vincent-leguen/stripe

Code for our NeurIPS 2020 paper "Probabilistic Time Series Forecasting with Structured Shape and Temporal Diversity"

stripe's Issues

reproduce Electricity data set results

I tried to reproduce the reported MSE on Electricity data set. I tried various parameters but I could not get close to the MSE reported in the original paper. Can you please let me know which set of parameters you used in the paper? Or can you please upload the corresponding python file?

similar position of Z for shape and time vector

STRIPE/trainer/trainer.py

Line 217 in fd63eff

z = torch.cat( (z_f, z), dim=1)

I noticed that you have sampled z code at the same position when sending to decoder during train of shape and time stripe model.

however in evaluation code you place z at the beginning or end of vector depending on what type of stripe (shape/time) we are working with.

would you be so kind to explain correct way building code-vector for decoder?
thank you.

asking for explanation about sampling from the vae

Dear authors,

Thank you for providing the code for the mode.

Why is it that the target is passed to the vae model during sampling as in line 103 in trainer.py?