Skip to content

gdmnl/Spectral-GNN-Benchmark

Repository files navigation

Benchmarking Spectral Graph Neural Networks

Docs arXiv License Contrib Python PyTorch

pyg_spectral is a PyTorch Geometric-based framework for analyzing, implementing, and benchmarking spectral GNNs with effectiveness and efficiency evaluations. Our preliminary paper is available on arXiv. Artifact and additional results can be found in the Appendix.

Important

Why this project?
We list the following highlights of our framework compared to PyG and similar works:

  • Unified Framework: We offer a plug-and-play collection for spectral models and filters in unified and efficient implementations, rather than a model-specific design. Our rich collection greatly extends the PyG model zoo.
  • Spectral-oriented Design: We decouple non-spectral designs and feature the pivotal spectral kernel being consistent throughout different settings. Most filters are thus easily adaptable to a wide range of model-level options, including those provided by PyG and PyG-based frameworks.
  • High scalability: As spectral GNNs are inherently suitable for large-scale learning, our framework is feasible to common scalable learning schemes and acceleration techniques. Several spectral-oriented approximation algorithms are also supported.

Installation

This package can be easily installed by running pip at package root path:

pip install -r requirements.txt
pip install -e .[benchmark]

The installation script already covers the following core dependencies:

For additional installation of the C++ backend, please refer to propagations/README.md.

Reproduce Experiments

Main Experiments

Acquire results on the effectiveness and efficiency of spectral GNNs. Datasets will be automatically downloaded and processed by the code.

Run full-batch models:

cd benchmark
bash scripts/runfb.sh

Run mini-batch models:

bash scripts/runmb.sh

Additional Experiments

Effect of graph normalization:

bash scripts/eval_degng.sh

Figures can be plotted by: benchmark/notebook/fig_degng.ipynb.

Effect of propagation hops:

bash scripts/eval_hop.sh

Figures can be plotted by: benchmark/notebook/fig_hop.ipynb.

Frequency response:

bash scripts/exp_regression.sh

Customization

Configure Experiment Parameters

Refer to the help text by:

python benchmark/run_single.py --help
usage: python run_single.py
options:
    --help                      show this help message and exit
    # Logging configuration
    --seed SEED                 random seed
    --dev DEV                   GPU id
    --suffix SUFFIX             Result log file name. None:not saving results
    -quiet                      File log. True:dry run without saving logs
    --storage {state_file,state_ram,state_gpu}
                                Checkpoint log storage scheme.
    --loglevel LOGLEVEL         Console log. 10:progress, 15:train, 20:info, 25:result
    # Data configuration
    --data DATA                 Dataset name
    --data_split DATA_SPLIT     Index or percentage of dataset split
    --normg NORMG               Generalized graph norm
    --normf [NORMF]             Embedding norm dimension. 0: feat-wise, 1: node-wise, None: disable
    # Model configuration
    --model MODEL               Model class name
    --conv CONV                 Conv class name
    --num_hops NUM_HOPS         Number of conv hops
    --in_layers IN_LAYERS       Number of MLP layers before conv
    --out_layers OUT_LAYERS     Number of MLP layers after conv
    --hidden_channels HIDDEN    Number of hidden width
    --dropout_lin  DP_LIN       Dropout rate for linear
    --dropout_conv DP_CONV      Dropout rate for conv
    # Training configuration
    --epoch EPOCH               Number of epochs
    --patience PATIENCE         Patience epoch for early stopping
    --period PERIOD             Periodic saving epoch interval
    --batch BATCH               Batch size
    --lr_lin  LR_LIN            Learning rate for linear
    --lr_conv LR_CONV           Learning rate for conv
    --wd_lin  WD_LIN            Weight decay for linear
    --wd_conv WD_CONV           Weight decay for conv
    # Model-specific
    --theta_scheme THETA_SCHEME Filter name
    --theta_param THETA_PARAM   Hyperparameter for filter
    --combine {sum,sum_weighted,cat}
                                How to combine different channels of convs
    # Conv-specific
    --alpha ALPHA               Decay factor
    --beta BETA                 Scaling factor
    # Test flags
    --test_deg                  Call TrnFullbatch.test_deg()

Add New Experiment Dataset

In benchmark/trainer/load_data.py, append the SingleGraphLoader._resolve_import() method to include new datasets under respective protocols.

