- Newest
- Most votes
- Most comments
Hello,
Unfortunately this is not fixing my problem. When I set the random seed in this way, I still end up with two different models, eventhough I have trained them with the exact same data and hyperparameters. I don't think that the locally set random seed will be transferred to the EC2 Instance where the DeepAR model is being trained. The random seed in the example provided is used for data preparation tasks, but not for the model training as it seems.
Are there any other ways to approach this problem? If not, are there any plans to implementing a random seed parameter for DeepAR?
Thanks!
Hello,
I understand that you want to provide a "random state" or a state in any sense to produce comparable results with the in-built DeepAR algorithm.
Kindly note that numpy can be used for setting random seeds for reproducibility using numpy.random.seed -
# set random seeds for reproducibility
np.random.seed(42)
random.seed(42)
Also, there is an example of SageMaker/DeepAR demo on electricity dataset from AWS which uses the same technique of setting random seeds. Here is the link to it -> https://sagemaker-examples.readthedocs.io/en/latest/introduction_to_amazon_algorithms/deepar_electricity/DeepAR-Electricity.html
Thanks :)
Relevant content
- asked 2 years ago
- AWS OFFICIALUpdated a year ago
- AWS OFFICIALUpdated 3 years ago
- AWS OFFICIALUpdated 3 years ago
- AWS OFFICIALUpdated 2 years ago
Hello,
Unfortunately this is not fixing my problem. When I set the random seed in this way, I still end up with two different models, eventhough I have trained them with the exact same data and hyperparameters. I don't think that the locally set random seed will be transferred to the EC2 Instance where the DeepAR model is being trained. The random seed in the example provided is used for data preparation tasks, but not for the model training as it seems.
Are there any other ways to approach this problem? If not, are there any plans to implementing a random seed parameter for DeepAR?
Thanks!