hello , do I need to make some normalize for my data before using model from neural forecast models ?

Anyone seen this before?

happens when trying to fit HINT/AUTOTFT

ah narrowed it down to number of trials, running fine with 1, but kind of defeats the purpose of auto xD

@Cristian (Nixtla) I think this documentation is missing a key step, no?

initialization of the NeuralForecast class with the model specified

Performing a fit of say 20 epochs or performing two fits of 10 epochs each is not equivalent because even though the network weights are preserved, the adaptive optimizer (Adam) state is not preserved between two different fits. Do you know if there is any way to preserve the optimizer state so that doing two consecutive fits of 10 epochs is equivalent to doing a single fit of 20 epochs?

Hi everyone, I'm currently working on a macroeconomic time series forecasting project at the Central Bank of Colombia. Our goal is to forecast the IPC (Consumer Price Index) based on the GDP (Gross Domestic Product) data. However, I am facing a challenge in introducing the GDP variable as an exogenous variable since it is observed on a quarterly basis. My question is, how can I handle this issue? thanks 🙂

does any one try transformer models and get killed after running ?

does any one know this reference where i try to find it in NHits paper it is mentions there but without referance Wu et al. 2021?AutoFormer model and also informer model @Cristian (Nixtla)

exogs dont need to be specified on the auto's? i see there's no parameter for them

Hi everyone, I'm hoping to pick you brains on this business problem I am trying to solve and the best modeling approach based on the tools you have all developed. I have a hierarchical system with roughly 130 daily timeseries. Although these timeseries follow a hierarchical structure, we are not interested in their mean forecast. We are more interested in their 95th quantile forecast. To add to the difficulty, we only have about 2.5 years ~ 914 days of historical data, yet we want to forecast 18 months in the future. horizon = 18*30 = 540 days. 1. Currently, I am trying to use a NHITS base network where the validation loss is

sCRPS

with quantiles = [0.95] with a HINT class. My hope is that reconciling the distributions across the hierarchy might result in better estimates of the 95th quantiles. 2. Since the 95th quantiles themselves can't be reconciled. i.e. the sum of the 95th quantiles of the bottom nodes is not equal to the 95th quantile of the parent nodes. Am I better off getting rid of the HINT layer altogether? 3. Can anyone provide a heuristic guide to tuning NHITS? I want to get a feel for the hyper-parameter space before I use Ray to fine tune a model?

How can I turn off the following output when calling the predict methods? Thanks The default for

verbose

is already

False

but I would prefer to suppress it further.

def predict(
        self,
        df: Optional[pd.DataFrame] = None,
        static_df: Optional[pd.DataFrame] = None,
        futr_df: Optional[pd.DataFrame] = None,
        sort_df: bool = True,
        verbose: bool = False,
        **data_kwargs,
    ):

I'm trying to create a new issue on Github and it's not letting me for some reason. I'm trying to load a HINT model and I'm getting the error:

---------------------------------------------------------------------------
KeyError                                  Traceback (most recent call last)
Cell In[32], line 1
----> 1 nf2 = NeuralForecast.load(path='./results/model5/')

File ~/pytorch_forecasting/silvertorch/pytorch/lib/python3.10/site-packages/neuralforecast/core.py:693, in NeuralForecast.load(path, verbose)
    690 for model in models_ckpt:
    691     model_name = model.split("_")[0]
    692     models.append(
--> 693         MODEL_FILENAME_DICT[model_name].load_from_checkpoint(f"{path}/{model}")
    694     )
    695     if verbose:
    696         print(f"Model {model_name} loaded.")

KeyError: 'hint'

hello , do i need to make my time series stationary before do the exogenous variable ? ?

Hello everyone, I'm continuing my experiments with DeepAR, and I have a few questions for you... Putting aside my suspicion of bug with the Negative Binomial, the performance I'm getting is still below the GluonTS implementation. One possibility would be the absence of lags_seq option. DeepAR being autoregressive, it's a waste to do without lag 1, for example. Or am I missing something? Second question, GPU predictions with mixtures mech loss (like PMM) require far too much GPU RAM (+90GO in my case). I suppose that by playing with

inference_windows_batch_size

I should be able to handle it, but does that mean that for inference DeepAR doesn't keep only the last

input_size

points as it does for training? Thanks in advance!

Is there a way to set up a validation set in neural forecast in which instead of testing on a portion of time, we instead test on different set of timeseries? I'm in a situation where I have many short timeseries. I'd like to use a validation set to tune my model but can't really afford using temporal cross validation. If instead I could train on a subset of timeseries and validation on another, I think I could maximize the value of my data better.

Hey, I’m trying to understand how BaseWindows, used by N-Beats and NHITS generates training examples. In particular I’m trying to understand what’s happening in

training_step

,

_create_windows

and

_parse_windows

. I’m a bit confused by the code. Is there a paper or some doc that describes the idea of what’s done there?

hello , How can i determine the cross validation type ? in the code ?

Hi, is there a way to visualize the convergence of a model i.e. plot the loss per training step after or during training?

are we using the expanding window or without expanding i mean the training data it will be the same size for each fold ?

