maxbrenner-ai / seq2seq-time-series-forecasting-fully-recurrent Goto Github PK

Hey Max!

Just wanted to thank you for the awesome notebook. I've got a question about handling dec_targets when they're out of sample and unknown. What's the correct approach to make the code work in this scenario? My plan is to have the first step of dec_inputs and leave all the remaining parts as np.nan. I've experimented with a couple of strategies:

1. Setting batch_dec_inputs[:,1:,:] = np.nan resulted in an error.
2. Trying batch_dec_inputs[:,1:,:] = 0.0 significantly reduced the quality of the output. I was surprised by this because with teacher_force_prob=None, these data points set to 0.0 shouldn't make any difference.

Any insights on how to tackle this would be greatly appreciated!