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

Dear Nixla Team, could you please help me with some questions regarding Temporal Normalization and Loss Functions: 1. Temporal Normalization will determine e.g. min and max on the past data points and ignore future data points so we don’t have look ahead bias, correct? 2. Are predicted values scaled back to the original domain before loss is calculated and back propagation happens? 3. If I train on two time series, 1. with magnitude of ~ 1000 and 2. with magnitude of ~0.1, both with the same relative errors. MAE will cause back propagation to update the weights more to suit time series 1, no matter what temporal normalization I use, correct?

Hi, in the documentation I can read that the loss can be "PyTorch module, instantiated train loss class from losses collection" - is there any way to define your own custom metric or it must be one of the predefined metrics from your losses collection? I currently use custom metric which is a subclass of

torchmetrics.metric.Metric

and thinking if it would be possible to use that.

Hi there. I am trying to train the nhits and autofomer models, but I have a few questions and am seeking assistance. 1. It seems that futr_exog_list and hist_exog_list are not working properly. For instance, when I give the futr_exog_list argument in nhits, all prediction results come out as nan values. 2. While training on the GPU, if GPU memory is insufficient, is it a correct decision to increase the training iterations when adjusting batch size or windows_batch_size? 3. I am curious if having data with around 40,000 records and column counts ranging from 16 to 20 is a suitable amount for training nhits or autofomer models. Thank you for checking my lengthy question. If you have any information on this, I would appreciate it if you could share it. Have a great day.

Dear Nixla Team, I have a question regarding normalization/scaling of different features. If I select a model that allows exogenous variables, say NBEATSx and I select a Scaler say minmax1. Is this Scaler applied to y only or to the exogenous vars aswell? Exogenous vars maybe of very different type e.g. int or booleans that I would not like to be normalized at all (because they have time invariant meaning that I do not want messed up) or time series, which I would like to have normalized. Is there a way to control how each input is normalized?

Hi all, Firstly, thank you for the wonderful library that makes forecasting with NN so much easier, especially with the Auto models. I am familiar with classifications and regressions, and am very new to time series forecasting. I spent the past 2 days reading the documentation and have the following questions below.My dataset contains 22 time series of different lengths, and with 100+ numerical features, 20 categorical features and 10 binary features. It is weekly data where every Monday is the start of week i.e. soweekdate and this is the 'ds'. The target variable is the Sales_Qty, which is what I am trying to forecast 8 weeks ahead and is the 'y'. I intend to perform feature selection using Light GBM regression tree. 1. If I am using the Auto models, can I specify the horizon, input size, futr_exog_list, hist_exog_list, stat_exog_list and etc in the configs of the Auto Models on top of the required model parameters? 2. Since I am forecasting 8 weeks ahead, I change the freq="W" in nf = NeuralForecast(models=models, freq='M')? 3. On scaler_type, what is the default for the Auto models? When it scales, does it do it for individual time series in the group or do it for all columns as a group? Is there a way to turn this off as I might want to scale by individual time times outside of Neural Forecast. 4. The 20 categorical features are time invariant so I will be creating dummy variables for them and putting them into the static_df. As I am forecasting 8 weeks ahead and I have both historical and future values for the 100+ numerical features. Should I put these numeric features into the hist_exog_list, futr_exog_list and futr_df accordingly? The column names might be the same but the underlying values are different for each id and each ds. 5. As for the 10 binary features, they are in the similar situation as in the numeric features where I have both historical and future values for each time series, should I also include these binary features into the hist_exog_list, futr_exog_list and futr_df accordingly? Thanks in advance, Will

Dear Nixtla-Team, in your Exogenous Variable Notebook (https://github.com/Nixtla/neuralforecast/blob/main/nbs/examples/Exogenous_Variables.ipynb) you use NHITS but from the NHITS Paper (https://arxiv.org/pdf/2201.12886.pdf) I understand that it does not make use of Exogenous Variables: "Our model, operating in the univariate regime and accepting only the predicted time-series’ history,". Have you just used the wrong model to demonstrate Exogenous Vars or am I misunderstanding the paper?

