How pass a name to models saved with sagemaker.sklearn.estimator.SKLearn

[plynch-chwy]

Please fill out the form below.

System Information

• Framework (e.g. TensorFlow) / Algorithm (e.g. KMeans): SKLearn
• Framework Version: 0.20.0
• Python Version: 3.6
• CPU or GPU: CPU
• Python SDK Version: 1.18.16.dev0
• Are you using a custom image: No

Describe the problem

I am creating an SKLearn estimator and specifying a base_job_name, but my saved models are being saved as sagemaker-scikit-learn-timestamp.

from sagemaker.sklearn.estimator import SKLearn

...

my_estimator = SKLearn(
    entry_point='entry.py',
    source_dir='src/sagemaker/',
    hyperparameters=hyperparams,
    py_version='py3',
    framework_version='0.20.0',
    role=role,
    train_instance_type=instance_type,
    base_job_name='my-classifier',
    output_path=outputs,
    code_location=code_location)

my_estimator.fit(inputs=inputs)

An example saved model that is saved from this training job is then sagemaker-scikit-learn-2019-05-08-12-50-06-661.

[ericangelokim]

Hello,

Just to clarify, your model file name should be determined from the entry point script (like so: https://github.com/awslabs/amazon-sagemaker-examples/blob/master/sagemaker-python-sdk/scikit_learn_iris/scikit_learn_iris.py#L60), but the job name should be created from the base_job_name. The code for this can be seen within the SDK:

Estimator _prepare_for_training: https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/estimator.py#L175

SKLearn estimator: https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/sklearn/estimator.py (method not overriden)

Is the job name correct?

[plynch-chwy]

Sorry, not sure if I made it clear that I have 2 scripts. The first launches the training job and saves the model (shown above). The other then looks up the training job name and deploys the endpoint:

import boto3
from sagemaker.sklearn.estimator import SKLearn

...

job_name = wait_for_training('my-classifier')
my_estimator = SKLearn.attach(job_name)
predictor = my_estimator.deploy(initial_instance_count=1,
                                       instance_type=instance_type,
                                       endpoint_name='my-classifier',
                                       update_endpoint=endpoint_exists('my-classifier'))

The wait_for_training waits until training is complete and returns the training job name and endpoint_exists returns True if endpoint already exists so I can update it.

My training job name does get named properly as my-classifier-timestamp . I do save multiple objects besides the model saved to /opt/ml/model though (category feature mappings needed to processing inputs). The model file is saved as my-classifier.joblib, but the tar file gets saved as model.tar.gz in S3.

When I deploy the model with the 2nd script, the SageMaker model name is then named as sagemaker-scikit-learn-timestamp. Even though I'm using SKLearn, this is an issue since I will have multiple deployed models from SKLearn in the future, and would potentially like to find them by name using boto3.

[plynch-chwy]

I've found that if I combine my 2 scripts to train and deploy, I don't have this problem:

my_estimator.fit(inputs=inputs, wait=True)
my_estimator.deploy(initial_instance_count=1)

I haven't tested further yet, but seems the model name is not getting attached and passed through properly.

[ericangelokim]

Hello

https://github.com/aws/sagemaker-python-sdk/blob/master/src/sagemaker/sklearn/estimator.py#L119

It looks like the problem is due to the Model creation setting the name as the _current_job_name instead of latest_job_name.name. The new estimator never set a current job due to not having run a training job, so the _current_job_name is None.

I will be driving a fix from the SDK side to fix this usecase. Thank you for your patience.

[ericangelokim]

Hello, the fix was pushed into master; you should be able to attach the correct name now.

https://github.com/aws/sagemaker-python-sdk/pull/808/files