```modelLSTM = AutoLSTM(h=h,
...
# neural-forecastb
Bersu T
03/18/2025, 11:31 AM Copy code
modelLSTM = AutoLSTM(h=h,
loss=MAE(),
backend='optuna',
num_samples=10)
nf = NeuralForecast(models=[modelLSTM], freq='ME')
nf.fit(df=df_encoded, val_size=18)m
Marco
03/18/2025, 2:05 PMHello! You do it in the LSTM model. Something like:
Copy code
config = dict(start_padding_enabled=True)
model = AutoLSTM(h=h, config=config)b
Bersu T
03/19/2025, 12:31 PMThanks! How can I perform cross validation with the best LSTM configuration found with AutoLSTM? My current code is this:
Copy code
def config_lstm(trial):
return {
"start_padding_enabled": True}
modelLSTM = AutoLSTM(h=h,
config=config_lstm,
loss=MAE(),
backend='optuna',
num_samples=10)
nf = NeuralForecast(models=[modelLSTM], freq='ME')
nf.fit(df=df_encoded, val_size=18)Bersu T
03/19/2025, 1:00 PMI have added the following, however training times are very long. I believe neural forecast is also global. MLforecast takes me max 3 mins for the same settings, whereas this takes me 2 hours. Does the following take the best configuration? And how can I improve my performance as currently it is performing really bad.
Copy code
cv_result_lstm = nf.cross_validation(
df=df_encoded,
n_windows=n_windows,
step_size=step_size,
val_size=18,
id_col='unique_id',
time_col='ds',
target_col='y')Bersu T
03/19/2025, 1:01 PMimage.png
m
Marco
03/19/2025, 1:42 PMYou can read our tutorial on hyperparameter optimization to see how you can get the config for each run and then select the best config. For better results, I would suggest using another model than LSTM, maybe something like NHITS would perform better, and would be faster to train.
b
Bersu T
03/19/2025, 1:57 PMI have looked at the website but I can't retrieve where I can get the best config. I used this for ML: config_lgb = auto_mlf.results_['my_lgb'].best_trial.user_attrs['config'], but this does not apply to neuralforecast
5 Views