使用 PyTorch 编写分布式应用程序

"""run.py:"""
#!/usr/bin/env python
importos
importsys
importtorch
importtorch.distributedasdist
importtorch.multiprocessingasmp

defrun(rank, size):
""" Distributed function to be implemented later. """
    pass

definit_process(rank, size, fn, backend='gloo'):
""" Initialize the distributed environment. """
    os.environ['MASTER_ADDR'] = '127.0.0.1'
    os.environ['MASTER_PORT'] = '29500'
    dist.init_process_group(backend, rank=rank, world_size=size)
    fn(rank, size)


if __name__ == "__main__":
    world_size = 2
    processes = []
    if "google.colab" in sys.modules:
        print("Running in Google Colab")
        mp.get_context("spawn")
    else:
        mp.set_start_method("spawn")
    for rank in range(world_size):
        p = mp.Process(target=init_process, args=(rank, world_size, run))
        p.start()
        processes.append(p)

    for p in processes:
        p.join()

"""Blocking point-to-point communication."""

defrun(rank, size):
    tensor = torch.zeros(1)
    if rank == 0:
        tensor += 1
        # Send the tensor to process 1
        dist.send(tensor=tensor, dst=1)
    else:
        # Receive tensor from process 0
        dist.recv(tensor=tensor, src=0)
    print('Rank ', rank, ' has data ', tensor[0])

"""Non-blocking point-to-point communication."""

defrun(rank, size):
    tensor = torch.zeros(1)
    req = None
    if rank == 0:
        tensor += 1
        # Send the tensor to process 1
        req = dist.isend(tensor=tensor, dst=1)
        print('Rank 0 started sending')
    else:
        # Receive tensor from process 0
        req = dist.irecv(tensor=tensor, src=0)
        print('Rank 1 started receiving')
    req.wait()
    print('Rank ', rank, ' has data ', tensor[0])

""" All-Reduce example."""
defrun(rank, size):
""" Simple collective communication. """
    group = dist.new_group([0, 1])
    tensor = torch.ones(1)
    dist.all_reduce(tensor, op=dist.ReduceOp.SUM, group=group)
    print('Rank ', rank, ' has data ', tensor[0])

""" Dataset partitioning helper """
classPartition(object):

    def__init__(self, data, index):
        self.data = data
        self.index = index

    def__len__(self):
        return len(self.index)

    def__getitem__(self, index):
        data_idx = self.index[index]
        return self.data[data_idx]


classDataPartitioner(object):

    def__init__(self, data, sizes=[0.7, 0.2, 0.1], seed=1234):
        self.data = data
        self.partitions = []
        rng = Random()  # from random import Random
        rng.seed(seed)
        data_len = len(data)
        indexes = [x for x in range(0, data_len)]
        rng.shuffle(indexes)

        for frac in sizes:
            part_len = int(frac * data_len)
            self.partitions.append(indexes[0:part_len])
            indexes = indexes[part_len:]

    defuse(self, partition):
        return Partition(self.data, self.partitions[partition])

""" Partitioning MNIST """
defpartition_dataset():
    dataset = datasets.MNIST('./data', train=True, download=True,
                             transform=transforms.Compose([
                                 transforms.ToTensor(),
                                 transforms.Normalize((0.1307,), (0.3081,))
                             ]))
    size = dist.get_world_size()
    bsz = 128 // size
    partition_sizes = [1.0 / size for _ in range(size)]
    partition = DataPartitioner(dataset, partition_sizes)
    partition = partition.use(dist.get_rank())
    train_set = torch.utils.data.DataLoader(partition,
                                         batch_size=bsz,
                                         shuffle=True)
    return train_set, bsz

""" Distributed Synchronous SGD Example """
defrun(rank, size):
    torch.manual_seed(1234)
    train_set, bsz = partition_dataset()
    model = Net()
    optimizer = optim.SGD(model.parameters(),
                          lr=0.01, momentum=0.5)

    num_batches = ceil(len(train_set.dataset) / float(bsz))
    for epoch in range(10):
        epoch_loss = 0.0
        for data, target in train_set:
            optimizer.zero_grad()
            output = model(data)
            loss = F.nll_loss(output, target)
            epoch_loss += loss.item()
            loss.backward()
            average_gradients(model)
            optimizer.step()
        print('Rank ', dist.get_rank(), ', epoch ',
              epoch, ': ', epoch_loss / num_batches)

""" Gradient averaging. """
defaverage_gradients(model):
    size = float(dist.get_world_size())
    for param in model.parameters():
        dist.all_reduce(param.grad.data, op=dist.ReduceOp.SUM)
        param.grad.data /= size