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

Farzad E

02/23/2023, 3:06 PM
When using AutoNHITS are we supposed to pass any option related to the window size? From neuralforecast/common/_base_windows.py I get
RuntimeError: maximum size for tensor at dimension 2 is 1463 but size is 1825
with 1463 being the size of my time series. The training continues though regardless of this error so I am not sure if this has any negative impact.
k

Kin Gtz. Olivares

02/23/2023, 3:10 PM
You might want to reduce the window size for it not to exceed the size of your series
f

Farzad E

02/23/2023, 3:15 PM
@Kin Gtz. Olivares is that controlled by windows_batch_size?
k

Kin Gtz. Olivares

02/23/2023, 3:19 PM
Yes, the windows_batch_size controls the size of lags+horizon
👍 1
c

Cristian (Nixtla)

02/23/2023, 4:08 PM
@Farzad E I think you should reduce
input_size
👍 1
f

Farzad E

02/23/2023, 4:28 PM
@Cristian (Nixtla) Thanks. Passing input_size explicitly in the config worked. Passing windows_batch_size by itself did not work. That one threw an error about missing input_size and then not finding the best trial. It seems that with AutoNHITS there are certain options that cannot be passed as stand alone so if we want to specify them in the config we have to pass others like input_size in the config as well (i.e., the algorithm is not going to take that one option from the config and apply all other defaults for anything else not passed in the config).
c

Cristian (Nixtla)

02/23/2023, 4:36 PM
When a hyperparameter is not specified in the config, it will use the default values of the
NHTIS
.
input_size
does not have a default value, so you should always include it in the config
👍 1
f

Farzad E

02/23/2023, 4:38 PM
Ahh that makes sense. Should have paid more attention to the default values.
2 Views