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

Pre-proceeding version

• shaDow-GNN paper accepted to NeurIPS'21 (updated title: Decoupling the Depth and Scope of Graph Neural Networks; camera ready coming soon)!
• We thank the Pytorch Geometric team for including the shaDow k-hop sampler in their library!
• Major updates to be released this month (Nov 2021).

We have implemented shaDow-GNN as a general and powerful pipeline for graph representation learning. The training of shaDow-GNN can be abstracted as three major steps:

The preprocessing steps may

• Smooth the node features
• Augment the node features with training labels.

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 that preprocessing is turned off in some cases (e.g., for results on ogbn-papers100M).

shaDow-GNN now supports six different backbone architectures, including: GCN, GraphSAGE, GAT, GIN, JK-Net and SGC.

In addition, we support the following architectural extensions:

• Samplers: k-hop, PPR, and the ensemble of them
• Subgraph pooling / READOUT: CenterPooling, SortPooling, MaxPooling, MeanPooling, SumPooling

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. In the paper, our discussion is based on the "end-to-end" ensemble.

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 sampling, where we construct a local subgraph for each target node independently. This part is parallelized on CPU by C++ and OpenMP.
• Forward / backward propagation of the GNN model. This part is parallelized on GPU via PyTorch.

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

If you have more powerful machines, you can simply scale up the performance by increasing the batch size. In our experiments, we have tested on GPUs ranging from NVIDIA GeForce GTX 1650 (4GB) to NVIDIA GeForce RTX 3090 (24GB).

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.

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).

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

• For the three 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 three other graphs (i.e., Flickr, Reddit, Yelp), these are listed in the officially 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 ./shaDow/data_converter.py. It is automatically invoked when you run training for the first time.

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 ./shaDow/para_sampler/ParallelSampler.* contains the C++ code for the PPR and k-hop samplers. The ./shaDow/para_sampler/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. May be installed by conda install -c anaconda cmake)
• ninja (our version is 1.10.2. May be installed by conda install -c conda-forge ninja)
• pybind11 (our version is 2.6.2. May 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

bash install.sh

On Windows machine, you could instead execute .\install.bat.

NOTE: if the above does not work, we provide an alternative script to compile the C++ sampler. See the Troubleshooting section for details.

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

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: You should be able to run the training now. 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 the GNN architecture, sampler parameters and other hyperparameters. The name of the graph should correspond to the sub-directory name under ./data/. E.g., the graphs used in the papers correspond to flickr, reddit, yelp, arxiv, products, papers100M (we use all lowercase and omit the ogbn- prefix).

Step 5 Check the logs of the training. We use the following protocol for logging. Our principle is that the logs enable you to completely reproduce your own 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 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.

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 for 10 runs.

shaDow-SAGE

Run the following:

python -m shaDow.main --configs config_train/papers100M/sage_5_ppr_ogb.yml --dataset papers100M --gpu 0

shaDow-GAT

Run the following:

python -m shaDow.main --configs config_train/papers100M/gat_5_ppr_ogb.yml --dataset papers100M --gpu 0

We don't fix any random seeds (for C++ sampler, the seed is set by the timestamp of launching the training. See std::srand(std::time(0)) of shaDow/para_samplers/ParallelSampler.h). Over 10 runs, we get 0.6653 test accuracy for the shaDow-GAT model.

According to the OGB leaderboard instruction, we need to repeat the same configuration for 10 times, and report the mean and std of accuracy. We provide a wrapper script for this purpose: ./scripts/train_multiple_runs.py.

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.

For example, to repeat the configuration for 10 times:

python scripts/train_multiple_runs.py --dataset papers100M --configs config_train/papers100M/gat_5_ppr_ogb.yml --gpu 0 --repetition 10

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

NOTE: due to the large graph size, each run on ogbn-papers100M takes quite sometime to converge. It may make more sense to manually launch the training 10 times and manually calculate the mean and std based on the log files.

This may be an issue specific to your machine. You can file an issue or contact [email protected]. Alternatively, you can try another compilation script we provided after manually build pybind11. See this installation instruction and do the following:

• Download pybind11 from the official GitHub repo. Enter the pybind11 root dir.
• Install pytest by, e.g., pip install -U pytest
• mkdir build
• cd build
• cmake ..
  • Note: if cmake is not installed, you can install it via Anaconda: conda install -c anaconda cmake
• make check -j 4 (Windows machine does not seem to be well supported)
• Install pybind11 package: python -m pip install pybind11

Then go back to the shaDow-GNN root directory and compile the C++ sampler by

./compile.sh

shaDow_GNN is released under an MIT license. Find out more about it here.

@article{shadow_GNN,
  author    = {Hanqing Zeng and
               Muhan Zhang and
               Yinglong Xia and
               Ajitesh Srivastava and
               Andrey Malevich and
               Rajgopal Kannan and
               Viktor K. Prasanna and
               Long Jin and
               Ren Chen},
  title     = {Deep Graph Neural Networks with Shallow Subgraph Samplers},
  journal   = {CoRR},
  volume    = {abs/2012.01380},
  year      = {2020},
  url       = {https://arxiv.org/abs/2012.01380},
  eprinttype = {arXiv},
  eprint    = {2012.01380},
  timestamp = {Fri, 04 Dec 2020 12:07:23 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2012-01380.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}