Dear Nixtla-Team, just starting with your fantastic library with a simple LSTM model: I have assured that the column 'ds' has a datetime.date format (pd.to_datetime). While the simple training (nf.fit) works, I get an error during training if I do a cross validation:

nf.cross_validation(df=Y_train_df, n_windows=10,step_size=1,val_size=horizon,sort_df=True,verbose=True)

`ValueError: You are trying to merge on datetime64[s] and object columns. If you wish to proceed you should use pd.concat`Actually I was not trying to merge anything at that point: is the method expecting another data format for 'ds' different from the simple fit? Thank you!

Dear Nixtla-Team, is it somehow possible to run a rolling prediction just like cross validation does without refitting the model? For example I want to load a model from file and evaluate its performance on some set of new time series each containing far more data points that the horizon of the model. I am already considering hacking up something myself from the core.py predict and crossvalidation methods.

I am trying to apply NHiTS to some new data I have but I get an error of

Out of bounds nanosecond timestamp: 2385-12-31 00:00:00

when I do nf.predict(futr_df=df_test). But my df_test only has 364 days in it with dates going from 2022-01-03 to 2023-01-01. There is no date even close to 2385! Also, the 'ds' column is properly converted to datetime type. It's weird. I am thinking that 2385 might be coming from adding a date to the last point in an input batch? I'm not sure. There is a line in the error stack pointing to core.py line 109 dates=np.tile(last_date_f(last_dates[0], len(dataset)). Has anyone seen this type of error before?

HI <!channel>, I am trying to use AutoLSTM with BayesOptSearch() and I don't know what ray tune function to use in config. I tried to use tune.randn and had this error: "OverflowError: Range exceeds valid bounds" and i can't solve it even if I use a very small SD, and I tried also tune.uniform and had this error: "TypeError: 'numpy.float64' object cannot be interpreted as an integer". Does anyone know how to solve this?

Hi Is there any guide on how to include a list of exogenous variables for transfer learning? And is it possible to make the model flexible enough such that the pre-trained model can receive varying number of exogenous variables? For example, during training, I may train it with 5 exogenous variables. However when predicting the model, for reasons like missing data, I may only have up to 3 exogenous variables while the other 2 are missing completely. https://nixtla.github.io/neuralforecast/examples/transfer_learning.html

I have a question about RNN Models do these models in the training stage predict one by one or all the sequences at once?

I'm attempting to modify the code here (https://nixtla.github.io/neuralforecast/examples/hierarchicalnetworks.html#fit-and-predict-hint) to also use a known covariate. To test, I modified the example code to have a random feature:

Y_df['noise'] = np.random.rand(len(Y_df))


# split to train and test
Y_train_df = Y_df[:-horizon]
Y_test_df = Y_df[-horizon:]
# Instantiate HINT
# BaseNetwork + Distribution + Reconciliation
nhits = NHITS(h=horizon,
           input_size=24,
           loss=GMM(n_components=10, level=level),
           hist_exog_list=['month'],
           futr_exog_list=['noise'],
           max_steps=2000,
           early_stop_patience_steps=10,
           val_check_steps=50,
           scaler_type='robust',
           learning_rate=1e-3,
           valid_loss=sCRPS(level=level))

model = HINT(h=horizon, S=S_df.values,
         model=nhits,  reconciliation='BottomUp')

# Fit and Predict
nf = NeuralForecast(models=[model], freq='MS')
nf.fit(df=Y_train_df, val_size=12)
Y_hat_df = nf.predict(Y_test_df)

When I run this code, I receive the error: "RuntimeError: normal expects all elements of std >= 0.0" Has this been seen before? Not sure exactly what's going on. Everything works fine when I comment out futr_exog_list

Has anyone benchmarked NHiTS performance on 1 vs. multiple GPUs? I am comparing it between EC2 g3.4xlarge (1 GPU) and g3.8xlarge (2 GPUs). The runtime slightly increases when running with more GPUs! I am testing on one time series right now. Maybe training for only one time series is not parallelized across multiple GPUs (i.e. I need to add more time series to see an effect)?! Also, I was first running in a Jupyter notebook and thought maybe notebooks have an issue with multiple GPUs but I see similar results out of a python script.

