train_cv_multi_gpu.py

import dgl
import numpy as np
import torch as th
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.multiprocessing as mp
import dgl.function as fn
import dgl.nn.pytorch as dglnn
import time
import argparse
import tqdm
import traceback
from _thread import start_new_thread
from functools import wraps
from dgl.data import RedditDataset
from torch.utils.data import DataLoader
from torch.nn.parallel import DistributedDataParallel

class SAGEConvWithCV(nn.Module):
    def __init__(self, in_feats, out_feats, activation):
        super().__init__()
        self.W = nn.Linear(in_feats * 2, out_feats)
        self.activation = activation
        self.reset_parameters()

    def reset_parameters(self):
        gain = nn.init.calculate_gain('relu')
        nn.init.xavier_uniform_(self.W.weight, gain=gain)
        nn.init.constant_(self.W.bias, 0)

    def forward(self, block, H, HBar=None):
        if self.training:
            with block.local_scope():
                H_src, H_dst = H
                HBar_src, agg_HBar_dst = HBar
                block.dstdata['agg_hbar'] = agg_HBar_dst
                block.srcdata['hdelta'] = H_src - HBar_src
                block.update_all(fn.copy_u('hdelta', 'm'), fn.mean('m', 'hdelta_new'))
                h_neigh = block.dstdata['agg_hbar'] + block.dstdata['hdelta_new']
                h = self.W(th.cat([H_dst, h_neigh], 1))
                if self.activation is not None:
                    h = self.activation(h)
                return h
        else:
            with block.local_scope():
                H_src, H_dst = H
                block.srcdata['h'] = H_src
                block.update_all(fn.copy_u('h', 'm'), fn.mean('m', 'h_new'))
                h_neigh = block.dstdata['h_new']
                h = self.W(th.cat([H_dst, h_neigh], 1))
                if self.activation is not None:
                    h = self.activation(h)
                return h

class SAGE(nn.Module):
    def __init__(self,
                 in_feats,
                 n_hidden,
                 n_classes,
                 n_layers,
                 activation):
        super().__init__()
        self.n_layers = n_layers
        self.n_hidden = n_hidden
        self.n_classes = n_classes
        self.layers = nn.ModuleList()
        self.layers.append(SAGEConvWithCV(in_feats, n_hidden, activation))
        for i in range(1, n_layers - 1):
            self.layers.append(SAGEConvWithCV(n_hidden, n_hidden, activation))
        self.layers.append(SAGEConvWithCV(n_hidden, n_classes, None))

    def forward(self, blocks):
        h = blocks[0].srcdata['features']
        updates = []
        for layer, block in zip(self.layers, blocks):
            # We need to first copy the representation of nodes on the RHS from the
            # appropriate nodes on the LHS.
            # Note that the shape of h is (num_nodes_LHS, D) and the shape of h_dst
            # would be (num_nodes_RHS, D)
            h_dst = h[:block.number_of_dst_nodes()]
            hbar_src = block.srcdata['hist']
            agg_hbar_dst = block.dstdata['agg_hist']
            # Then we compute the updated representation on the RHS.
            # The shape of h now becomes (num_nodes_RHS, D)
            h = layer(block, (h, h_dst), (hbar_src, agg_hbar_dst))
            block.dstdata['h_new'] = h
        return h

    def inference(self, g, x, batch_size, device):
        """
        Inference with the GraphSAGE model on full neighbors (i.e. without neighbor sampling).
        g : the entire graph.
        x : the input of entire node set.

