Skip to content

NeurIPS 2021: Improve the GNN expressivity and scalability by decoupling the depth and receptive field of state-of-the-art GNN architectures

License

Notifications You must be signed in to change notification settings

facebookresearch/shaDow_GNN

Repository files navigation

Decoupling the Depth and Scope of Graph Neural Networks

Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Andrey Malevich, Rajgopal Kannan, Viktor Prasanna, Long Jin, Ren Chen

Contact: Hanqing Zeng (zengh@usc.edu)

Latest version of the paper

(Note: There is an old version named "Deep Graph Neural Networks with Shallow Subgraph Samplers". Please only refer to the new version and disgard the old one. )

News

  • Major updates (code refactoring; link prediction support; all training configs) released in Jan 2022.
  • We thank the DGL team for including the shaDow k-hop sampler in their library.
  • shaDow-GNN paper accepted to NeurIPS'21!
  • We thank the Pytorch Geometric team for including the shaDow k-hop sampler in their library.

Overview

We propose a design principle of "decoupling the depth and scope" when constructing GNN models. This is a simple way to surpass 1-WL, overcome oversmoothing and avoid neighborhood explosion at the same time.

We call the practical implementation of our design principle as shaDow-GNN (Deep GNNs on shallow subgraphs).

This repo implements:

  • 6 backbone message passing layers (GCN, GraphSAGE, GIN, GAT, JK-Net, SGC)
  • 4 pooling layers (sort, max, mean, sum)
  • 4 subgraph extractors / sampler (IID node, k-hop, PPR, stochastic PPR)
  • Pre-processing (feature smoothening; label propagation)
  • Post-processing (C&S)
  • Subgraph ensemble (either during training or post-processing)

This repo supports:

  • Inductive node classification (Flickr, Reddit, Yelp)
  • Transductive node classification (ogbn-arxiv, ogbn-products, ogbn-papers100M)
  • Link prediction (ogbl-collab)

The training pipeline of shaDow-GNN can be abstracted as three major steps:

Preprocessing (optional)

Expand to see details...

The preprocessing steps may augment the input node features with

  • Smoothened node features
  • Ground-truth labels in the training set

The first point is similar to what SGC and SIGN did (it's just we convert the original algorithm into the shaDow version). The second point is inspired by the methods on the OGB leaderboard (only applicable under the transductive setting).

Note: preprocessing is turned off in all experiment in our main paper.

Training

All shaDow-GNN are trained in the minibatch fashion. For each training batch, we first perform subgraph extraction, and then build a multi-layer GNN on the subgraph to perform message passing.

For any nodes u and v in the same batch, we treat the two subgraphs as completely isolated. i.e., when a node w of the original graph is included in both subgraphs, we rename w of u's subgraph as w1 and w of v's subgraph as w2 so that the two subgraphs don't talk to each other. See _node_induced_subgraph() function in para_graph_sampler/graph_engine/backend/ParallelSampler.cpp.

Note: unlike other graph sampling based methods, shaDow-GNN allows much smaller batch size (can be as small as 1) since the subgraph degree of shaDow-GNN does not drop with batch size. This property makes shaDow-GNN easily portable on GPUs of limited memory.

Postprocessing (optional)

Expand to see details...

After the training is finished, we can reload the stored checkpoint to perform the following post-processing steps: * *C&S* (transductive only): we borrow the DGL implementation of C&S to perform smoothening of the predictions generated by shaDow-GNN. * *Ensemble*: Ensemble can be done either in an "end-to-end" fashion during the above training step, or as a postprocessing step.

Hardware requirements

Due to its flexibility in minibatching, shaDow-GNN requires the minimum hardware for training and inference computation. Most of our experiments can be run on a desktop machine. Even the largest graph of 111 million nodes can be trained on a low-end server.

The main computation operations include:

  • Subgraph extraction / sampling: parallelized on CPU by C++ and OpenMP.
  • GNN model propagation: accelerated on GPU via PyTorch.

We summarize the recommended minimum hardware spec for the three OGB graphs:

Graph Num. nodes CPU cores CPU RAM GPU memory
ogbn-arxiv 0.2M 4 8GB 4GB
ogbn-products 2.4M 4 32GB 4GB
ogbn-papers100M 111.1M 4 128GB 4GB

Data format

When you run shaDow-GNN for the first time, we will convert the graph data from the OGB or GraphSAINT format into the shaDow-GNN format. The converted data files are (by default) stored in the ./data/<graph_name> directory.

NOTE: the initial data conversion may take a while for large graphs (e.g., for ogbn-papers100M). Please be patient.

