Bugfix, weights shape during length randomization

[bohdan-safoniuk]

Description

This PR fixes RuntimeError during calculation of weighted MultiHorizonMetric loss due to losses and weights shape mismatch.

Setup

training_dataset = TimeSeriesDataSet(
    dataset.loc[dataset["date"] <= cutoff_],
    time_idx="time_idx",
    target="target",
    weight="weight",
    group_ids=["group_id_1", "group_id_2", "group_id_3"],
    min_encoder_length=max_encoder_length // 2,
    max_encoder_length=max_encoder_length,
    min_prediction_length=1,
    max_prediction_length=max_prediction_length,
    static_categoricals=static_categoricals,
    time_varying_known_reals=time_varying_known_reals,
    time_varying_known_categoricals=time_varying_known_categoricals,
    time_varying_unknown_reals=time_varying_unknown_reals,
    target_normalizer=target_normalizer,
    categorical_encoders=categorical_encoders,
    scalers=scalers,
    add_relative_time_idx=True,
    add_encoder_length=True,
    randomize_length=(0.1, 0.15),
    allow_missing_timesteps=False,
)

tft = TemporalFusionTransformer.from_dataset(
    training_dataset,
    learning_rate=0.0001,
    hidden_size=256,
    lstm_layers=2,
    attention_head_size=4,
    dropout=0.5,
    hidden_continuous_size=128,
    output_size=7,
    loss=QuantileLoss(),
    reduce_on_plateau_patience=4,
)

Error

File /opt/conda/envs/forecaster-env/lib/python3.9/site-packages/pytorch-forecasting/pytorch_forecsting/metrics/base_metrics.py in MultiHorizonMetric.update(self, y_pred, target)
     786   if weight is not None:
---> 787       losses = losses * unsqueeze_like(weight, losses)
     789   self._update_losses_and_lengths(losses, lengths)

RuntimeError: The size of tensor a (12) must match the size of tensor b (22) at non-singleton dimension 1

Issues

Possible linked issues (they have exact same error, but different setup)

• MultiHorizonMetric weighted loss not working #942
• when use weight param in TimeSeriesDataSet, raised RuntimeError: The size of tensor a (128) must match the size of tensor b (12) at non-singleton dimension 1 #1099

[codecov-commenter]

Codecov Report

Patch coverage: 50.00% and project coverage change: -0.02% ⚠️

Comparison is base (9995d0a) 90.13% compared to head (4149e9a) 90.11%.
Report is 2 commits behind head on master.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1318      +/-   ##
==========================================
- Coverage   90.13%   90.11%   -0.02%     
==========================================
  Files          30       30              
  Lines        4712     4714       +2     
==========================================
+ Hits         4247     4248       +1     
- Misses        465      466       +1     

Flag	Coverage Δ
cpu	90.11% <50.00%> (-0.02%)	⬇️
pytest	90.11% <50.00%> (-0.02%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files Changed	Coverage Δ
pytorch_forecasting/data/timeseries.py	93.19% <50.00%> (-0.13%)	⬇️

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.