        The inference code is written in a fashion that it could handle any number of nodes and
        layers.
        """
        # During inference with sampling, multi-layer blocks are very inefficient because
        # lots of computations in the first few layers are repeated.
        # Therefore, we compute the representation of all nodes layer by layer.  The nodes
        # on each layer are of course splitted in batches.
        # TODO: can we standardize this?
        nodes = th.arange(g.number_of_nodes())
        ys = []
        for l, layer in enumerate(self.layers):
            y = th.zeros(g.number_of_nodes(), self.n_hidden if l != len(self.layers) - 1 else self.n_classes)

            for start in range(0, len(nodes), batch_size):
                end = start + batch_size
                batch_nodes = nodes[start:end]
                block = dgl.to_block(dgl.in_subgraph(g, batch_nodes), batch_nodes)
                induced_nodes = block.srcdata[dgl.NID]

                h = x[induced_nodes].to(device)
                h_dst = h[:block.number_of_dst_nodes()]
                h = layer(block, (h, h_dst))

                y[start:end] = h.cpu()

            ys.append(y)
            x = y
        return y, ys



class NeighborSampler(object):
    def __init__(self, g, fanouts):
        self.g = g
        self.fanouts = fanouts

    def sample_blocks(self, seeds):
        seeds = th.LongTensor(seeds)
        blocks = []
        hist_blocks = []
        for fanout in self.fanouts:
            # For each seed node, sample ``fanout`` neighbors.
            frontier = dgl.sampling.sample_neighbors(self.g, seeds, fanout)
            # For history aggregation we sample all neighbors.
            hist_frontier = dgl.in_subgraph(self.g, seeds)
            # Then we compact the frontier into a bipartite graph for message passing.
            block = dgl.to_block(frontier, seeds)
            hist_block = dgl.to_block(hist_frontier, seeds)
            # Obtain the seed nodes for next layer.
            seeds = block.srcdata[dgl.NID]

            blocks.insert(0, block)
            hist_blocks.insert(0, hist_block)
        return blocks, hist_blocks

# According to https://github.com/pytorch/pytorch/issues/17199, this decorator
# is necessary to make fork() and openmp work together.
#
# TODO: confirm if this is necessary for MXNet and Tensorflow.  If so, we need
# to standardize worker process creation since our operators are implemented with
# OpenMP.
def thread_wrapped_func(func):
    @wraps(func)
    def decorated_function(*args, **kwargs):
        queue = mp.Queue()
        def _queue_result():
            exception, trace, res = None, None, None
            try:
                res = func(*args, **kwargs)
            except Exception as e:
                exception = e
                trace = traceback.format_exc()
            queue.put((res, exception, trace))

        start_new_thread(_queue_result, ())
        result, exception, trace = queue.get()
        if exception is None:
            return result
        else:
            assert isinstance(exception, Exception)
            raise exception.__class__(trace)
    return decorated_function

def prepare_mp(g):
    """
    Explicitly materialize the CSR, CSC and COO representation of the given graph
    so that they could be shared via copy-on-write to sampler workers and GPU
    trainers.

    This is a workaround before full shared memory support on heterogeneous graphs.
    """
    g.in_degree(0)
    g.out_degree(0)
    g.find_edges([0])

def compute_acc(pred, labels):
    """
    Compute the accuracy of prediction given the labels.
    """
    return (th.argmax(pred, dim=1) == labels).float().sum() / len(pred)

def evaluate(model, g, labels, val_mask, batch_size, device):
    """
    Evaluate the model on the validation set specified by ``val_mask``.
    g : The entire graph.
    inputs : The features of all the nodes.
    labels : The labels of all the nodes.
    val_mask : A 0-1 mask indicating which nodes do we actually compute the accuracy for.
    batch_size : Number of nodes to compute at the same time.
    device : The GPU device to evaluate on.
    """
    model.eval()
    with th.no_grad():
        inputs = g.ndata['features']
        pred, _ = model.inference(g, inputs, batch_size, device)
    model.train()
    return compute_acc(pred[val_mask], labels[val_mask])

def load_subtensor(g, labels, blocks, hist_blocks, dev_id, aggregation_on_device=False):
    """
    Copys features and labels of a set of nodes onto GPU.
    """
    blocks[0].srcdata['features'] = g.ndata['features'][blocks[0].srcdata[dgl.NID]].to(dev_id)
    blocks[-1].dstdata['label'] = labels[blocks[-1].dstdata[dgl.NID]].to(dev_id)
    for i, (block, hist_block) in enumerate(zip(blocks, hist_blocks)):
        hist_col = 'features' if i == 0 else 'hist_%d' % i
        block.srcdata['hist'] = g.ndata[hist_col][block.srcdata[dgl.NID]].to(dev_id)