I have just started to use it and thanks for the really nice package. I am running the "end to end walkthrough" on my windows with PC with my own data which only has 6 unique_id and ~4000 data points/ID. After calling the nf.fit(df=Y_df) I can see all the progress but the ray tuner was reporting sync errors not being able to sync the files to the desired places (could not upload checkpoint). In the end, it was reporting the .json file was not the correct format (it must be a path to an experiment checkpoint file). There are sub-folders created in the tune folder according to the num_samples and the final json files seems to contain some info of all the experiments. What went wrong with my setup? I have already updated ray and will that be the issue windows instead of linux? Any suggestion is appreciated. 🙂

How can I update the

storage_path

for the

AutoNHITS

or

HyperOptSearch

? I'm trying to get the

ray

or

tune

working with

neuralforecast

and I keep getting this error below.

(_train_tune pid=21000) 2023-09-13 08:02:48,610 ERROR syncer.py:466 -- Caught sync error: Sync process failed: GetFileInfo() yielded path 'C:/Users/bubu/ray_results/_train_tune_2023-09-13_08-02-34/_train_tune_7a885f1f_1_h=12,input_size=24,learning_rate=0.0032,loss=ref_ph_de895953,max_steps=10,n_freq_downsample=24_12_1,n_pool__2023-09-13_08-02-39/error.pkl', which is outside base dir 'C:\Users\bubu\ray_results\_train_tune_2023-09-13_08-02-34\_train_tune_7a885f1f_1_h=12,input_size=24,learning_rate=0.0032,loss=ref_ph_de895953,max_steps=10,n_freq_downsample=24_12_1,n_pool__2023-09-13_08-02-39\'. Retrying after sleeping for 1.0 seconds...
(_train_tune pid=21000) 2023-09-13 08:02:49,617 ERROR syncer.py:466 -- Caught sync error: Sync process failed: GetFileInfo() yielded path 'C:/Users/bubu/ray_results/_train_tune_2023-09-13_08-02-34/_train_tune_7a885f1f_1_h=12,input_size=24,learning_rate=0.0032,loss=ref_ph_de895953,max_steps=10,n_freq_downsample=24_12_1,n_pool__2023-09-13_08-02-39/error.pkl', which is outside base dir 'C:\Users\bubu\ray_results\_train_tune_2023-09-13_08-02-34\_train_tune_7a885f1f_1_h=12,input_size=24,learning_rate=0.0032,loss=ref_ph_de895953,max_steps=10,n_freq_downsample=24_12_1,n_pool__2023-09-13_08-02-39\'. Retrying after sleeping for 1.0 seconds...

It seems this is a known bug and a suggested workaround is to predefine the storage path as below https://github.com/ray-project/ray/issues/36448#issuecomment-1673274452

from ray import air, tune

tuner = tune.Tuner(
    train_fn,
    run_config=air.RunConfig(storage_path="./")
)
tuner.fit()

However, I'm not sure where to define

run_config=air.RunConfig(storage_path="./")

with

AutoNHITS

Really appreciate any tips on this issue, thanks!

Hi Nixtla Team. Thanks for developing a wonderful library for forecasting. I am getting an error when I try to fit AutoLSTM and AutoNHITS. The error is as follows: "

--> 168             raise RuntimeError(error_msg)
    169 
    170         return self._trial_to_result(best_trial)

RuntimeError: No best trial found for the given metric: loss. This means that no trial has reported this metric, or all values reported for this metric are NaN. To not ignore NaN values, you can set the

