https://github.com/nixtla logo
#neural-forecast
Title
# neural-forecast
m

Matthieu ANTUNES -LYON-

11/17/2023, 7:19 PM
Hi and first of all thanks for your job on Nixtla library 🙌 I'm currently fitting
neuralforecast
Temporal Fusion Transformers to some retail dataset and doing kind of a parallel with DeepAR training procedure (requiring
context_length
+
prediction_length
parameters) I wonder what would happen during training with series that are just the length of the
input_size
parameter. E.g. : if a series is 156-week long and my
input_size
model parameter is
156
, how are such series handled during training ? Are they padded with
0
in the past? Are they discarded? Cheers 🤓
c

Cristian (Nixtla)

11/17/2023, 8:38 PM
Hi @Matthieu ANTUNES -LYON-! Yes, each batch is padded with 0s on the left based on the longest time series. You will need to also set
start_padding_enabled
to True so that those very short ts (with length < input + h) are used during training
m

Matthieu ANTUNES -LYON-

11/20/2023, 8:56 AM
Hi Christian, thanks for your prompt reply. Crystal clear, much appreciated 😄