TLDR:

scaler_type

is behaving weirdly for the validation set loss functions. If it's not

identity

. The validation loss is way off. What is going on? I'm observing some really weird behavior out of the NBEATS model and I wanted to see if I am out to lunch or what. To get a sense of the model landscape. I started small. Out of my dataset of 100 timeseries, I wanted to find a basic set of hyper-parameters where I could overfit 10 timeseries really well. Once I felt comfortable with the complexity and relative space of model parameters, I would reduce the complexity, add regularization and try to get a model which could perform well on a validation set. After some trial and error I found I could overfit my data very well with this model configuration. I plotted the training loss there.

model_name = "NBEATSx"
stacks = 2
params = dict(
    h=180,                         
    input_size=360,              
    loss=HuberLoss(),                
    max_steps=2000,                          
    dropout_prob_theta=0.0,
    stack_types=stacks*["identity"],
    n_blocks=stacks*[15],
    mlp_units=[[16, 16] for _ in range(stacks)], 
    scaler_type="standard",                 
    learning_rate=1e-3,
    random_seed=200,
    alias=model_name,
    batch_size=5,
)
nf = NeuralForecast(models=[NBEATSx(**params)], freq="D")
nf.fit(df=Y_train_df.reset_index())

Epoch 999: 100%
2/2 [00:00<00:00, 6.45it/s, v_num=788, train_loss_step=0.0128, train_loss_epoch=0.0122]

Next, I wanted to check the performance on a validation set. The main issue is no matter how I try to reduce the complexity. example: changing the number of stacks from 2 -> 1or number of blocks from 15 -> 1. or even number of mlp_units from 16 -> 2. Or adding dropout, The validation loss of MAPE or any other loss for that matter does not decrease significantly at all!!! See the second plot. The scale makes it look like it's decreasing but it gets to 1 - 1e-6.

stacks = 1
params = dict(
    h=MODEL_HORIZON,                         
    input_size=CONTEXT_LENGTH,              
    loss=HuberLoss(),   
    max_steps=2000,                          
    dropout_prob_theta=0.5,
    stack_types=stacks*["identity"],
    n_blocks=stacks*[1],
    mlp_units=[[16, 16] for _ in range(stacks)], 
    scaler_type="standard",                 
    learning_rate=1e-3,
    random_seed=200,
    alias=model_name,
    batch_size=5,
    # Validation params
    val_check_steps=1,
    valid_loss=MAPE(),
    early_stop_patience_steps=3000
)

However, the biggest change comes from changing the

scaler_type

. If I change it to

identity

. At least the MAPEs are in a decent range and I see some decrease and typical curve behavior. What is going on?

A few months ago I asked how time series shorter than input_size were handled (whether they were padded or ignored) and was told that zero-padding was used. @Cristian (Nixtla) Could you explain what changed with this commit https://github.com/Nixtla/neuralforecast/pull/715 ? Thank you very much

Quick question regarding validation, I dont want to use cross-validation, but just train on 8 months and validate on one month. Is this possible? Can I insert my validation dataframe to fit? Something like this:

nf.fit(df_train, df_val)

Is there target_transform for neuralforecast? Target transformation has been implemented in mlforecast.

hello triad.collections.schema.SchemaError: Invalid syntax: 'unique_idstring,metricstring,cutofftimestamp[us],AutoNHITSdouble,AutoMLP:double' Expecting ',' delimiter: line 1 column 61 (char 60) i am getting this error when using the accuracy function from the library does any one this happened with you ? thank you

Hi, I ran the below AutoNHITS model using neuralforecast version == 1.5.0: The results look fine.

from ray.tune.search.basic_variant import BasicVariantGenerator
models = [AutoNHITS(h=61,
          config=nhits_config,
          loss=MAPE(),
          search_alg=BasicVariantGenerator(random_state=42),
          num_samples=100)]
nf = NeuralForecast(
  models=models,
  freq='D')
Y_hat_df = nf.cross_validation(df=Y_df, val_size=val_size,
                test_size=test_size, n_windows=None)

When I run the same AutoNHITS model using neuralforecast version 1.6.2 (the latest version), the train loss step, train loss epoch for all configs turn to zero. The predictions are also zero. Did something change in the latest version? I have attached the pics of the same. Thanks.

Hi everybody, I am new here and I might be asking a question which was already asked. I couldn't find any matches though, so I hope you'll forgive me. Here is the question: I am intending to develop a time series model in the fashion of our old econometric "distributed lags models". That means, that the forecast (even a nowcast) ist done only with past values of other variables. The own time series is not available for prediction. LSTMs and TCNs can handle it without difficulty but I would like to implement a TFT and a nbeats for I have the feeling that the structures might be interesting for the task. Any hint on how to move on will be much appreciated!

Awesome! I will try it asap. Thank you very much for the quick reply!

Hey everyone, I trained a model on a work cluster with a GPU. When I try to load it on my local machine (CPU only). I get the following error message:

nf2 = NeuralForecast.load(path='../cluster_results/results/b79f6fd0c50099a7519da800805bf436d4f4f4a6/')

RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

I'm trying to debug this error on my own. I found this stackoverflow post. https://stackoverflow.com/questions/56369030/runtimeerror-attempting-to-deserialize-object-on-a-cuda-device But I think this needs to be handle as part of the load function?https://github.com/Nixtla/neuralforecast/blob/1d78f503ac4cd0fdfa7aae9156a876fe94eb1db4/neuralforecast/core.py#L665 I think this can be solved if you have an extra argument to the load function

map_location

MODEL_FILENAME_DICT[model_name].load_from_checkpoint(f"{path}/{model}",  map_location=map_location)

If I specify

map_location=torch.device('cpu'))

, it works