This message was deleted.

Hi @Layon Hu, thanks for using our library! I see that you are working on long-horizon, but you set

horizon=1

. The model is only learning to forecast the next timestamp, and could explain the behavior on the plot. With

cross_validation

the forecasts are rolled, using the latest true value available (t-1 in this case because horizon is 1.

Thank you for your reply. The following is the prediction result of the model I tried to use AutoNHITS. (Try to ensure that the hyperparameters in the config are consistent with those in DeepAR.) I assigned h to 1 before mainly because it can only be set in this way. When the value is assigned to 2 or larger, an error will be reported (cuda out of memory QAQ). In addition, I have a question. When I checked the loss information of each trail during the training process, I found that in the lightning_logs folder under the same path as the code, the train_loss_step and train_loss_epoch of the training set at each step were recorded, and valid_loss of the validation set. The number of train_loss_step is consistent with the number of steps. However, I found that valid_loss records data once every one hundred steps. I would like to ask if it can be adjusted so that valid_loss is recorded once every step? If possible, in which code source file should I modify it?

import <http://plotly.io|plotly.io> as pio

pio.renderers.default = 'browser'

import optuna

import pandas as pd

import datetime

import warnings

warnings.filterwarnings("ignore")

import matplotlib.pyplot as plt

from neuralforecast.core import NeuralForecast

from neuralforecast.models import NHITS

from neuralforecast.losses.pytorch import DistributionLoss, MQLoss

data = pd.read_csv("IMF6.csv")

data['ds'] = <http://pd.to|pd.to>_datetime(data['ds'], unit='s')

val_size = 64*320

test_size = 64*320

models = [NHITS(

h = horizon,

input_size = 40,

futr_exog_list= [

'railway',

'SPEED',

],

hist_exog_list=None,

stat_exog_list=None,

exclude_insample_y=False,

stack_types = ["identity", "identity", "identity"],

n_blocks = [1, 1, 1],

mlp_units = 2 * [[128, 128]],

n_pool_kernel_size = [16, 8, 1],

n_freq_downsample = [24, 12, 1],

pooling_mode = "MaxPool1d",

interpolation_mode = "linear",

dropout_prob_theta= 0.0,

activation="ReLU",

loss=DistributionLoss(distribution="Normal", level=[60, 90], return_params=False),

valid_loss=MQLoss(level=[60, 90]),

max_steps = 1600,

learning_rate = 0.00011156,

num_lr_decays = 3,

early_stop_patience_steps = -1,

val_check_steps = 100,

batch_size = 20,

valid_batch_size = None,

windows_batch_size = 64,

inference_windows_batch_size = -1,

start_padding_enabled=False,

step_size = 1,

scaler_type = "robust",

random_seed = 1,

num_workers_loader=0,

drop_last_loader=False,

)

]

nf = NeuralForecast(models=models, freq='s')

Y_hat_df = nf.cross_validation(df=data, val_size=val_size,

test_size=test_size, n_windows=None)