate. Default: ``0.``.
weight_decay : float, optional
Weight decay (L2 penalty). Default: ``0.``.
act : callable activation function or None, optional
Activation function if not None.
Default: ``torch.nn.functional.relu``.
contamination : float, optional
Valid in (0., 0.5). The proportion of outliers in the data set.
Used when fitting to define the threshold on the decision
function. Default: ``0.1``.
lr : float, optional
Learning rate. Default: ``0.004``.
epoch : int, optional
Maximum number of training epoch. Default: ``100``.
gpu : int
GPU Index, -1 for using CPU. Default: ``0``.
batch_size : int, optional
Minibatch size, 0 for full batch training. Default: ``0``.
num_neigh : int, optional
Number of neighbors in sampling, -1 for all neighbors.
Default: ``-1``.
verbose : bool
Verbosity mode. Turn on to print out log information.
Default: ``False``.
Examples
--------
>>> from pygod.models import GCNAE
>>> model = GCNAE()
>>> model.fit(data) # PyG graph data object
>>> prediction = model.predict(data)
"""
def __init__(self,
hid_dim=64,
num_layers=4,
dropout=0.3,
weight_decay=0.,
act=F.relu,
contamination=0.1,
lr=5e-3,
epoch=100,
gpu=0,
batch_size=0,
num_neigh=-1,
verbose=False):
super(GCNAE, self).__init__(contamination=contamination)
# model param
self.hid_dim = hid_dim
self.num_layers = num_layers
self.dropout = dropout
self.weight_decay = weight_decay
self.act = act
# training param
self.lr = lr
self.epoch = epoch
self.device = validate_device(gpu)
self.batch_size = batch_size
self.num_neigh = num_neigh
# other param
self.verbose = verbose
self.model = None
[docs] def fit(self, G, y_true=None):
"""
Fit detector with input data.
Parameters
----------
G : torch_geometric.data.Data
The input data.
y_true : numpy.ndarray, optional
The optional outlier ground truth labels used to monitor
the training progress. They are not used to optimize the
unsupervised model. Default: ``None``.
Returns
-------
self : object
Fitted estimator.
"""
G.node_idx = torch.arange(G.x.shape[0])
if self.batch_size == 0:
self.batch_size = G.x.shape[0]
loader = NeighborLoader(G,
[self.num_neigh] * self.num_layers,
batch_size=self.batch_size)
self.model = GCN(in_channels=G.x.shape[1],
hidden_channels=self.hid_dim,
num_layers=self.num_layers,
out_channels=G.x.shape[1],
dropout=self.dropout,
act=self.act).to(self.device)
optimizer = torch.optim.Adam(self.model.parameters(),
lr=self.lr,
weight_decay=self.weight_decay)
self.model.train()
decision_scores = np.zeros(G.x.shape[0])
for epoch in range(self.epoch):
epoch_loss = 0
for sampled_data in loader:
batch_size = sampled_data.batch_size
node_idx = sampled_data.node_idx
x, edge_index = self.process_graph(sampled_data)
x_ = self.model(x, edge_index)
score = torch.mean(F.mse_loss(x_, x, reduction='none'), dim=1)
decision_scores[node_idx[:batch_size]] = score.detach()\
.cpu().numpy()
loss = torch.mean(score)
epoch_loss += loss.item() * batch_size
optimizer.zero_grad()
loss.backward()
optimizer.step()
if self.verbose:
print("Epoch {:04d}: Loss {:.4f}"
.format(epoch, epoch_loss / G.x.shape[0]), end='')
if y_true is not None:
auc = eval_roc_auc(y_true, decision_scores)
print(" | AUC {:.4f}".format(auc), end='')
print()
self.decision_scores_ = decision_scores
self._process_decision_scores()
return self
[docs] def decision_function(self, G):
"""
Predict raw anomaly score using the fitted detector. Outliers
are assigned with larger anomaly scores.
Parameters
----------
G : PyTorch Geometric Data instance (torch_geometric.data.Data)
The input data.
Returns
-------
outlier_scores : numpy.ndarray
The anomaly score of shape :math:`N`.
"""
check_is_fitted(self, ['model'])
G.node_idx = torch.arange(G.x.shape[0])
loader = NeighborLoader(G,
[self.num_neigh] * self.num_layers,
batch_size=self.batch_size)
self.model.eval()
outlier_scores = np.zeros(G.x.shape[0])
for sampled_data in loader:
batch_size = sampled_data.batch_size
node_idx = sampled_data.node_idx
x, edge_index = self.process_graph(sampled_data)
x_ = self.model(x, edge_index)
score = torch.mean(F.mse_loss(x_, x, reduction='none'), dim=1)
outlier_scores[node_idx[:batch_size]] = score.detach() \
.cpu().numpy()
return outlier_scores
def process_graph(self, G):
"""
Process the raw PyG data object into a tuple of sub data
objects needed for the model.
Parameters
----------
G : PyTorch Geometric Data instance (torch_geometric.data.Data)
The input data.
Returns
-------
x : torch.Tensor
Attribute (feature) of nodes.
edge_index : torch.Tensor
Edge list of the graph.
"""
edge_index = G.edge_index.to(self.device)
x = G.x.to(self.device)
return x, edge_index