        # Aggregate history
        hist_block.srcdata['hist'] = g.ndata[hist_col][hist_block.srcdata[dgl.NID]]
        if aggregation_on_device:
            hist_block.srcdata['hist'] = hist_block.srcdata['hist'].to(dev_id)
        hist_block.update_all(fn.copy_u('hist', 'm'), fn.mean('m', 'agg_hist'))
        block.dstdata['agg_hist'] = hist_block.dstdata['agg_hist']
        if not aggregation_on_device:
            block.dstdata['agg_hist'] = block.dstdata['agg_hist'].to(dev_id)

def init_history(g, model, dev_id):
    with th.no_grad():
        history = model.inference(g, g.ndata['features'], 1000, dev_id)[1]
        for layer in range(args.num_layers + 1):
            if layer > 0:
                hist_col = 'hist_%d' % layer
                g.ndata['hist_%d' % layer] = history[layer - 1]

def update_history(g, blocks):
    with th.no_grad():
        for i, block in enumerate(blocks):
            ids = block.dstdata[dgl.NID]
            hist_col = 'hist_%d' % (i + 1)

            h_new = block.dstdata['h_new'].cpu()
            g.ndata[hist_col][ids] = h_new

@thread_wrapped_func
def run(proc_id, n_gpus, args, devices, data):
    dropout = 0.2

    dev_id = devices[proc_id]
    if n_gpus > 1:
        dist_init_method = 'tcp://{master_ip}:{master_port}'.format(
            master_ip='127.0.0.1', master_port='12345')
        world_size = n_gpus
        th.distributed.init_process_group(backend="nccl",
                                          init_method=dist_init_method,
                                          world_size=world_size,
                                          rank=proc_id)
    th.cuda.set_device(dev_id)

    # Unpack data
    train_mask, val_mask, in_feats, labels, n_classes, g = data
    train_nid = th.LongTensor(np.nonzero(train_mask)[0])
    val_nid = th.LongTensor(np.nonzero(val_mask)[0])
    train_mask = th.BoolTensor(train_mask)
    val_mask = th.BoolTensor(val_mask)

    # Split train_nid
    train_nid = th.split(train_nid, len(train_nid) // n_gpus)[proc_id]

    # Create sampler
    sampler = NeighborSampler(g, [int(_) for _ in args.fan_out.split(',')])

    # Create PyTorch DataLoader for constructing blocks
    dataloader = DataLoader(
        dataset=train_nid.numpy(),
        batch_size=args.batch_size,
        collate_fn=sampler.sample_blocks,
        shuffle=True,
        drop_last=False,
        num_workers=args.num_workers_per_gpu)

    # Define model
    model = SAGE(in_feats, args.num_hidden, n_classes, args.num_layers, F.relu)

    # Move the model to GPU and define optimizer
    model = model.to(dev_id)
    if n_gpus > 1:
        model = DistributedDataParallel(model, device_ids=[dev_id], output_device=dev_id)
    loss_fcn = nn.CrossEntropyLoss()
    loss_fcn = loss_fcn.to(dev_id)
    optimizer = optim.Adam(model.parameters(), lr=args.lr)

    # Compute history tensor and their aggregation before training on CPU
    model.eval()
    if n_gpus > 1:
        init_history(g, model.module, dev_id)
        th.distributed.barrier()
    else:
        init_history(g, model, dev_id)
    model.train()

    # Training loop
    avg = 0
    iter_tput = []
    for epoch in range(args.num_epochs):
        tic = time.time()
        model.train()
        for step, (blocks, hist_blocks) in enumerate(dataloader):
            if proc_id == 0:
                tic_step = time.time()

