error with hyperparameters tuning

[guilloufre]

Please fill out the form below.

System Information

• Framework (e.g. TensorFlow) / Algorithm (e.g. KMeans): Factorization Machine
• Framework Version:
• Python Version: 2.7
• CPU or GPU: CPU
• Python SDK Version: 1.4.1
• Are you using a custom image:

Describe the problem

Hello,
I tried to use the newly released hyperparameters tuning described on the aws blog, but some error was thrown, when I launch the following command:
tuner.fit({'train': records_train, 'test': records_val})

This is the error:

Traceback (most recent call last):

  File "<ipython-input-20-9c3ba798ac1a>", line 1, in <module>
    tuner.fit({'train': s3_train_data, 'test': s3_val_data})

  File "/usr/local/lib/python2.7/dist-packages/sagemaker/tuner.py", line 144, in fit
    self.estimator._prepare_for_training(job_name)

  File "/usr/local/lib/python2.7/dist-packages/sagemaker/amazon/amazon_estimator.py", line 117, in _prepare_for_training
    feature_dim = records.feature_dim

AttributeError: 'NoneType' object has no attribute 'feature_dim'

Both records_train and records_val are RecordSet objects. For example, this is records_train:
(<class 'sagemaker.amazon.amazon_estimator.RecordSet'>, {'s3_data_type': 'S3Prefix', 'feature_dim': 2229, 'num_records': 7923, 'channel': 'train', 's3_data': 's3://###############'})

The training of the Factorization machine works if I launch:
fm_estimator.fit(records_train, mini_batch_size = 1000)
I also tried by providing direct links to s3 instead of RecordSet object with
tuner.fit({'train': s3_train_data, 'test': s3_val_data})
like in the example on the blog, but it throws the same error.

Thanks for helping me about this issue!

[laurenyu]

hi @guilloufre, thanks for trying out the new hyperparameter tuning feature!

The error is because you need to pass a list with the RecordSet objects instead of a dict to fit(). (the channel names are already specified in the RecordSet class, so no need to write them again.)

[guilloufre]

Thanks, it was actually easy! :)

[phschimm]

When using the record_set() method referenced here, the created RecordSet does not use the required S3DataDistributionType=FullyReplicated but instead uses ShardedByS3Key.

I'm trying to use the high level Python API to get metrics about my created models:

from sagemaker import RandomCutForest
from sagemaker.tuner import HyperparameterTuner, IntegerParameter

rcf = RandomCutForest(..., eval_metrics=['accuracy', 'precision_recall_fscore'])

train_set = rcf.record_set(features,
                           channel='train')

test_set = rcf.record_set(features,
                          labels=labels,
                          channel='test')

tuner = HyperparameterTuner(estimator=rcf,
                            objective_metric_name='test:f1',
                            hyperparameter_ranges={'num_samples_per_tree': IntegerParameter(32, 512),
                                                   'num_trees': IntegerParameter(50, 1000)},
                            max_jobs=1,
                            max_parallel_jobs=1)

tuner.fit([train_set, test_set])

When I execute this code, I get the following error in the AWS SageMaker console:

Failure reason
ClientError: Unable to initialize the algorithm. Failed to validate input data configuration. (caused by ValidationError) Caused by: 'ShardedByS3Key' is not one of ['FullyReplicated'] Failed validating 'enum' in schema['properties']['test']['properties']['S3DistributionType']: {'enum': ['FullyReplicated'], 'type': 'string'} On instance['test']['S3DistributionType']: 'ShardedByS3Key'

Is there a manual way to create a RecordSet that has the correct S3DataDistributionType via the Pyton API?