def test_glue(self):
# create estimator
estimator = self.create_estimator()
# run training
estimator.fit()
# result dataframe
result_metrics_df = TrainingJobAnalytics(estimator.latest_training_job.name).dataframe()
# extract kpis
eval_accuracy = list(result_metrics_df[result_metrics_df.metric_name == "eval_accuracy"]["value"])
eval_loss = list(result_metrics_df[result_metrics_df.metric_name == "eval_loss"]["value"])
# get train time from SageMaker job, this includes starting, preprocessing, stopping
train_runtime = (
Session().describe_training_job(estimator.latest_training_job.name).get("TrainingTimeInSeconds", 999999)
)
# assert kpis
assert train_runtime <= self.results["train_runtime"]
assert all(t >= self.results["eval_accuracy"] for t in eval_accuracy)
assert all(t <= self.results["eval_loss"] for t in eval_loss)
# dump tests result into json file to share in PR
with open(f"{estimator.latest_training_job.name}.json", "w") as outfile:
json.dump({"train_time": train_runtime, "eval_accuracy": eval_accuracy, "eval_loss": eval_loss}, outfile)
def test_scripz(self, instance_count):
# create estimator
estimator = self.create_estimator(instance_count)
# run training
estimator.fit()
# save csv
self.save_results_as_csv(estimator.latest_training_job.name)
# result dataframe
result_metrics_df = TrainingJobAnalytics(
estimator.latest_training_job.name).dataframe()
# extract kpis
train_runtime = list(result_metrics_df[result_metrics_df.metric_name ==
"train_runtime"]["value"])
eval_accuracy = list(result_metrics_df[result_metrics_df.metric_name ==
"eval_accuracy"]["value"])
eval_loss = list(result_metrics_df[result_metrics_df.metric_name ==
"eval_loss"]["value"])
# assert kpis
assert all(t <= self.results["train_runtime"] for t in train_runtime)
assert all(t >= self.results["eval_accuracy"] for t in eval_accuracy)
assert all(t <= self.results["eval_loss"] for t in eval_loss)
def _report_extended_online_metrics(self, describe_response):
self._report_secondary_transitions(describe_response)
# No reason to start reading metrics before the job is actually starting
start_time = self._job_start_time(describe_response)
if start_time:
if self._analytics is None:
self._analytics = TrainingJobAnalytics(
training_job_name=self._job_name,
metric_names=self._metric_names_for_training_job(),
start_time=start_time)
metrics_df = self._analytics.dataframe(force_refresh=True)
if not metrics_df.empty:
for index, row in metrics_df.iterrows():
Report.job_metric(
row.get(SMApiConstants.Estimator.DF_METRIC_NAME,
"Unknown"),
row.get(SMApiConstants.Estimator.DF_METRIC_VALUE, 0))
def save_results_as_csv(self, job_name):
TrainingJobAnalytics(job_name).export_csv(f"{self.env.test_path}/{job_name}_metrics.csv")
class JobMonitorEstimator(JobMonitorBase):
def __init__(self, sagemaker_client, job_name, logger):
super(self.__class__, self).__init__(sagemaker_client, job_name,
logger)
self._metric_names = None
self._analytics = None
def _describe_job(self):
return self._sagemaker_client.describe_training_job(
TrainingJobName=self._job_name)
def _job_status(self, describe_response):
return describe_response[SMApiConstants.Estimator.JOB_STATUS]
def _job_start_time(self, describe_response):
return describe_response.get(SMApiConstants.Estimator.START_TIME)
def _job_end_time(self, describe_response):
