Created by: HugoLian

程序单机训练的auc是对的，集群训练上无论训练集（event.metrics）还是测试集（results.metrics）打印出来都是0是为什么呢？训练数据和测试数据肯定是正常的，已经检验过了，不存在连续1或者0分布。

集群训练程序打印auc的方式如下：

    def event_handler(event):
        if isinstance(event, paddle.event.EndIteration):
            if (event.batch_id + 1) % 400 == 0:
                print "\nPass %d Batch %d Cost %.4f %s" % (
                    event.pass_id, event.batch_id, event.cost, event.metrics)
                result = trainer.test(
                    reader=paddle.batch(cluster_data_reader(cluster_test_dir, node_id), batch_size = 1024),
                    feeding=feeding)
                print "\nTest %d, Cost %6f, %s" % (event.pass_id, result.cost, result.metrics)
             
    trainer.train(
        reader=paddle.batch(
            paddle.reader.shuffle(cluster_data_reader(cluster_train_dir, node_id), buf_size=102400),
            batch_size=1024),
        event_handler=event_handler,
        feeding=feeding,
        num_passes=20)

cluster_data_reader()是读取文件yield数据的方法，没有问题。trainer的写法是：

    auc_layer = paddle.layer.slope_intercept(input=inference, name='auc_layer', slope=0.5, intercept=0.5)
    cost = paddle.layer.regression_cost(input=auc_layer, label=label)

    parameters = paddle.parameters.create(cost)

    trainer = paddle.trainer.SGD(
        cost = cost,
        extra_layers = paddle.evaluator.auc(input=auc_layer, label=label),
        parameters = parameters,
        update_equation = paddle.optimizer.Adam(learning_rate=2e-4),
        is_local = False)

这样每轮递归结束后，训练集（Pass 0）和测试集(Test 0）的结果都是0，是什么地方出了问题呢？

Thu Aug  9 15:04:48 2018[1,0]<stdout>:Pass 0 Batch 399 Cost 0.1073 {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:04:48 2018[1,3]<stdout>:
Thu Aug  9 15:04:48 2018[1,3]<stdout>:Pass 0 Batch 399 Cost 0.1220 {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:04:48 2018[1,2]<stdout>:
Thu Aug  9 15:04:48 2018[1,2]<stdout>:Pass 0 Batch 399 Cost 0.1143 {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:04:48 2018[1,1]<stdout>:
Thu Aug  9 15:04:48 2018[1,1]<stdout>:Pass 0 Batch 399 Cost 0.1155 {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:04:48 2018[1,4]<stdout>:
Thu Aug  9 15:04:48 2018[1,4]<stdout>:Pass 0 Batch 399 Cost 0.1208 {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:17:22 2018[1,2]<stdout>:
Thu Aug  9 15:17:22 2018[1,2]<stdout>:Test 0, Cost 0.114977, {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:17:26 2018[1,0]<stdout>:
Thu Aug  9 15:17:26 2018[1,0]<stdout>:Test 0, Cost 0.114977, {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:17:28 2018[1,4]<stdout>:
Thu Aug  9 15:17:28 2018[1,4]<stdout>:Test 0, Cost 0.114977, {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:17:30 2018[1,3]<stdout>:
Thu Aug  9 15:17:30 2018[1,3]<stdout>:Test 0, Cost 0.114977, {'__auc_evaluator_0__': 0.0}
Thu Aug  9 15:17:40 2018[1,1]<stdout>:
Thu Aug  9 15:17:40 2018[1,1]<stdout>:Test 0, Cost 0.114977, {'__auc_evaluator_0__': 0.0}