Single Machine Multi-GPU Minibatch Graph Classification

In this tutorial, you will learn how to use multiple GPUs in training a graph neural network (GNN) for graph classification. This tutorial assumes knowledge in GNNs for graph classification and we recommend you to check Training a GNN for Graph Classification otherwise.

(Time estimate: 8 minutes)

To use a single GPU in training a GNN, we need to put the model, graph(s), and other tensors (e.g. labels) on the same GPU:

import torch

# Use the first GPU
device = torch.device("cuda:0")
model = model.to(device)
graph = graph.to(device)
labels = labels.to(device)

The node and edge features in the graphs, if any, will also be on the GPU. After that, the forward computation, backward computation and parameter update will take place on the GPU. For graph classification, this repeats for each minibatch gradient descent.

Using multiple GPUs allows performing more computation per unit of time. It is like having a team work together, where each GPU is a team member. We need to distribute the computation workload across GPUs and let them synchronize the efforts regularly. PyTorch provides convenient APIs for this task with multiple processes, one per GPU, and we can use them in conjunction with DGL.

Intuitively, we can distribute the workload along the dimension of data. This allows multiple GPUs to perform the forward and backward computation of multiple gradient descents in parallel. To distribute a dataset across multiple GPUs, we need to partition it into multiple mutually exclusive subsets of a similar size, one per GPU. We need to repeat the random partition every epoch to guarantee randomness. We can use GraphDataLoader(), which wraps some PyTorch APIs and does the job for graph classification in data loading.

Once all GPUs have finished the backward computation for its minibatch, we need to synchronize the model parameter update across them. Specifically, this involves collecting gradients from all GPUs, averaging them and updating the model parameters on each GPU. We can wrap a PyTorch model with DistributedDataParallel() so that the model parameter update will invoke gradient synchronization first under the hood.

That’s the core behind this tutorial. We will explore it more in detail with a complete example below.

Note

See this tutorial from PyTorch for general multi-GPU training with DistributedDataParallel.

Distributed Process Group Initialization

For communication between multiple processes in multi-gpu training, we need to start the distributed backend at the beginning of each process. We use world_size to refer to the number of processes and rank to refer to the process ID, which should be an integer from 0 to world_size - 1.

import torch.distributed as dist

def init_process_group(world_size, rank):
    dist.init_process_group(
        backend='nccl',
        init_method='tcp://127.0.0.1:12345',
        world_size=world_size,
        rank=rank)

Data Loader Preparation

We split the dataset into training, validation and test subsets. In dataset splitting, we need to use a same random seed across processes to ensure a same split. We follow the common practice to train with multiple GPUs and evaluate with a single GPU, thus only set use_ddp to True in the GraphDataLoader() for the training set, where ddp stands for DistributedDataParallel().

from dgl.data import split_dataset
from dgl.dataloading import GraphDataLoader

def get_dataloaders(dataset, seed, batch_size=32):
    # Use a 80:10:10 train-val-test split
    train_set, val_set, test_set = split_dataset(dataset,
                                                 frac_list=[0.8, 0.1, 0.1],
                                                 shuffle=True,
                                                 random_state=seed)
    train_loader = GraphDataLoader(train_set, use_ddp=True, batch_size=batch_size, shuffle=True)
    val_loader = GraphDataLoader(val_set, batch_size=batch_size)
    test_loader = GraphDataLoader(test_set, batch_size=batch_size)

    return train_loader, val_loader, test_loader

Model Initialization

For this tutorial, we use a simplified Graph Isomorphism Network (GIN).

import torch.nn as nn
import torch.nn.functional as F
from dgl.nn.pytorch import GINConv, SumPooling

class GIN(nn.Module):
    def __init__(self, input_size=1, num_classes=2):
        super(GIN, self).__init__()

        self.conv1 = GINConv(nn.Linear(input_size, num_classes), aggregator_type='sum')
        self.conv2 = GINConv(nn.Linear(num_classes, num_classes), aggregator_type='sum')
        self.pool = SumPooling()

    def forward(self, g, feats):
        feats = self.conv1(g, feats)
        feats = F.relu(feats)
        feats = self.conv2(g, feats)

        return self.pool(g, feats)

To ensure same initial model parameters across processes, we need to set the same random seed before model initialization. Once we construct a model instance, we wrap it with DistributedDataParallel().

import torch
from torch.nn.parallel import DistributedDataParallel

def init_model(seed, device):
    torch.manual_seed(seed)
    model = GIN().to(device)
    model = DistributedDataParallel(model, device_ids=[device], output_device=device)

    return model

Main Function for Each Process

Define the model evaluation function as in the single-GPU setting.

def evaluate(model, dataloader, device):
    model.eval()

    total = 0
    total_correct = 0

    for bg, labels in dataloader:
        bg = bg.to(device)
        labels = labels.to(device)
        # Get input node features
        feats = bg.ndata.pop('attr')
        with torch.no_grad():
            pred = model(bg, feats)
        _, pred = torch.max(pred, 1)
        total += len(labels)
        total_correct += (pred == labels).sum().cpu().item()

    return 1.0 * total_correct / total

Define the main function for each process.

from torch.optim import Adam

def main(rank, world_size, dataset, seed=0):
    init_process_group(world_size, rank)
    # Assume the GPU ID to be the same as the process ID
    device = torch.device('cuda:{:d}'.format(rank))
    torch.cuda.set_device(device)

    model = init_model(seed, device)
    criterion = nn.CrossEntropyLoss()
    optimizer = Adam(model.parameters(), lr=0.01)

    train_loader, val_loader, test_loader = get_dataloaders(dataset,
                                                            seed)
    for epoch in range(5):
        model.train()
        # The line below ensures all processes use a different
        # random ordering in data loading for each epoch.
        train_loader.set_epoch(epoch)

        total_loss = 0
        for bg, labels in train_loader:
            bg = bg.to(device)
            labels = labels.to(device)
            feats = bg.ndata.pop('attr')
            pred = model(bg, feats)

            loss = criterion(pred, labels)
            total_loss += loss.cpu().item()
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
        loss = total_loss
        print('Loss: {:.4f}'.format(loss))

        val_acc = evaluate(model, val_loader, device)
        print('Val acc: {:.4f}'.format(val_acc))

    test_acc = evaluate(model, test_loader, device)
    print('Test acc: {:.4f}'.format(test_acc))
    dist.destroy_process_group()

Finally we load the dataset and launch the processes.

Note

You will need to use dgl.multiprocessing instead of the Python multiprocessing package. dgl.multiprocessing is identical to Python’s built-in multiprocessing except that it handles the subtleties between forking and multithreading in Python.

if __name__ == '__main__':
    import dgl.multiprocessing as mp

    from dgl.data import GINDataset

    num_gpus = 4
    procs = []
    dataset = GINDataset(name='IMDBBINARY', self_loop=False)
    for rank in range(num_gpus):
        p = mp.Process(target=main, args=(rank, num_gpus, dataset))
        p.start()
        procs.append(p)
    for p in procs:
        p.join()

# Thumbnail credits: DGL
# sphinx_gallery_thumbnail_path = '_static/blitz_5_graph_classification.png'