基于ray的分布式机器学习（二）

基本思路：基于parameter server + multiple workers模式。同步方式：parameter server负责网络参数的统一管理，每次迭代均将参数发送给每一个worker，多个worker同时迭代数据集，计算当前批次的损失和梯度，当所有worker全部完成当前批次的计算后，将每个worker的梯度回传给parameter server，parameter server使用该梯度进行参数优化。异步方式：与同步方式不同的是，parameter server不需要每次等待所有worker全部完成一个批次的计算后再利用所有worker的梯度更新网络参数，而是每当有一个worker完成一个批次的计算时，立刻进行网络参数的更新，并将新的参数下发给该worker。1、定义模型
class ConvNet(nn.Module):
    def __init__(self):
        super(ConvNet, self).__init__()
        self.conv1 = nn.Conv2d(1, 3, kernel_size=3)
        self.fc = nn.Linear(192, 10)

    def forward(self, x):
        x = F.relu(F.max_pool2d(self.conv1(x), 3))
        x = x.view(-1, 192)
        x = self.fc(x)
        return F.log_softmax(x, dim=1)

    def get_weights(self):
        return {k: v.cpu() for k, v in self.state_dict().items()}

    def set_weights(self, weights):
        self.load_state_dict(weights)

    def get_gradients(self):
        grads = []
        for p in self.parameters():
            grad = None if p.grad is None else p.grad.data.cpu().numpy()
            grads.append(grad)
        return grads

    def set_gradients(self, gradients):
        for g, p in zip(gradients, self.parameters()):
            if g is not None:
                p.grad = torch.from_numpy(g)

2、定义parameter server
@ray.remote
class ParameterServer(object):
    def __init__(self, lr):
        self.model = ConvNet()
        self.optimizer = torch.optim.SGD(self.model.parameters(), lr=lr)

    def apply_gradients(self, *gradients):
        summed_gradients = [
            np.stack(gradient_zip).sum(axis=0)
            for gradient_zip in zip(*gradients)
        ]
        self.optimizer.zero_grad()
        self.model.set_gradients(summed_gradients)
        self.optimizer.step()
        return self.model.get_weights()

    def get_weights(self):
        return self.model.get_weights()

3、定义worker
@ray.remote
class DataWorker(object):
    def __init__(self):
        self.model = ConvNet()
        self.data_iterator = iter(get_data_loader()[0])

    def compute_gradients(self, weights):
        self.model.set_weights(weights)
        try:
            data, target = next(self.data_iterator)
        except StopIteration:  # When the epoch ends, start a new epoch.
            self.data_iterator = iter(get_data_loader()[0])
            data, target = next(self.data_iterator)
        self.model.zero_grad()
        output = self.model(data)
        loss = F.nll_loss(output, target)
        loss.backward()
        return self.model.get_gradients()

4、同步训练
iterations = 200
num_workers = 2

ray.init(ignore_reinit_error=True)
ps = ParameterServer.remote(1e-2)
workers = [DataWorker.remote() for i in range(num_workers)]

model = ConvNet()
test_loader = get_data_loader()[1]

print("Running synchronous parameter server training.")
current_weights = ps.get_weights.remote()
for i in range(iterations):
    gradients = [
        worker.compute_gradients.remote(current_weights) for worker in workers
    ]

    current_weights = ps.apply_gradients.remote(*gradients)

    if i % 10 == 0:
        model.set_weights(ray.get(current_weights))
        accuracy = evaluate(model, test_loader)
        print("Iter {}: \taccuracy is {:.1f}".format(i, accuracy))

print("Final accuracy is {:.1f}.".format(accuracy))
ray.shutdown()

5、异步训练
print("Running Asynchronous Parameter Server Training.")

ray.init(ignore_reinit_error=True)
ps = ParameterServer.remote(1e-2)
workers = [DataWorker.remote() for i in range(num_workers)]

current_weights = ps.get_weights.remote()

gradients = {}
for worker in workers:
    gradients[worker.compute_gradients.remote(current_weights)] = worker

for i in range(iterations * num_workers):
    ready_gradient_list, _ = ray.wait(list(gradients))
    ready_gradient_id = ready_gradient_list[0]
    worker = gradients.pop(ready_gradient_id)

    current_weights = ps.apply_gradients.remote(*[ready_gradient_id])
    gradients[worker.compute_gradients.remote(current_weights)] = worker

    if i % 10 == 0:
        model.set_weights(ray.get(current_weights))
        accuracy = evaluate(model, test_loader)
        print("Iter {}: \taccuracy is {:.1f}".format(i, accuracy))

print("Final accuracy is {:.1f}.".format(accuracy))