Hello, maybe you know how to solve this error? MisconfigurationException: The provided lr scheduler

StepLR

doesn't follow PyTorch's LRScheduler API. You should override the

<http://LightningModule.lr|LightningModule.lr>_scheduler_step

hook with your own logic if you are using a custom LR scheduler. Thanks

Is there a good tutorial/paper of

neuralforecast

application to the sales forecast? If there a comparison of

neuralforecast

result with other approaches that would be great

Hi everyone, I’m wondering if there’s a way to obtain in-sample predictions on the training set for a neural forecaster, similar to the forecast_fitted_values() function. Can anyone suggest a method to replicate this behavior? Thank you!

Could any one suggest loss and scaler type for TFT for negative values in data

Have anyone else experienced that a notebook freeze after hyperparameter search using AutoNBEATSx? I running a hyperparameter search based of the example notebook. The search runs fine but once the search is done the terminal or notebook just freezes and nothing happens. No error no feedback 😕. I managed to find this error in JupyterLab: ” Error displaying widget: model not found” I am using neuralforecast 1.4.0 with ray 2.3

Hi, I keep getting this error I have all the code in Google Colab 🥲 The provided lr scheduler

StepLR

does not follow PyTorch's LRScheduler API. You should replace the

<http://LightningModule.lr|LightningModule.lr>_scheduler_step

hook with your own logic if you are using a custom LR scheduler. Thanks

If sequences shorter than input_size are automatically padded with 0 at the beginning, and future exogenous variables are also automatically padded with 0 at the beginning (is this true?), in order to distinguish those zeros from real zeros of the time series, could it make sense to create a "real_data" exogenous variable with value 1 for the time series points, so that its value is automatically set to 0 for padding data and we can tell to the model that padding data have real_data=0 and true data have real_data=1? Basically we would set that variable to 1 for all time series points and then leverage the all-zero padding to create real_data=0. Could the scaler (e.g. robust) negatively impact this "real_data" indicator since future exogenous variables get scaled? Is the exogenous variables scaling performed before or after the automatic zero-padding? Thanks

In pytorch loss levels are quantiles or confidence interval.....say instead of predicting median I want the model to predict the 0.25 to 0.10 quantile on data which params should I tweak ? And what is high low in that case

Thread for practical insights of NHITS and TCN for Power Markets. (Courtesy of @Chris Gervais)

Hi 🙂, I have testing Temporal Fusion Transformer in the package, but I tend to run out of cuda memory.

OutOfMemoryError: CUDA out of memory.

Is there a parameter to limit that or any recommendations on how to avoid that?

Hi Team, I am facing below issue while doing .fit, I have installed all the requirements/dependencies as per git+https://github.com/nixtla/neuralforecast.git@main nf = NeuralForecast(models=models, freq='H') nf.fit(df=df, static_df=static_df)

Hello team. Is there an implementation of the DeepAR Forecasting Algorithm in the neuralforecast ? Thanks!

Hi all, quick question. In the TCN model, if I followed well there is an assumption that past values of the target are used as predictors correct ? I was wondering if it could adapt to a form where the TCN provide prediction of the target, based only on historic covariates for example. Thank you very much !

Using NHITS I get monthly forecasts like this. As you can see there is yearly seasonality, but the minimum point in the forecast is too high. To handle cases like this in models like Prophet I would set the seasonality mode as "multiplicative" with respect to the trend. Do you have any general suggestions on how to set hyperparameters to get better results with NHITS? I have tried adding exogenous variables to indicate the month, but the prediction is even worse. Thanks

Hello. I receive the following error when running the NBEATS model with horizon = 1. RuntimeError: einsum(): subscript p has size 2 for operand 1 which does not broadcast with previously seen size 0 Is there a fix on that? Thanks!!

I'm trying to use

neuralforecast

with

hierarchicalforecast

for hierarchical reconciliation (I've also seen

HINT

, but the provided reconciliation methods,

MinTrace

and

BottomUp

, are not suitable for my case). Now the problem is this: since time series in the dataset have a hierarchical relationship and time series at the higher levels tend to have higher values (because they are the sum of the hierarchically lower time series), loss functions such as plain

RMSE

and

MAE

seem to be unsuitable because the errors at the lower levels of the hierarchy, being smaller in absolute value, tend to be penalized too little. A workaround might be to use scale invariant loss functions such as

SMAPE

, which being a percentage error might mitigate the problem. Another solution might be to give more weight in the loss function (e.g., RMSE) to hierarchically lower time series. Do you know if with neuralforecast (and in particular with

NHITS

) there is a way to give a higher weight to some unique_id during model training? Do you have any other ideas about this? I also tried using losses such as PMM, GMM and NBMM but they did not give good results. Thanks

Hey, it’s mentioned in this paper that N-BEATS could do zero-shot forecasting. Is this implemented in neural forecast?