Add New Spectral Filter

New spectral filters to pyg_spectral/nn/conv/ can be easily implemented by only three steps, then enjoys a range of model architectures, analysis utilities, and training schemes.

Step 1: Define propagation matrix

The base class BaseMP provides essential methods for building spectral filters. We can define a new filter class SkipConv by inheriting from it:

from torch import Tensor
from pyg_spectral.nn.conv.base_mp import BaseMP

class SkipConv(BaseMP):
    def __init__(self, num_hops, hop, cached, **kwargs):
        kwargs['propagate_mat'] = 'A-I'
        super(SkipConv, self).__init__(num_hops, hop, cached, **kwargs)

The propagation matrix is specified by the propagate_mat argument as a string. Each matrix can be the normalized adjacency matrix (A) or the normalized Laplacian matrix (L), with optional diagonal scaling, where the scaling factor can either be a number or an attribute name of the class. Multiple propagation matrices can be combined by ,. Valid examples: A, L-2*I, L,A+I,L-alpha*I.

Step 2: Prepare representation matrix

Similar to PyG modules, our spectral filter class takes the graph attribute x and edge index edge_index as input. The _get_convolute_mat() method prepares the representation matrices used in recurrent computation as a dictionary:

    def _get_convolute_mat(self, x, edge_index):
        return {'x': x, 'x_1': x}

The above example overwrites the method for SkipConv, returning the input feature x and a placeholder x_1 for the representation in the previous hop.

Step 3: Derive recurrent forward

The _forward() method implements recurrent computation of the filter. Its input/output is a dictionary combining the propagation matrices defined by propagate_mat and the representation matrices prepared by _get_convolute_mat().

    def _forward(self, x, x_1, prop):
        if self.hop == 0:
            # No propagation for k=0
            return {'x': x, 'x_1': x, 'prop': prop}

        h = self.propagate(prop, x=x)
        h = h + x_1
        return {'x': h, 'x_1': x, 'prop': prop}

Similar to PyG modules, the propagate() method conducts graph propagation by the given matrices. The above example corresponds to the graph propagation with a skip connection to the previous representation: $H^{(k)} = (A-I)H^{(k-1)} + H^{(k-2)}$.

Build the model!

Now the SkipConv filter is properly defined. The following snippet use the DecoupledVar model composing 10 hops of SkipConv filters, which can be used as a normal PyTorch model:

from pyg_spectral.nn.models import DecoupledVar

model = DecoupledVar(conv='SkipConv', num_hops=10, in_channels=x.size(1), hidden_channels=x.size(1), out_channels=x.size(1))
out = model(x, edge_index)

Framework Arrangement

Covered Models

Category Model
Fixed Filter GCN, SGC, gfNN, GZoom, S²GC, GLP, APPNP, GCNII, GDC, DGC, AGP, GRAND+
Variable Filter GIN, AKGNN, DAGNN, GPRGNN, ARMAGNN, ChebNet, ChebNetII, HornerGCN / ClenshawGCN, BernNet, LegendreNet, JacobiConv, FavardGNN / OptBasisGNN
Filter Bank AdaGNN, FBGNN, ACMGNN, FAGCN, G²CN, GNN-LF/HF, FiGURe

Covered Datasets

Source Graph
PyG cora, citeseer, pubmed, flickr, actor
OGB ogbn-arxiv, ogbn-mag, ogbn-products
LINKX penn94, arxiv-year, genius, twitch-gamer, snap-patients, pokec, wiki
Yandex chameleon, squirrel, roman-empire, minesweeper, amazon-ratings, questions, tolokers

Code Structure

  • benchmark/: codes for benchmark experiments.
  • pyg_spectral/: core codes for spectral GNNs designs, arranged in PyG structure.
  • log/: experiment log files and parameter search results.
  • data/: raw and processed datasets arranged following different protocols.

Code structure of this framework and relation to PyG.

Roadmap

  • Support C++ propagation backend with efficient algorithms.
    • Unifews
    • SGC
    • GBP, AGP
  • Support more transformation operations.
    • Generalize ACMGNN
  • Support iterative eigen-decomposition for full-spectrum spectral filters.
    • Jacobi method
    • Lanczos method

Misc

  • This project is licensed under the MIT LICENSE.
  • Please export CITATION by using "Cite this repository" in the right sidebar.

Footnotes

  1. Please refer to the official guide if a specific CUDA version is required for PyTorch.