Getting error about batch_size being zero when i have batch_size set to non-zero

[dhuang-apex]

horizon=58

m1 = NeuralProphet(
    n_changepoints=30,
    changepoints_range=0.90,
    yearly_seasonality=False,
    weekly_seasonality=True,
    daily_seasonality=False,
    n_forecasts=horizon,
    n_lags=horizon,
    epochs=100,
    batch_size=16,
)
metrics = m1.fit(train, freq="D", validation_df=valid)

When i instantiate the model with these settings, I get:
ValueError: batch_size should be a positive integer value, but got batch_size=0

But as soon as I comment out the n_lags hyperparameter (which turns off AR-Net?), the model runs with no issue. Anyone else ran into this as well?

Some more details about the error:

ValueError                                Traceback (most recent call last)
<command-2469282746511484> in <module>
     11     batch_size=16,
     12 )
---> 13 metrics = m1.fit(train, freq="D", validation_df=valid)
     14 plot_metrics(metrics)

/databricks/python/lib/python3.8/site-packages/neuralprophet/forecaster.py in fit(self, df, freq, validation_df, epochs, local_modeling, progress_bar, plot_live_loss, progress_print, minimal)
    937             validation_df = self._check_dataframe(validation_df, check_y=False, exogenous=False)
    938             validation_df = self.handle_missing_data(validation_df, freq=self.data_freq)
--> 939             metrics_df = self._train(
    940                 df,
    941                 validation_df,

/databricks/python/lib/python3.8/site-packages/neuralprophet/forecaster.py in _train(self, df, df_val, progress_bar, plot_live_loss, progress_print)
    659         val = df_val is not None
    660         if val:
--> 661             val_loader = self._init_val_loader(df_val)
    662             val_metrics = metrics.MetricsCollection([m.new() for m in self.metrics.batch_metrics])
    663 

/databricks/python/lib/python3.8/site-packages/neuralprophet/forecaster.py in _init_val_loader(self, df)
    532         df = df_utils.normalize(df, self.data_params, local_modeling=self.local_modeling)
    533         dataset = self._create_dataset(df, predict_mode=False)
--> 534         loader = DataLoader(dataset, batch_size=min(1024, len(dataset)), shuffle=False, drop_last=False)
    535         return loader
    536 

/databricks/python/lib/python3.8/site-packages/torch/utils/data/dataloader.py in __init__(self, dataset, batch_size, shuffle, sampler, batch_sampler, num_workers, collate_fn, pin_memory, drop_last, timeout, worker_init_fn, multiprocessing_context, generator, prefetch_factor, persistent_workers)
    272         if batch_size is not None and batch_sampler is None:
    273             # auto_collation without custom batch_sampler
--> 274             batch_sampler = BatchSampler(sampler, batch_size, drop_last)
    275 
    276         self.batch_size = batch_size

/databricks/python/lib/python3.8/site-packages/torch/utils/data/sampler.py in __init__(self, sampler, batch_size, drop_last)
    216         if not isinstance(batch_size, int) or isinstance(batch_size, bool) or \
    217                 batch_size <= 0:
--> 218             raise ValueError("batch_size should be a positive integer value, "
    219                              "but got batch_size={}".format(batch_size))
    220         if not isinstance(drop_last, bool):

ValueError: batch_size should be a positive integer value, but got batch_size=0

[ourownstory]

Hi @apex-dhuang may it be that your time series is shorter than 116 days (58 lags+58 targets) long?

[dhuang-apex]

That's not it unfortunately. The time series data I'm testing on is only 346 days long, but just now I've tried to adjust the training vs validation vs test window to make sure training set is several times longer than lag + forecast length, but still getting the same result

[ourownstory]

likely your test/val sets are shorter than 116? or are you getting the error on train df?

[ourownstory]

BTW, if you were to create a dummy version of your problem (reproducible code with dummy data), I could help you better.