General shaDow format

Expand to see details if you want to prepare your own dataset

We briefly describe the shaDow data format. You should not need to worry about the details unless you want to prepare your own dataset. Each graph is defined by the following files:

  • adj_full_raw.npz / adj_full_raw.npy: The adjacency matrix of the full graph (consisting of all the train / valid / test nodes). It can either be a *.npz file of type scipy.sparse.csr_matrix, or a *.npy file containing the dictionary {'indptr': numpy.ndarray, 'indices': numpy.ndarray, 'data': numpy.ndarray}.
  • adj_train_raw.npz / adj_train_raw.npy: The adjacency matrix induced by all training nodes (ONLY used in inductive learning).
  • label_full.npy: The numpy.ndarray representing the labels of all the train / valid / test nodes. If this matrix is 2D, then a row is a one-hot encoding of the label(s) of a node. If this is 1D, then an element is the label index of a node. In any case, the first dimension equals the total number of nodes.
  • feat_full.npy: The numpy.ndarray representing the node features. The first dimension of the matrix equals the total number of nodes.
  • split.npy: The file stores a dictionary representing the train / valid / test splitting. The keys are train / valid / test. The values are numpy array of the node indices for the corresponding split.
  • (Optional) adj_full_undirected.npy: This is a cache file storing the graph after converting adj_full_raw into undirected (e.g., the raw graph of ogbn-arxiv is directed).
  • (Optional) adj_train_undirected.npy: Similar as above. Converted from adj_train_raw into undirected.
  • (Optional) cpp/adj_<full|train>_<indices|indptr|data>.bin: These are the cache files for the C++ sampler. We store the corresponding *.npy / *.npz files as binary files so that the C++ sampler can directly load the graph without going through the layer of PyBind11 (see below). For gigantic graphs such as ogbn-papers100M, the conversion from numpy.ndarray to C++ vector seems to be slow (maybe an issue of PyBind11).
  • (Optional) ppr_float/<neighs|scores>_<transductive|inductive>_<ppr params>.bin: These are the cache files for the C++ PPR sampler. We store the PPR values and node indices for the close neighbors of each target as the external binary files. Therefore, we do not need to run PPR multiple times when we perform parameter tuning (even through running PPR from scratch is still much cheaper than the model training).

Graphs tested

To train shaDow-GNN on the 7 graphs evaluated in the paper:

  • For the 4 OGB graphs (i.e., ogbn-arxiv, ogbn-products, ogbn-papers100M), you don't need to manually download anything. Just execute the training command (see below).
  • For the 3 other graphs (i.e., Flickr, Reddit, Yelp), the source data files are listed in the official GraphSAINT repo. Please manually download from the link provided by GraphSAINT, and place all the downloaded files under the ./data/saint/<graph name>/ directory.
    • E.g., for Flickr, the directory should look something like (note the lower case for graph name)
data/
└───saint/
    └───flickr/
        └───adj_full.npz
            class_map.json
            ...

The script for converting from OGB / SAINT into shaDow format is ./para_graph_sampler/graph_engine/frontend/data_converter.py. It is automatically invoked when you run training for the first time.

Build and Run

Clone the repo by (you need the --recursive flag to download pybind11 as submodule):

git clone <URL FOR THIS REPO> --recursive

Step 0: Make sure you create a virtual environment with Python 3.8 (lower version of python may not work. The version we use is 3.8.5).

Step 1: We need PyBind11 to link the C++ based sampler with the PyTorch based trainer. The ./para_graph_sampler/graph_engine/backend/ParallelSampler.* contain the C++ code for the PPR and k-hop samplers. The ./para_graph_sampler/graph_engine/backend/pybind11/ directory contains a copy of PyBind11.

Before training, we need to build the C++ sampler as a python package, so that it can be directly imported by the PyTorch trainer (just like we import any other python module). To do so, you need to install the following:

  • cmake (our version is 3.18.2. Can be installed by conda install -c anaconda cmake)
  • ninja (our version is 1.10.2. Can be installed by conda install -c conda-forge ninja)
  • pybind11 (our version is 2.6.2. Can be installed by pip install pybind11)
  • OpenMP: normally openmp should already be included in the C++ compiler. If not, you may need to install it manually based on your C++ compiler version.

Then build the sampler. Run the following in your terminal

cd para_graph_sampler
bash install.sh
cd ..

On Windows machine, you could instead replace the bash install.sh command by .\install.bat.

Step 2: Install all the other Python packages in your virtual environment.

  • pytorch==1.7.1 (CUDA 11)
  • Pytorch Geometric and its dependency packages (torch-scatter, torch-sparse, etc.)
    • Follow the official instructions (see the "Installation via Binaries" section)
    • We also explicitly use the torch_scatter functions to perform some graph operations for shaDow.
  • ogb>=1.2.4
  • dgl>=0.5.3 (only used by the postprocessing of C&S). Can be installed by pip or conda. See the official instruction
  • numpy>=1.19.2
  • scipy>=1.6.0
  • scikit-learn>=0.24.0
  • pyyaml>=5.4.1
  • argparse
  • tqdm

(Optional) Step 3: Record your system information. We use the CONFIG.yml file to keep track of the meta information of your hardware / software system. Copy CONFIG_TEMPLATE.yml and name it CONFIG.yml. Edit the fields based on your machine specs.

