How to Configure Metrics Collector

This guide describes how Katib metrics collector works.

Overview

There are two ways to collect metrics:

1. Pull-based: collects the metrics using a sidecar container. A sidecar is a utility container that supports the main container in the Kubernetes Pod.

2. Push-based: users push the metrics directly to Katib DB in the training scripts.

In the metricsCollectorSpec section of the Experiment YAML configuration file, you can define how Katib should collect the metrics from each Trial, such as the accuracy and loss metrics.

Pull-based Metrics Collector

Your training code can record the metrics into StdOut or into arbitrary output files.

To define the pull-based metrics collector for your Experiment:

1. Specify the collector type in the .collector.kind field. Katib’s metrics collector supports the following collector types:

   • StdOut: Katib collects the metrics from the operating system’s default output location (standard output). This is the default metrics collector.

   • File: Katib collects the metrics from an arbitrary file, which you specify in the .source.fileSystemPath.path field. Training container should log metrics to this file in TEXT or JSON format. If you select JSON format, metrics must be line-separated by epoch or step as follows, and the key for timestamp must be timestamp:

     {"epoch": 0, "foo": "bar", "fizz": "buzz", "timestamp": "2021-12-02T14:27:51"}
     {"epoch": 1, "foo": "bar", "fizz": "buzz", "timestamp": "2021-12-02T14:27:52"}
     {"epoch": 2, "foo": "bar", "fizz": "buzz", "timestamp": "2021-12-02T14:27:53"}
     {"epoch": 3, "foo": "bar", "fizz": "buzz", "timestamp": "2021-12-02T14:27:54"}
     

     Check the file metrics collector example for TEXT and JSON format. Also, the default file path is /var/log/katib/metrics.log, and the default file format is TEXT.

   • TensorFlowEvent: Katib collects the metrics from a directory path containing a tf.Event. You should specify the path in the .source.fileSystemPath.path field. Check the TFJob example. The default directory path is /var/log/katib/tfevent/.

   • Custom: Specify this value if you need to use a custom way to collect metrics. You must define your custom metrics collector container in the .collector.customCollector field. Check the custom metrics collector example.

2. Write code in your training container to print or save to the file metrics in the format specified in the .source.filter.metricsFormat field. The default metrics format value is:

   ([\w|-]+)\s*=\s*([+-]?\d*(\.\d+)?([Ee][+-]?\d+)?)
   

   Each element is a regular expression with two sub-expressions. The first matched expression is taken as the metric name. The second matched expression is taken as the metric value.

   For example, using the default metrics format and StdOut metrics collector, if the name of your objective metric is loss and the additional metrics are recall and precision, your training code should print the following output:

   epoch 1:
   loss=3.0e-02
   recall=0.5
   precision=.4
   
   epoch 2:
   loss=1.3e-02
   recall=0.55
   precision=.5
   

Push-based Metrics Collector

Your training code needs to call report_metrics() function in Python SDK to record metrics. The report_metrics() function works by parsing the metrics in metrics field into a gRPC request, automatically adding the current timestamp for users, and sending the request to Katib DB Manager.

But before that, kubeflow-katib package should be installed in your training container.

To define the push-based metrics collector for your Experiment, you have two options:

• YAML File

  1. Specify the collector type Push in the .collector.kind field.

  2. Write code in your training container to call report_metrics() to report metrics.

• tune function

  Use tune function and specify the metrics_collector_config field. You can reference to the following example:

  import kubeflow.katib as katib
  
  def objective(parameters):
    import time
    import kubeflow.katib as katib
    time.sleep(5)
    result = 4 * int(parameters["a"])
    # Push metrics to Katib DB.
    katib.report_metrics({"result": result})
  
  katib.KatibClient(namespace="kubeflow").tune(
    name="push-metrics-exp",
    objective=objective,
    parameters= {"a": katib.search.int(min=10, max=20)}
    objective_metric_name="result",
    max_trial_count=2,
    metrics_collector_config={"kind": "Push"},
    # When SDK is released, replace it with packages_to_install=["kubeflow-katib==0.18.0"].
    # Currently, the training container should have `git` package to install this SDK. 
    packages_to_install=["git+https://github.com/kubeflow/katib.git@master#subdirectory=sdk/python/v1beta1"],
  )