filter_nan_and_inf

arg to False." My config looks like this: %%capture from ray import tune config_nhits = { "input_size": tune.choice([4, 8]), # Length of input window "start_padding_enabled": True, "n_blocks": 3*[1], # Length of input window "activation": "relu", "mlp_units": 3 * [[3, 3]], # Length of input window "n_pool_kernel_size": tune.choice([3*[1], 3*[2], 3*[4], [8, 4, 2, 1, 1]]), # MaxPooling Kernel size "n_freq_downsample": tune.choice([[8, 4, 2, 1, 1], [1, 1, 1, 1, 1]]), # Interpolation expressivity ratios "learning_rate": tune.loguniform(1e-4, 1e-2), # Initial Learning rate "scaler_type": tune.choice(['robust', 'standard']), # Scaler type "max_steps": tune.choice([500, 1000]), # Max number of training iterations "batch_size": tune.choice([1,2,4,8,16,32,64,128]), # Number of series in batch "windows_batch_size": tune.choice([1,2,4,8,16,32,64,128]), # Number of windows in batch "stack_types": ['trend', 'seasonality'], "random_seed": tune.randint(1, 20), # Random seed "hist_exog_list": ['VOICE_TRAFFIC','SUCCESSFULL_CALLS'] } config_lstm = { "input_size": tune.choice([4,8]), # Length of input window "encoder_hidden_size": tune.choice([4,8,16,32,64,128]), # Hidden size of LSTM cells "encoder_activation": tune.choice(['relu']), "encoder_n_layers": tune.choice([1,2,3]), # Number of layers in LSTM "learning_rate": tune.loguniform(1e-4, 1e-2), # Initial Learning rate "scaler_type": tune.choice(['robust', 'standard']), # Scaler type "max_steps": tune.choice([100, 200]), # Max number of training iterations "batch_size": tune.choice([2,4,8,16,32,64,128]), # Number of series in batch "random_seed": tune.randint(1, 20), # Random seed "hist_exog_list": ['VOICE_TRAFFIC','SUCCESSFULL_CALLS'] } config_nbeatsx = { "input_size": tune.choice([4,8]), # Length of input window "learning_rate": tune.loguniform(1e-4, 1e-2), "scaler_type": tune.choice(['robust', 'standard']), "max_steps": tune.choice([100, 200]), "n_polynomials": 0, "n_blocks": [1,1], "stack_types": ['trend', 'seasonality'], "dropout_prob_theta": 0.5, "mlp_units": [[4,4],[8,8],[16,16],[32,32]], "random_seed": tune.randint(1, 20), "hist_exog_list": ['VOICE_TRAFFIC','SUCCESSFULL_CALLS'] } #config_tft = { # "input_size": tune.choice([4,8]), # "hidden_size": tune.choice([8, 16]), # "learning_rate": tune.loguniform(1e-4, 1e-2), # "scaler_type": tune.choice(['robust']), # "max_steps": tune.choice([250, 500]), # "batch_size": tune.choice([8, 16]), # "random_seed": tune.randint(1, 20), # "hist_exog_list": ['VOICE_TRAFFIC','SUCCESSFULL_CALLS'] My Prediction Code looks like this: import numpy as np import pandas as pd import pytorch_lightning as pl import matplotlib.pyplot as plt from neuralforecast import NeuralForecast from neuralforecast.auto import AutoLSTM,AutoNHITS,AutoTFT, AutoNBEATSx from neuralforecast.losses.pytorch import MQLoss,MAE,HuberMQLoss from ray.tune.search.hyperopt import HyperOptSearch levels=[80,90] h = Y_test_df['ds'].nunique() #search_alg=HyperOptSearch(metric="loss", mode="min"), modelNHITS = AutoNHITS(h=h, loss=MAE(), config=config_nhits, num_samples=50, cpus=1) modelLSTM = AutoLSTM(h=h, loss=MAE(), config=config_lstm, num_samples=50, cpus=1) #modelNBEATSx = AutoNBEATSx(h=h, config=config_nbeatsx, loss=MAE(),valid_loss=MAE(), num_samples=50, cpus=1) #modelTFT = AutoTFT(h=h, config=config_tft, loss=MAE(),valid_loss=MAE(), num_samples=50, cpus=1) nf = NeuralForecast(models=[modelNHITS,modelLSTM],freq='W') nf.fit(df=Y_train_df) #Y_hat_df = nf.predict() Y_hat_df = nf.predict_insample(step_size=h) #Y_hat_df = Y_hat_df.reset_index(drop=False).drop(columns=['unique_id','ds']) #plot_df = pd.concat([Y_test_df, Y_hat_df], axis=1) #plot_df = pd.concat([Y_train_df, plot_df]) #plot_df = plot_df[plot_df.unique_id=='WEEKLY'].drop('unique_id', axis=1) #plt.plot(plot_df['ds'], plot_df['y'], c='black', label='Actual') #plt.plot(plot_df['ds'], plot_df['AutoNHITS'], c='blue', label='NHITS-Prediction') #plt.plot(plot_df['ds'], plot_df['AutoLSTM'], c='purple', label='LSTM-Prediction') ##plt.plot(plot_df['ds'], plot_df['AutoTFT-median'], c='green', label='AutoTFT-Prediction') ##plt.plot(plot_df['ds'], plot_df['AutoNBEATSx'], c='Green', label='NBeatsx-Prediction') ##plt.fill_between(x=plot_df['ds'][-40:], ## y1=plot_df['AutoLSTM-lo-90'][-40:].values, ## y2=plot_df['AutoLSTM-hi-90'][-40:].values, ## alpha=0.4, label='level 90') #plt.legend() #plt.grid() #plt.plot() plt.figure(figsize=(10, 5)) plt.plot(Y_hat_df['ds'], Y_hat_df['y'], label='True') plt.plot(Y_hat_df['ds'], Y_hat_df['AutoNHITS'], c='blue', label='NHITS-Forecast') plt.plot(Y_hat_df['ds'], Y_hat_df['AutoLSTM'], c='purple', label='LSTM-Forecast') #plt.plot(Y_hat_df['ds'], Y_hat_df['AutoNBEATSx'], c='green', label='NBeatsx-Prediction') plt.axvline(Y_hat_df['ds'].iloc[-20], color='black', linestyle='--', label='Train-Test Split') plt.xlabel('Timestamp [t]') plt.ylabel('Demand for Voice Services(Call Attempts)') plt.grid() plt.legend() Can someone please guide me what I am doing wrong here? Any help/guidance will help me..

