Sample_len Value

[gretelai]

Sample_len Value

[kboyd]

Thanks for the question @Frankie0609! I'm guessing this question is for the timeseries_dgan model. Let me know if that's not the case.

In our DGAN model, sample_len controls some internals of how we model a time series. The max_sequence_len parameter is how many time points are in each of your example time series. sample_len needs to divide max_sequence_len evenly, and is used to implicitly split the sequence into smaller chunks for the model to work with. Specifically, DGAN uses an RNN architecture and sample_len is how many time points are generated from each cell of the RNN.

We recommend using sample_len=1 for shorter time sequences, say up to ~20 (max_sequence_len=20). In longer sequences, being able to experiment with different values for sample_len allows you to explore the tradeoffs between a larger model that probably requires more data to train(small sample_len) and a smaller model with faster per epoch training (larger sample_len). It can also be very useful if you know there's periodicity in your data, e.g., use sample_len=7 for daily data with weekly patterns, though this is not required.

There's a few places to learn more about this model. For this particular implementation, see our blog posts https://gretel.ai/blog/create-synthetic-time-series-with-doppelganger-and-pytorch and https://gretel.ai/blog/generate-time-series-data-with-gretels-new-dgan-model. And our PyTorch implementation is based on the DoppelGANger model published in https://arxiv.org/abs/1909.13403. This paper has some discussion about including sample_len as a configurable parameter for the model.

Hope that information helps! Let me know if you have any other questions.

[Frankie0609]

Much thanks for this clarification