aws-samples / amazon-sagemaker-gluonts-entrypoint

Sample GluonTS entrypoint scripts for Amazon SageMaker.

Example Entrypoint Scripts to run GluonTS on Amazon SageMaker

This repository contains the example that supplement the blog post on using GluonTS to develop custom product demand forecasting on Amazon SageMaker, and generate demand forecasts for each products with custom models.

UPDATES:

• the main branch has been updated to gluonts-0.8 and mxnet-1.7.
• the blog-version branch contains the code referred by the blog post. This branch uses gluonts-0.5 and mxnet-1.6

The sample notebooks demonstrate faster experimentation with multiple custom models. The changes involves only parameteric changes to the entrypoint train script.

Implementation notes

The entrypoint train script utilizes GluonTS to work with time series data in a simpler fashion, and to access many state-of-the-art custom models in the GluonTS library. The novelties demonstrated by the script include:

• a single generic train script that supports multiple built-in algorithms in GluonTS.

• Guide modular style implementation in notebooks and entrypoint scripts.

• The train script passes through hyperparameters which it receives as command line arguments, directly to the specified GluonTS estimator.

  • The entrypoint train script does not need to explicitly declare all those hyperpameters in its body.

  • This functionality utilizes another sample repository: aws-samples/amazon-sagemaker-entrypoint-utilities, a dependency that the entrypoint train script declares in its requirements.txt

• At the end of training, the train script performs backtesting on a test split. Then, the script outputs to files the backtest performance metrics, the forecasts resulted from test split, and backtest plots of all timeseries as montages and individual plots.

  • This is done by design, and the aim is to let data scientists quickly analyze their model performance without having to separately run inferences (and additional downstream tasks for plotting etc.).

  • This functionality utilizes another sample repository: aws-samples/smallmatter-package, a dependency that the entrypoint train script declares in its requirements.txt

Security

See CONTRIBUTING for more information.

License

This library is licensed under the MIT-0 License. See the LICENSE file.

Authors

• Verdi March GitHub
• Beibit Baktygaliyev
• Zmnako Awrahman GitHub

amazon-sagemaker-gluonts-entrypoint's Issues

Training job failed when executing train.py

Hi,
Thanks for this great sample code.

I followed all the instructions. But the training job fails.

I'm getting the following message:
AlgorithmError: ExecuteUserScriptError: Command "/usr/local/bin/python3.6 train.py --algo gluonts.model.deepar.DeepAREstimator --cardinality [151] --cell_type gru --distr_output.__class__ gluonts.distribution.gaussian.GaussianOutput --num_cells 46 --num_layers 4 --num_samples 1000 --plot_transparent 0 --prediction_length 30 --trainer.__class__ gluonts.trainer.Trainer --trainer.epochs 200 --trainer.learning_rate 0.0005490034989544101 --use_feat_static_cat True"

The Traceback in cloud watch provides me with this error

Traceback (most recent call last):
  File "train.py", line 197, in <module>
    train(args, smepu.argparse.to_kwargs(train_args))
  File "/usr/local/lib/python3.6/site-packages/smepu/argparse.py", line 83, in to_kwargs
    args_round_2 = _round_2(args_round_1)  # Custom class in IR
  File "/usr/local/lib/python3.6/site-packages/smepu/argparse.py", line 211, in _round_2
    desered = {k: decode(v.klass_dict) for k, v in ir.items()}
  File "/usr/local/lib/python3.6/site-packages/smepu/argparse.py", line 211, in <dictcomp>
    desered = {k: decode(v.klass_dict) for k, v in ir.items()}
  File "/usr/local/lib/python3.6/site-packages/smepu/_gluonts_core_serde.py", line 49, in decode
    return cls(*args, **kwargs)  # type: ignore

TypeError: 'NoneType' object is not callable

I'm using latest gluonts release 0.7.6