In most cases, the only thing you need to overwrite is the max_threads field. This is used to control the parallelism of the C++ sampler. You can also set it to -1 so that OpenMP will automatically decide the number of threads for you.

Step 4: Now you should be able to run the training / inference. In general, just type:

python -m shaDow.main --configs <your config *.yml file> --dataset <name of the graph> --gpu <index of the available GPU>

where the *.yml file specifies all the hyperparameters (e.g., GNN architecture, sampler, etc.). The name of the graph should correspond to the sub-directory name under ./data/ (we use all lowercase and omit the ogbn- or ogbl- prefix).

Step 5 Check the logs of the training. We use the following protocol for logging. Our principle is to enable complete reproductivity of the previous runs.

  • Each run gets its own subdirectory in the format of ./<log dir>/<data>/<running|done|crashed|killed>/<timestamp>-<githash>/..., where the subdirectory indicates the status of the run:
    • running/: the training is still in progress.
    • finished/: the training finishes normally. The logs will be moved from running/ to finished/.
    • killed/: the training is killed (e.g., by CTRL-C).
    • crashed/: the training crashes (e.g., bugs in the code, GPU / CPU out-of-memory, etc.).
  • In the subdirectory we should find the following files:
    • *.yml: a copy of the *.yml file to launch the training
    • epoch_<train|valid|test>.csv: CSV file logging the accuracy and loss of each epoch
    • final.csv: CSV file logging the final accuracy on the full train / valid / test sets.
    • pytorch checkpoint: the model weights and optimizer states.

Reproducing the paper results

We first describe the command for a single run. At the end of this section, we show the wrapper script for repeating the same configuration multiple times.

Table 1

The configs are under ./config_train/<dataset>/<vanilla|pool>/<arch>_<depth>_<sampler>.yml, where

  • dataset corresponds to the 5 graphs in Table 1: flickr, reddit, yelp, arxiv, products.
  • vanilla means no subgraph pooling is performed. So we take the target node embedding for classification and disgard embeddings for all other subgraph nodes.
  • pool means we add an extra subgraph pooling layer onto the vanilla arch. Table 1 only evaluated mean and max pooling. You can also try sort and sum.
  • arch in Table 1 is restricted to gcn, sage and gat. Figure 3 corresponds to sgc and Table 12 corresponds to gin.
  • depth in Table 1 is either 3 or 5.
  • sampler is chosen from ppr and khop.

Note: for ogbn-products, since its test set is especially large, you can skip evaluating test accuracy during training by additional flags. e.g.,

python -m shaDow.main --configs config_train/products/pool/gat_3_ppr.yml --dataset products --gpu <gpu idx> --log_test_convergence -1 --nocache test

Table 2

Run:

python -m shaDow.main --configs config_train/papers100M/leaderboard/gat_ppr.yml --dataset papers100M --gpu <gpu idx>

Table 3

Run:

python -m shaDow.main --configs config_train/collab/leaderboard/sage_ppr.yml --dataset collab --gpu <gpu idx>

Repeat the same configuration multiple times

Expand to see details

Table 1 results are all repeated 5 times. Table 2 and 3 results are repeated 10 times. All without fixing random seeds. In C++, not fixing random seed is achieved by `std::srand(std::time(0))` in `para_graph_sampler/graph_engine/backend/ParallelSampler.h`.

We also provide a wrapper script for repeat the training. See ./scripts/train_multiple_runs.py.

General command:

python scripts/train_multiple_runs.py --dataset <dataset> --configs <config yml> --gpu <gpu idx> --repetition 10

where all the command line arguments of train_multiple_runs.py are the same as the original training script (i.e., the shaDow.main module). The only additional flag is --repetition.

NOTE: the wrapper script uses python subprocess to launch multiple runs. There seems to be some issue on redirecting the print-out messages of the training subprocess. It may appear that the program stucks without any outputs. This is due to the buffering of output. However, the training should actually be running in the background. You can check the corresponding log files in the running/ directory to see the accuracy per epoch being updated.

License

shaDow-GNN is released under an MIT license. Find out more about it here.

Citation

NeurIPS 2021

@inproceedings{
    shaDow,
    title={Decoupling the Depth and Scope of Graph Neural Networks},
    author={Hanqing Zeng and Muhan Zhang and Yinglong Xia and Ajitesh Srivastava and Andrey Malevich and Rajgopal Kannan and Viktor Prasanna and Long Jin and Ren Chen},
    booktitle={Advances in Neural Information Processing Systems},
    editor={A. Beygelzimer and Y. Dauphin and P. Liang and J. Wortman Vaughan},
    year={2021},
    url={https://openreview.net/forum?id=d0MtHWY0NZ}
}

About

NeurIPS 2021: Improve the GNN expressivity and scalability by decoupling the depth and receptive field of state-of-the-art GNN architectures

Resources

License

Code of conduct

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages