Hua Tang
06/19/2024, 6:13 PMHua Tang
06/20/2024, 6:02 AMHuseyn Zeynalov
06/20/2024, 7:39 AMGiovanni Perri
06/21/2024, 9:43 AMBill
06/22/2024, 3:40 AMBrijesh Mantri
06/24/2024, 3:11 PMTal Zaquin
06/25/2024, 9:17 AMavailable_mask
column. But I couldn't find any documentation about it.
a. Should I add it to my train df?
b. Is there a need to define it somewhere in the model or the fit command?
c. I also have some exogenous features that I would like to use, but they also have missing values in certain timestamps (the same as the target). Should I leave it as nan (let's say I'm using the TSMixerx model)?
2. This is more of a theoretical question, and continuing with the use of the TSMixerx model: I have multiple time series that I want to run multivariate forecasting. However, due to complexities in data acquisition, the time series have different lengths, and they do not start or end on the same timestamp. The only good thing about it is that most of the time stamps are found in all of the series.
a. Could you suggest a way to deal with it?
b. Can I do multivariate forecasting, or should I go for univariate forecasting and choose a single model for each series?Vidar Ingason
06/25/2024, 12:49 PMmodel = AutoNHITS(h=12,
loss=MAE(),
config=nhits_config,
search_alg=HyperOptSearch(),
backend='ray',
num_samples=10)
nf = NeuralForecast(models=[model], freq='M')
nf.fit(df=Y_df, val_size=24)
Banafsheh Nikbakht
06/25/2024, 4:38 PMYag ger Phone
06/26/2024, 8:08 AMJasmine Rienecker
06/27/2024, 11:42 PMJonathan
06/28/2024, 12:31 PMD N
07/03/2024, 3:45 PMAli Muqtadir
07/04/2024, 7:31 AMRodrigo Sodré
07/07/2024, 9:36 PMmodel = AutoLSTM(h=12, config=config, num_samples=1, cpus=1)
model.fit(dataset=dataset)
and it throws an AttributeError: 'DataFrame' object has no attribute 'temporal_cols'
But when i put the fit line inside a NeuralForecast it works with the same model and dataset:
models = [ model ]
nf = NeuralForecast(models=models, freq='M')
nf.fit(df=data)
My dataframe thas the unique_id, ds, y
columns and no empty or nan row
Could anyone explain what's happening?Banafsheh Nikbakht
07/08/2024, 7:40 AMMech Engineer
07/08/2024, 6:51 PMlink▾
Tracy Teal
07/08/2024, 9:32 PMRodrigo Sodré
07/09/2024, 12:13 AM# Use your own config or AutoNHITS.default_config
config = dict(max_steps=1, val_check_steps=1, input_size=12, lstm_hidden_size=8)
model = AutoDeepAR(h=12, config=config, num_samples=1, cpus=1)
That breaks the code. As fair as I understood, each AutoModel has its own get_default_config. That's happening for these model documentations: AutoDeepAR, AutoBiTCN and all transformer models.Rodrigo Sodré
07/09/2024, 1:14 AM| Name | Type | Params | Mode
--------------------------------------------------------
0 | loss | MAE | 0 | eval
1 | padder_train | ConstantPad1d | 0 | train
2 | scaler | TemporalNorm | 0 | train
3 | decomp | SeriesDecomp | 0 | train
4 | enc_embedding | DataEmbedding | 192 | train
5 | dec_embedding | DataEmbedding | 192 | train
6 | encoder | Encoder | 50.8 K | train
7 | decoder | Decoder | 53.1 K | train
--------------------------------------------------------
104 K Trainable params
0 Non-trainable params
104 K Total params
0.417 Total estimated model params size (MB)
`Trainer.fit` stopped: `max_steps=500` reached.
CPU times: user 51.3 s, sys: 4.2 s, total: 55.5 s
Wall time: 18min 2s
yh xu
07/11/2024, 2:17 AMYag ger Phone
07/11/2024, 11:28 AMTarik Jakan
07/11/2024, 12:44 PMCándido Otero
07/12/2024, 2:43 PMh
after some delay d
between my past data p
and the start of h
, since not all my past data may be available at prediction time.
I.e.:
|------------ p ------------|- d -|----- h ------|
·····································································
I believe this is equivalent to the "Forecast Start Shifting" in Darts.
I cannot find an analogous option in neuralforecast. Of course, I could simply predict a new horizon h' = d + h
, but I expect this approach to perform suboptimally in cases where d
is of similar magnitude to (or even larger than) h
.
Is there an out-of-the-box solution for doing this that I may have overlooked?
Otherwise, what would be the most elegant solution? Would the approach vary between models or can I fix it in the TimeSeriesDataset
class (or the BaseWindows
class, depending on the model family)?
Thanks in advance,
CándidoAssing Yang
07/14/2024, 6:39 PMTal Zaquin
07/15/2024, 1:57 PMcross_validation(val_size, test_size)
?
From what I found, I can get the best score based on the loss function that I defined, but it will be on all series, and I would like to calculate different metrics for each series by itself.
Thank youVincent Beast
07/16/2024, 10:59 AMTina Sedaghat
07/18/2024, 6:54 PMlocal_scaler_type
scale each target (unique_id) individually or does it scale the ‘y’ column at once?Tarik Jakan
07/23/2024, 6:08 AMhist_exog
or futr_exog
parameters in models like NHITS does that mean the model will not use any of the features other than the target?
Thank you so muchMilo
07/23/2024, 8:38 PM