Has anyone ever received the attached error when using predict_insample? My "big" dataframe has columns=["unique_id", "ds", "y"] with about 150 unique "unique_id"s, the smallest time-series has ~30 datapoints, the biggest time-series has ~1800 datapoints, ~90k datapoints in total. No NaNs. No exogenous covariates.

Does anyone know why `neuralforecast`'s

TemporalNorm

uses one set of normalization per sequence (basically InstanceNorm for sequence data without learnable parameters)? I would have expected the offset and scale to be calculated over the entire training dataset and kept constant for every sequence @Cristian (Nixtla)

Thanks for the great framework you implemented. Here's a small issue. I want to run the program on one GPU while the program calls all GPUs. I tried to set

accelerator="cuda:0"

but errors were raised. Any solutions except setting "CUDA_VISIBLE_DIVICE" in the command line?

Hello Sorry in advance if my questions are really basic for some of you : 1/ while using a pretrained model (transfer learning), I am not able to pass horizon h with predicton function, why ? use case : how can I produce a prediction for next 36 months if the trained models was trained on 24 past months ? 2/ may I finetune a model with exogenous variables while the trained model was not trained with exogenous variables ? Thank you

I got an issue while training M4 Monthly dataset with default nhits config : raise ValueError("batch_size should be a positive integer value, " ValueError: batch_size should be a positive integer value, but got batch_size=7.0 here hyperparameter config nhits_config = { "input_size": tune.choice([horizon * x for x in [1, 2, 3, 4, 5]]), "h": None, "n_pool_kernel_size": tune.choice( [[2, 2, 1], 3 * [1], 3 * [2], 3 * [4], [8, 4, 1], [16, 8, 1]] ), "step_size":tune.choice([1, horizon]), "n_freq_downsample": tune.choice( [ [168, 24, 1], [24, 12, 1], [180, 60, 1], [60, 8, 1], [40, 20, 1], [1, 1, 1], ] ), "learning_rate": tune.loguniform(1e-4, 1e-1), "scaler_type": tune.choice([None, "robust", "standard"]), "max_steps": tune.quniform(lower=500, upper=1500, q=100), "batch_size": tune.qloguniform( lower=5, upper=9, base=2, q=1 ), # [32, 64, 128, 256] "windows_batch_size": tune.qloguniform( lower=7, upper=10, base=2, q=1 ), # [128, 256, 512, 1024] "loss": None, "random_seed": tune.randint(lower=1, upper=20), } it seems that i am not the only one : https://github.com/Nixtla/neuralforecast/issues/716 Any idea on what could create this issue ?

Hello. Sometimes cross_validation produces NaN forecasts and the best model has loss = 0.0 when using Recurrent Auto* Models. My TimeseriesDataset has 54 unique time-series and every time-series is exactly of length 800. There are no NaNs in my data. I use h=1, step_size=1, AutoRNN + AutoLSTM models, and the configs are the default configs. num_samples=50. It seems like sometimes the combination of hyperparameters leads to loss=0.0, as can be seen in the screenshots. The last two screenshots are the two best hyperparameter combinations (loss=0.0, next ist loss=43.91..) extracted from the ray ResultGrid. I also checked all the other samples where

input_size=-1 && inference_input_size=-1

but they have loss is >0.0 so it must be something else. Any help is appreciated!

