No metrics when I use tensorflow estimator

[xzy0223]

hi

I'm trying step functions data science sdk, but I ran into a issue

I created tf estimator in my jupyter notebook like this:
tf_estimator = TensorFlow( entry_point='my_train_1.py', model_dir=model_dir, output_path=s3_output_location, train_instance_type=train_instance_type, train_instance_count=1, hyperparameters=hyperparameters, role=sagemaker.get_execution_role(), #base_job_name='tf-scriptmode-mnist', framework_version='2.0.0', py_version='py3', metric_definitions=metric_definitions, script_mode=True)

I really defined the "metric_definitions", but when I chained all steps as a workflow, and executed it.

there is no customer metrics in sagemaker console

but when I just create the estimator and fit without step functions data science sdk, I can get the metric.

what did I miss? or any suggestion?

thanks a lot

[hanaan-yousaf]

I had the same issue using sagemaker.sklearn.estimator.SKLearn

I had to manually add them to the step function definition under AlgorithmSpecification in the training state like so

"AlgorithmSpecification": {
  "TrainingImage": "someimage",
  "MetricDefinitions": [ 
         { 
            "Name": "foo",
            "Regex": "bar=(.*?);"
         }
      ]
}

[xzy0223]

nually add them to the step function definit

thanks, bro!!!

[shunjd]

It happens to be an issue in SageMaker SDK, I've created an issue on their repo.