unit8co/darts

AI Summary: This issue proposes handling cases where scaled forecasting metrics (like MASE) encounter a zero denominator due to a perfectly seasonal or constant in-sample series. Instead of raising an error, the metrics should return NaN. Additionally, if the model's prediction also achieves a zero error score, it should be considered equivalent to the naive forecast and return 1.

AI Summary: This issue proposes to enhance the `StatsForecastModel` to accept model specifications as strings or classes, similar to `NeuralForecastModel`. This change aims to simplify model configuration, especially when using external config files, by allowing users to pass model names and their corresponding keyword arguments (`model_kwargs`) directly, rather than requiring pre-instantiated model objects.

AI Summary: This issue requests the addition of `save()` and `load()` methods to the `KMeansScorer` class. The goal is to enable persistence of the scorer, allowing users to avoid retraining the model by saving and loading its state.

AI Summary: This issue proposes adding an optional `insample` parameter to the `inverse_transform()` method of Darts' data transformers and pipelines. This change aims to simplify the process of inverse-transforming forecasts generated by the `Diff` transformer, which currently requires manual prepending of the transformed insample target series.

AI Summary: The `TimeSeries.add_datetime_attribute` method in the `darts` library has a `tz` parameter with a restrictive type hint (`str | None`). This parameter is ultimately passed to Pandas' `DatetimeIndex.tz_convert`, which supports a broader range of types, including `tzinfo` objects and integers. The bug is that `darts`'s type hints prevent users from utilizing the full capabilities of the underlying Pandas method, and the proposed fix is to broaden the type hints in the `darts` call chain to match Pandas' flexibility.

AI Summary: The user discovered that `TorchForecastingModel.fit_from_dataset()` and `predict_from_dataset()` do not utilize encoders, nor do they warn the user about this omission. This silent behavior can lead to unexpected results, particularly for models like TFTModel which rely on encoders for future covariates and categorical feature embeddings. The issue suggests either documenting this behavior or integrating encoder logic directly into the `_from_dataset` methods.

AI Summary: This feature request aims to enhance the `TimeSeries.to_json()` and `from_json()` methods to correctly handle and serialize/deserialize static covariates, metadata, and hierarchy information. Currently, these components are not supported, leading to potential data loss or incorrect representation when converting time series data to and from JSON format.

AI Summary: This feature request proposes adding a new function `to_group_dataframe()` to the Darts library. This function will convert a list of `TimeSeries` objects into a single "long" DataFrame, incorporating optional static covariates and metadata. The generated DataFrame should be compatible with `TimeSeries.from_group_df()` for reconstruction.

AI Summary: The TFTExplainer in Darts incorrectly indexes the attention_heads matrix when dealing with a list of multiple independent timeseries. This leads to an IndexError because it uses indices derived from iterating over the timeseries list, which can exceed the bounds of the attention_heads matrix size. The issue is reproducible with large datasets like TrafficDataset and requires a fix in the `TFTExplainer.explain()` method's indexing logic.

AI Summary: This bug report highlights an inconsistency in the type hinting for the `lags` parameter in the CatBoost model within the Darts library. Unlike other Scikit-learn models which accept a dictionary for default lags, CatBoost currently only supports integers or lists, leading to a discrepancy in the API documentation and functionality.

AI Summary: The user is encountering an issue where `darts.models.XGBModel`, particularly with its `historical_forecasts` method, fails to utilize GPU acceleration for XGBoost, despite having a CUDA-enabled XGBoost installation that works when called directly. The error messages indicate that no GPU is found and XGBoost is not compiled with CUDA support, suggesting a problem with how `darts` initializes or passes GPU-related parameters to the underlying XGBoost library.

AI Summary: The task is to investigate why the `darts` Python library's wheel file includes an `examples` package, which is considered a potential source of naming conflicts with other packages. The solution likely involves removing the `examples` directory from the wheel's contents during the build process, ensuring that example code is distributed separately (e.g., via a repository or documentation site) and not packaged within the main library.

AI Summary: The task is to improve the parameter validation in the Darts Python library's `torch_forecasting_model.py` file. The current validation breaks subclassing due to hardcoded class checks. The solution involves modifying the validation to dynamically inspect the inheritance chain of a model class to determine valid parameters, ensuring compatibility with subclasses.

AI Summary: The task is to debug and fix a bug in the Darts Python library where saving global baseline models (GlobalNaiveSeasonal, GlobalNaiveAggregate, GlobalNaiveDrift) fails due to an AttributeError related to PyTorch Lightning's Trainer. The solution involves investigating why the Trainer isn't properly attached to these models during the saving process and modifying the `save()` method to ensure compatibility.