            # The nodes for input lies at the LHS side of the first block.
            # The nodes for output lies at the RHS side of the last block.
            input_nodes = blocks[0].srcdata[dgl.NID]
            seeds = blocks[-1].dstdata[dgl.NID]

            load_subtensor(g, labels, blocks, hist_blocks, dev_id, True)

            # forward
            batch_pred = model(blocks)
            # update history
            update_history(g, blocks)
            # compute loss
            batch_labels = blocks[-1].dstdata['label']
            loss = loss_fcn(batch_pred, batch_labels)
            # backward
            optimizer.zero_grad()
            loss.backward()
            if n_gpus > 1:
                for param in model.parameters():
                    if param.requires_grad and param.grad is not None:
                        th.distributed.all_reduce(param.grad.data,
                                                  op=th.distributed.ReduceOp.SUM)
                        param.grad.data /= n_gpus
            optimizer.step()
            if proc_id == 0:
                iter_tput.append(len(seeds) * n_gpus / (time.time() - tic_step))
            if step % args.log_every == 0 and proc_id == 0:
                acc = compute_acc(batch_pred, batch_labels)
                print('Epoch {:05d} | Step {:05d} | Loss {:.4f} | Train Acc {:.4f} | Speed (samples/sec) {:.4f}'.format(
                    epoch, step, loss.item(), acc.item(), np.mean(iter_tput[3:])))

        if n_gpus > 1:
            th.distributed.barrier()

        toc = time.time()
        if proc_id == 0:
            print('Epoch Time(s): {:.4f}'.format(toc - tic))
            if epoch >= 5:
                avg += toc - tic
            if epoch % args.eval_every == 0 and epoch != 0:
                model.eval()
                eval_acc = evaluate(
                    model if n_gpus == 1 else model.module, g, labels, val_nid, args.val_batch_size, dev_id)
                print('Eval Acc {:.4f}'.format(eval_acc))

    if n_gpus > 1:
        th.distributed.barrier()
    if proc_id == 0:
        print('Avg epoch time: {}'.format(avg / (epoch - 4)))

if __name__ == '__main__':
    argparser = argparse.ArgumentParser("multi-gpu training")
    argparser.add_argument('--gpu', type=str, default='0')
    argparser.add_argument('--num-epochs', type=int, default=20)
    argparser.add_argument('--num-hidden', type=int, default=16)
    argparser.add_argument('--num-layers', type=int, default=2)
    argparser.add_argument('--fan-out', type=str, default='1,1')
    argparser.add_argument('--batch-size', type=int, default=1000)
    argparser.add_argument('--val-batch-size', type=int, default=1000)
    argparser.add_argument('--log-every', type=int, default=20)
    argparser.add_argument('--eval-every', type=int, default=5)
    argparser.add_argument('--lr', type=float, default=0.003)
    argparser.add_argument('--num-workers-per-gpu', type=int, default=0)
    args = argparser.parse_args()
    
    devices = list(map(int, args.gpu.split(',')))
    n_gpus = len(devices)

    # load reddit data
    data = RedditDataset(self_loop=True)
    train_mask = data.train_mask
    val_mask = data.val_mask
    features = th.Tensor(data.features)
    in_feats = features.shape[1]
    labels = th.LongTensor(data.labels)
    n_classes = data.num_labels
    # Construct graph
    g = dgl.graph(data.graph.all_edges())
    g.ndata['features'] = features
    prepare_mp(g)
    # Pack data
    data = train_mask, val_mask, in_feats, labels, n_classes, g

    if n_gpus == 1:
        run(0, n_gpus, args, devices, data)
    else:
        procs = []
        for proc_id in range(n_gpus):
            p = mp.Process(target=run, args=(proc_id, n_gpus, args, devices, data))
            p.start()
            procs.append(p)
        for p in procs:
            p.join()