[BUG] too short time series give a ZeroDivisionError

Describe the bug When training some models with a short time series (eg. 8 points), a ZeroDivisionError occur.

---------------------------------------------------------------------------
ZeroDivisionError                         Traceback (most recent call last)
<ipython-input-29-7c617aa60865> in <module>
     12 model = NBEATSModel(input_chunk_length=1, output_chunk_length=1)
     13 
---> 14 model.fit(series)

~/.local/lib/python3.8/site-packages/darts/utils/torch.py in decorator(self, *args, **kwargs)
     63         with fork_rng():
     64             manual_seed(self._random_instance.randint(0, high=MAX_TORCH_SEED_VALUE))
---> 65             decorated(self, *args, **kwargs)
     66     return decorator

~/.local/lib/python3.8/site-packages/darts/models/torch_forecasting_model.py in fit(self, series, covariates, val_series, val_covariates, verbose)
    292         logger.info('Train dataset contains {} samples.'.format(len(train_dataset)))
    293 
--> 294         self.fit_from_dataset(train_dataset, val_dataset, verbose)
    295 
    296     @random_method

~/.local/lib/python3.8/site-packages/darts/utils/torch.py in decorator(self, *args, **kwargs)
     63         with fork_rng():
     64             manual_seed(self._random_instance.randint(0, high=MAX_TORCH_SEED_VALUE))
---> 65             decorated(self, *args, **kwargs)
     66     return decorator

~/.local/lib/python3.8/site-packages/darts/models/torch_forecasting_model.py in fit_from_dataset(self, train_dataset, val_dataset, verbose)
    345 
    346         # Train model
--> 347         self._train(train_loader, val_loader, tb_writer, verbose)
    348 
    349         # Close tensorboard writer

~/.local/lib/python3.8/site-packages/darts/models/torch_forecasting_model.py in _train(self, train_loader, val_loader, tb_writer, verbose)
    518 
    519             if epoch % self.nr_epochs_val_period == 0:
--> 520                 training_loss = total_loss / len(train_loader)
    521                 if val_loader is not None:
    522                     validation_loss = self._evaluate_validation_loss(val_loader)

ZeroDivisionError: division by zero

Seems like this is the case only for global models. Tried with:

• NBEATSModel
• TCNModel
• RNNModel
• TransformerModel

To Reproduce Run the following code snippet to get the error

from darts.models.transformer_model import TransformerModel
import pandas as pd
from darts import TimeSeries

df = pd.DataFrame({
    'time': ['2010-01-01', '2012-01-01'], # change 2012 to 2019, and the error disappear
    'values': [1, 2],
})

df['time'] =  pd.PeriodIndex(df['time'], freq='Q').to_timestamp()
df.index = df['time']
df = df.resample('Q').sum().interpolate('time').reset_index()

series = TimeSeries.from_dataframe(df, time_col='time', value_cols=['values'])

model = TransformerModel(input_chunk_length=2, output_chunk_length=1)
model.fit(series)

Expected behavior The model should train on short time series.

System (please complete the following information):

• Python version: 3.8.8
• darts version: 0.6.1

[kaosdev]

Ok, found the problem. This has to do with the batch size. If you create the model passing the batch_size as an argument, the error goes away.

Maybe it’s better to get a clear error instead of a ZeroDivisionError? Or maybe handling it. E.g.

if batch_size > len(train_dataset):
   batch_size = len(train_dataset)

[kaosdev]

Thanks @kaosdev. I had the same error and reducing the batch size helped. I tried figuring out how the train_dataset length is calculated. If I have a dataset with 63 rows in total and my output_chunk_length is 12 and my input_chunk_length is 24. My model’s batch_size is 8 (LSTM - RNN). How does it calculate that I have 15 samples? Thanks. I tried figuring it out from here line 292, and here from the SequentialDataset Class. If you can help, please let me know. Thanks.

Mhh, I may be wrong but I think you should have more than 15 samples.

The formula should be (if I am right): samples = total - input - output

So in your case: 63 - 24 - 12 = 27

Batch size has nothing to do with the number of samples, it should just be less than or equal to the number of samples.