Hello everyone. Thank you for your amazing work!!! I am facing an error using azure function and NHits model. Locally it works, but once published in the cloud, I have this message : « you tried to log -1 which is not currently supported. Try a dict or a scalar/tensor. ». It seems to be due to tensorboard that tries to log file without success. Do you know if there is a way to set log_tensorboard=False (it seems to be possible with darts implementation). Thank you very much for your help and once again amazing work!

Hello, got a quick question to ask. I am trying to use exogenous variables in NHITS but am not clear on the definition of historic vs. Future. I'm doing a daily fuel sales forecast, with price of our own fuel and competitors price as exogenous variables. Both own prices and competitors' prices are only known on the day of the sales itself. Are they considered historic or future? Also I'm not quite sure what future and historic variables mean from the electricity price example. Is there another example to explain what is considered historic/future?

Just thought of another question w.r.t hierarchical forecast network. If I have exogenous variables that are specific to the lowest level, how will that work with HINT? E.g. I have a hierarchy of supermarket sales by region made up of supermarket sales by individual stores. Individual stores have categorical exogenous variables that describe location (rural, urban). I'm not sure how that will aggregate at the higher level, does it allow for exogenous variables to be used only for the lowest level in the hierarchy?

Hello, I want to apply NHITS using multivariate variables. And I'm trying to use SHAP(KernelExplainer) for variable interpretation, but I get an error... Please let us know if SHAP cannot be applied to NHITS. Thanks!!

Hi everyone. I want to use Neuralforecast to predict the offset produced by a ground-based telescope in different time series, but these time series are unevenly spaced time series. Is it possible to use models (like AutoNBEATSx or AutoNHITS) with these kind of time series?

Hi everyone, just wanted to ask if autoNhits can handle time series of different lengths when fitting. For example when forecasting product sales, new products may be introduced at different times.

Hi everyone! Firstly, thanks for providing this excellent package – I hope to also contribute once I get a better grasp of the underlying code. My question is regarding the memory usage:

model = [RNN(h = test_length,

input_size = test_length*3,

scaler_type = 'robust')]

nf = NeuralForecast(models=model, freq='H')

nf.fit(df=df)

where test_length=1000. The training data consist of a single length 9000 time series. Using the default hidden layer numbers and sizes, a very rough estimate of the memory usage gives me: 1,280,000 parameters for 5.12 MB. Of course this would be multiplied by a few times for forward and backward pass memory and storing optimizer state in order to obtain total memory usage. What I don’t understand is that when I run:

nf.predict()

I receive the error message “OutOfMemoryError: CUDA out of memory. Tried to allocate 6.71 GiB (GPU 0; 11.75 GiB total capacity; 7.10 GiB already allocated; 3.88 GiB free; 7.13 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.” I have checked to make sure there is negligble GPU memory usage before executing nf.predict(). What do you think might be causing this issue? I am happy to provide the dataframe df I used if needed. I should also mention that this error does not occur when I restrict the RNN to single encoder/decoder hidden layers of size 100, rather than the default double layers of size 200.

hello i am using neuralforcast and i want to know if there way to add earlystopping or the library contains that?

Hey Guys, Joined the community recently and I look forward to learning from the discussions. I am starting a Time Series project with NeuralForecast where I am trying to train a model which will be used for One-shot learning on other time series (Transfer Learning). I read the article (https://nixtla.github.io/neuralforecast/examples/transfer_learning.html) and now I have two questions: 1. I also read this article which says I need to provide data in a long format and there needs to be a "unique_id" associated with each time series. (https://nixtla.github.io/neuralforecast/examples/getting_started_complete.html). The long format is fine and there would be unique ids to distinguish between different time series but while predicting new data there would be new unique IDs, so would that affect the model? So essentially does the model learn unique_ids as well? 2. If there is a big dataset with thousands of time series then there would be too many samples for the model so is there a way in which the model only uses some 'x' number of samples from each series? I came to know about the "max_samples_per_ts" hyperparameter which does the same in the Darts library ( https://unit8co.github.io/darts/examples/14-transfer-learning.html).

With AutoNHITS, is there any way to include the loss (with a different val_loss) in the search space as well? For example, I would like to search for the best loss among GMM, PMM and NBMM (also optimizing their "n_components" parameter, e.g. between 1 and 10) so as to minimize the sCRPS validation loss. Thanks.

i have questions : do we treat the problem of non stationary time series? in the library also why we put the input size -1 as default

Hi there! We want to compare forecasting methods including some that are not implement in neuralforecast and want to treat these on a similar basis. One of the key aspects is to ensure that the hyperparameter tuning is done in a consistent manner. For the methods in neuralforecast we are using the automatic tuning functionality to find good hyper parameters and we would like to apply the same/similar methodology to the other methods. Good you provide some information/link on the hyper-parameter tuning policy? For example, how are training/validtion/test sets selected and processed? Or do you tune based on backtesting results?

Does Neuralforecast provide something similar to Neuralprophet's Global-Local model where a global model can be used along with local models for some components such as trend and seasonality (https://neuralprophet.com/how-to-guides/feature-guides/global_local_modeling.html)? Thanks