return describe_response.get(SMApiConstants.Estimator.END_TIME)
@cached_property
def _host_metrics_defs(self):
return [
JobMonitorBase.MetricMeta('cpuavg_{}',
SMApiConstants.METRIC_CPU_UTILIZATION,
SMApiConstants.STAT_AVG),
JobMonitorBase.MetricMeta('cpumin_{}',
SMApiConstants.METRIC_CPU_UTILIZATION,
SMApiConstants.STAT_MIN),
JobMonitorBase.MetricMeta('cpumax_{}',
SMApiConstants.METRIC_CPU_UTILIZATION,
SMApiConstants.STAT_MAX),
JobMonitorBase.MetricMeta('memavg_{}',
SMApiConstants.METRIC_MEMORY_UTILIZATION,
SMApiConstants.STAT_AVG),
JobMonitorBase.MetricMeta('memmin_{}',
SMApiConstants.METRIC_MEMORY_UTILIZATION,
SMApiConstants.STAT_MIN),
JobMonitorBase.MetricMeta('memmax_{}',
SMApiConstants.METRIC_MEMORY_UTILIZATION,
SMApiConstants.STAT_MAX),
JobMonitorBase.MetricMeta('diskavg_{}',
SMApiConstants.METRIC_MEMORY_UTILIZATION,
SMApiConstants.STAT_AVG),
JobMonitorBase.MetricMeta('diskmin_{}',
SMApiConstants.METRIC_MEMORY_UTILIZATION,
SMApiConstants.STAT_MIN),
JobMonitorBase.MetricMeta('diskmax_{}',
SMApiConstants.METRIC_MEMORY_UTILIZATION,
SMApiConstants.STAT_MAX)
]
def _metrics_namespace(self):
return SMApiConstants.Estimator.NAMESPACE
def _report_extended_online_metrics(self, describe_response):
self._report_secondary_transitions(describe_response)
# No reason to start reading metrics before the job is actually starting
start_time = self._job_start_time(describe_response)
if start_time:
if self._analytics is None:
self._analytics = TrainingJobAnalytics(
training_job_name=self._job_name,
metric_names=self._metric_names_for_training_job(),
start_time=start_time)
metrics_df = self._analytics.dataframe(force_refresh=True)
if not metrics_df.empty:
for index, row in metrics_df.iterrows():
Report.job_metric(
row.get(SMApiConstants.Estimator.DF_METRIC_NAME,
"Unknown"),
row.get(SMApiConstants.Estimator.DF_METRIC_VALUE, 0))
def _report_secondary_transitions(self, describe_response):
secondary_transitions = describe_response[
SMApiConstants.Estimator.SECONDARY_TRANSITIONS]
rows = []
for transition in secondary_transitions:
start_time = transition['StartTime']
end_time = transition.get('EndTime', datetime.now(pytz.UTC))
status = transition['Status']
message = transition['StatusMessage']
time_span = (end_time - start_time).total_seconds()
rows.append([
start_time.strftime("%Y-%m-%d, %H:%M:%S"),
end_time.strftime("%Y-%m-%d, %H:%M:%S"),
Report.seconds_fmt(time_span), status, message
])
if rows:
Report.job_secondary_transitions(rows)
def _metric_names_for_training_job(self):
if self._metric_names is None:
training_description = self._sagemaker_client.describe_training_job(
TrainingJobName=self._job_name)
metric_definitions = training_description[
SMApiConstants.Estimator.ALGO_SPEC][
SMApiConstants.Estimator.METRIC_DEFS]
self._metric_names = [
md[SMApiConstants.Estimator.METRIC_DEF_NAME]
for md in metric_definitions
if md[SMApiConstants.Estimator.METRIC_DEF_NAME].startswith(
SMApiConstants.Estimator.TRAIN_PREFIX)
]
return self._metric_names
def _report_extended_final_metrics(self, describe_response):
final_metrics = describe_response.get(
SMApiConstants.Estimator.FINAL_METRIC_DATA_LIST)
if final_metrics:
for metric in final_metrics:
Report.job_metric(
metric.get(SMApiConstants.Estimator.METRIC_NAME,
"Unknown"),
metric.get(SMApiConstants.Estimator.METRIC_VALUE, 0))
