使用ml-engine调优超参数返回状态:失败

Question

我正在尝试使用ml-engine调优我的模型超参数，但我不太确定它是否可以工作。

我没有在HyperparameterSpec中指定algorithm标记，根据文档，它应该默认使用贝叶斯优化方法。Im也没有设置maxFailedTrials，根据文档，如果第一个跟踪失败，它应该结束所有跟踪。

这是我的配置

trainingInput:
  scaleTier: CUSTOM
  masterType: standard_gpu
  hyperparameters:
    goal: MAXIMIZE
    maxTrials: 8
    maxParallelTrials: 2
    hyperparameterMetricTag: test_accuracy
    params:
    - parameterName: dropout_rate
      type: DOUBLE
      minValue: 0.3
      maxValue: 0.7
      scaleType: UNIT_LINEAR_SCALE
    - parameterName: lr
      type: DOUBLE
      minValue: 0.0001
      maxValue: 0.0003
      scaleType: UNIT_LINEAR_SCALE

这是训练结果：

{
  "completedTrialCount": "8",
  "trials": [
    {
      "trialId": "1",
      "hyperparameters": {
        "lr": "0.00014959385395050048",
        "dropout_rate": "0.42217149734497067"
      },
      "startTime": "2019-10-07T09:40:02.143968039Z",
      "endTime": "2019-10-07T09:47:50Z",
      "state": "FAILED"
    },
    {
      "trialId": "2",
      "hyperparameters": {
        "dropout_rate": "0.62217149734497068",
        "lr": "0.00028292718728383382"
      },
      "startTime": "2019-10-07T09:40:02.144192681Z",
      "endTime": "2019-10-07T09:47:19Z",
      "state": "FAILED"
    },
    {
      "trialId": "3",
      "hyperparameters": {
        "lr": "0.00014846909046173097",
        "dropout_rate": "0.31717863082885739"
      },
      "startTime": "2019-10-07T09:48:09.266596472Z",
      "endTime": "2019-10-07T09:55:26Z",
      "state": "FAILED"
    },
    {
      "trialId": "4",
      "hyperparameters": {
        "lr": "0.00018741662502288819",
        "dropout_rate": "0.34178204536437984"
      },
      "startTime": "2019-10-07T09:48:10.761305330Z",
      "endTime": "2019-10-07T09:55:58Z",
      "state": "FAILED"
    },
    {
      "trialId": "5",
      "hyperparameters": {
        "dropout_rate": "0.6216828346252441",
        "lr": "0.00010192830562591553"
      },
      "startTime": "2019-10-07T09:56:15.904704865Z",
      "endTime": "2019-10-07T10:04:04Z",
      "state": "FAILED"
    },
    {
      "trialId": "6",
      "hyperparameters": {
        "dropout_rate": "0.42288427352905272",
        "lr": "0.000230206298828125"
      },
      "startTime": "2019-10-07T09:56:17.895067636Z",
      "endTime": "2019-10-07T10:04:05Z",
      "state": "FAILED"
    },
    {
      "trialId": "7",
      "hyperparameters": {
        "lr": "0.00019101441543291624",
        "dropout_rate": "0.36415641310447144"
      },
      "startTime": "2019-10-07T10:05:22.147233194Z",
      "endTime": "2019-10-07T10:13:09Z",
      "state": "FAILED"
    },
    {
      "trialId": "8",
      "hyperparameters": {
        "dropout_rate": "0.69955616224911532",
        "lr": "0.00029989311482522672"
      },
      "startTime": "2019-10-07T10:05:22.147396438Z",
      "endTime": "2019-10-07T10:13:30Z",
      "state": "FAILED"
    }
  ],
  "consumedMLUnits": 2.29,
  "isHyperparameterTuningJob": true,
  "hyperparameterMetricTag": "test_accuracy"
}

所有的跟踪都是运行的，所以我相信是搜索算法因为某些原因而失败了。我还没有找到更多的信息，为什么它通过运行另一个冗长的搜索算法返回这个或任何日志。

对我来说，它似乎无法在tensorflow事件文件中找到指标，但我不明白为什么，由于名称完全相同，使用tensorboard打开事件文件我能够看到数据。也许对日志结构有一些我不知道的要求？

日志记录指标的代码：

from tensorflow.contrib.summary import summary as summary_ops

# in __init__
self.tf_board_writer = summary_ops.create_file_writer(self.save_path)
....

# During training
with self.tf_board_writer.as_default(), summary_ops.always_record_summaries():
    summary_ops.scalar(name=name, tensor=value, step=step)

一个小问题，如果ml-engine团队中的任何一个最终出现在这里，现在TF2已经稳定并发布了，你知道它什么时候可以在运行时环境中使用吗？

无论如何，希望有人能帮助我:)

Answer 1

这个问题可以通过使用python包cloudml-hypertune和以下代码来解决：

self.hpt.report_hyperparameter_tuning_metric(
            hyperparameter_metric_tag=hypeparam_metric_name,
            metric_value=value,
            global_step=step)

然后将HyperparameterSpec中的hyperparameterMetricTag设置为hypeparam_metric_name