I have a model understanding question concerning the NHITS model. I am hoping to get some clarifications about the connection between two hyper-parameters:

n_freq_downsample

and

n_pool_kernel_size

. My question may also stem from how these parameters interact with one another. From my understanding, the higher the value in

n_freq_downsample

, the less datapoints we will have after downsampling. i.e. if we have daily data and we downsample with

n_freq_downsample = 7

, we will have 1/7 the amount of data.

n_pool_kernel_size

also has a "downsampling effect", if we have an input size of

N

and

n_pool_kernel_size = 2

, then we will now have

N / 2

datapoints. First of all, is this understanding correct? And if so, it seems like we would want these two hyperparameters be inversely proportional. I would not want to downsample by a large number and then apply large kernel size pooling layer. These would dramatically decrease the amount of information being fed into my model? The reason I am asking this question is because I was looking at the AutoNHITS parameter space

"n_pool_kernel_size": tune.choice(
            [[2, 2, 1], 3 * [1], 3 * [2], 3 * [4], [8, 4, 1], [16, 8, 1]]
        ),
        "n_freq_downsample": tune.choice(
            [
                [168, 24, 1],
                [24, 12, 1],
                [180, 60, 1],
                [60, 8, 1],
                [40, 20, 1],
                [1, 1, 1],
            ]
        ),

Intuitively, I would have assumed that the last two choices for

n_pool_kernel_size

would have been reversed. I.E.

[1, 4, 8], [1, 8, 16]

?

Hi everyone, I am facing an issue training all kinds of

Auto

models. It starts training and gets stuck, and never recovers. I am using Python 3.10.1 on Ubuntu 20.04.6 LTS.

In [1]: from neuralforecast import NeuralForecast                                                                                                     
                                                                                                                                                      
In [2]: from neuralforecast.auto import AutoNBEATS
                                                                           
In [3]: from neuralforecast.utils import AirPassengersDF
   ...:                                                                                                                                               
   ...: Y_df = AirPassengersDF
   ...: Y_df.head()
Out[3]: 
   unique_id         ds      y
0        1.0 1949-01-31  112.0
1        1.0 1949-02-28  118.0
2        1.0 1949-03-31  132.0
3        1.0 1949-04-30  129.0
4        1.0 1949-05-31  121.0

In [4]: models = [AutoNBEATS(h=12)]

In [5]: nf = NeuralForecast(models=models, freq='M')

In [6]: nf.fit(df=Y_df)
2023-09-25 18:10:00,972 INFO worker.py:1621 -- Started a local Ray instance.
2023-09-25 18:10:01,764 INFO tune.py:226 -- Initializing Ray automatically. For cluster usage or custom Ray initialization, call `ray.init(...)` befor
e `Tuner(...)`.
2023-09-25 18:10:01,767 INFO tune.py:666 -- [output] This will use the new output engine with verbosity 0. To disable the new output and use the legac
y output engine, set the environment variable RAY_AIR_NEW_OUTPUT=0. For more information, please see <https://github.com/ray-project/ray/issues/36949>
+--------------------------------------------------------------------+
| Configuration for experiment     _train_tune_2023-09-25_18-09-58   |
+--------------------------------------------------------------------+
| Search algorithm                 BasicVariantGenerator             |
| Scheduler                        FIFOScheduler                     |
| Number of trials                 10                                |
+--------------------------------------------------------------------+

View detailed results here: <$USER>/ray_results/_train_tune_2023-09-25_18-09-58
To visualize your results with TensorBoard, run: `tensorboard --logdir <$USER>/ray_results/_train_tune_2023-09-25_18-09-58`

(_train_tune pid=401980) Global seed set to 10
(_train_tune pid=401980) [rank: 0] Global seed set to 10
(_train_tune pid=401980) Initializing distributed: GLOBAL_RANK: 0, MEMBER: 1/3

It does not progress beyond this point. Issuing

$ ps aux | grep ray

points out to multiple

ray::IDLE

processes. Any help regarding this would be great! Thanks.

Hi, is it possible to load a huggingface model weight for prediction ?

I am trying to test HuberLoss() using the latest neural forecast version. However, I'm running into an error. Can you please help me resolve it? Thanks!