[BUG] DeepAR: Inverted condition causes n_plotting_samples to remain None when n_validation_samples is None

[Gitanaskhan26]

Describe the bug

In the DeepAR model initialization, there is a logic error where the condition for setting n_plotting_samples is inverted, causing it to remain None when n_validation_samples is also None.

The buggy code in pytorch_forecasting/models/deepar/_deepar.py lines 143-147:

if n_plotting_samples is None:
    if n_validation_samples is None:
        n_plotting_samples = n_validation_samples  # BUG: assigns None to None!
    else:
        n_plotting_samples = 100

When both n_plotting_samples and n_validation_samples are None, the code assigns n_validation_samples (which is None) to n_plotting_samples, leaving it as None instead of assigning the default value of 100.

The condition if n_validation_samples is None: should be if n_validation_samples is not None:.

To Reproduce

from pytorch_forecasting.models import DeepAR

# Create DeepAR without specifying n_validation_samples or n_plotting_samples
# (Both default to None in function signature)
model = DeepAR(cell_type="LSTM")

# Check the resulting n_plotting_samples value
print(f"n_plotting_samples: {model.hparams.n_plotting_samples}")
# Output: n_plotting_samples: None
# Expected: n_plotting_samples: 100

# This None value is later used in create_log():
# n_samples = [self.hparams.n_validation_samples, self.hparams.n_plotting_samples][self.training]
# When training=True, n_samples will be None, which can cause downstream errors

Expected behavior

When n_validation_samples is None, the code should default n_plotting_samples to 100. The corrected code should be:

if n_plotting_samples is None:
    if n_validation_samples is not None:  # FIXED: inverted condition
        n_plotting_samples = n_validation_samples
    else:
        n_plotting_samples = 100

Additional context

The n_plotting_samples value is used in the create_log method (line 479) where it is passed as n_samples to prediction and quantiles functions. When None, this can cause issues in downstream calculations that expect an integer.

This bug affects all users who instantiate DeepAR without explicitly setting n_validation_samples or n_plotting_samples, which are both None by default in the function signature (line 65-66).

Versions

Details

System:
    python: 3.12.0
executable: /Library/Frameworks/Python.framework/Versions/3.12/bin/python3
   machine: macOS-26.3.1-arm64-arm-64bit

Python dependencies:
          pip: 23.2.1
pytorch-forecasting: 1.6.1
        torch: 2.9.1
    lightning: 2.6.1
        numpy: 2.3.5
        scipy: 1.15.3
       pandas: 2.3.3

[Gitanaskhan26]

@benHeid @fkiraly

I have identified and reported a critical logic bug in the DeepAR model initialization that I believe warrants immediate attention and resolution.

Issue Summary:
The bug is located in pytorch_forecasting/models/deepar/_deepar.py at lines 143-147, where an inverted conditional statement causes n_plotting_samples to remain None when n_validation_samples is also None, instead of defaulting to the intended value of 100. This occurs due to the condition checking if n_validation_samples is None: when it should check if n_validation_samples is not None:.

Technical Impact:

• The None value is subsequently utilized in the create_log() method (line 479) where it is passed as n_samples parameter to downstream prediction and quantile calculation functions
• This causes type errors and runtime failures for any user instantiating DeepAR without explicitly specifying these parameters
• The default function signature (lines 65-66) sets both parameters to None, making this a common failure scenario

Request:
Could you please review this issue and, if you determine it to be valid, consider assigning it to me for implementation of the fix? I am prepared to submit a pull request with the necessary corrections to resolve this issue.

Thank you for your attention to this matter.