def test_quiver_ogbnproducts()

in benchmarks/api/bench_feature.py [0:0]

• 34 lines of code
• 2 McCabe index (conditional complexity)

def test_quiver_ogbnproducts(split_ratio):
  import quiver
  cache_size = str(950 * split_ratio) + 'M'
  root = osp.join(osp.dirname(osp.dirname(osp.realpath(__file__))),
                  '..', 'data', 'products')
  dataset = PygNodePropPredDataset('ogbn-products', root)
  train_idx = dataset.get_idx_split()["train"]
  train_loader = torch.utils.data.DataLoader(train_idx,
                                             batch_size=1024,
                                             pin_memory=True,
                                             shuffle=True)
  csr_topo = quiver.CSRTopo(dataset[0].edge_index)
  quiver_sampler = quiver.pyg.GraphSageSampler(csr_topo, [15, 10, 5],
                                               device=0,
                                               mode="GPU")
  quiver_feature = quiver.Feature(rank=0,
                                  device_list=[0],
                                  device_cache_size=cache_size,
                                  cache_policy="device_replicate",
                                  csr_topo=csr_topo)
  quiver_feature.from_cpu_tensor(dataset[0].x)
  total_num = 0
  total_time = 0
  for seeds in train_loader:
    nid, _, _ = quiver_sampler.sample(seeds)
    torch.cuda.synchronize()
    start = time.time()
    res = quiver_feature[nid]
    torch.cuda.synchronize()
    total_time += time.time() - start
    total_num += res.numel()
  torch.cuda.synchronize()
  print('Lookup {} ids, takes {} secs, Throughput {} GB/s.'\
    .format(total_num, total_time, total_num * 4 / total_time/ (1024**3)))