- Custom `kernel_size` support added to all base models
0.4
1. Liquid base model `TimeDiffuionLiquid` added. Lightweight, fast, but less accurate than the main model. 2. Abstract class for all base models `TimeDiffusionModel` added.
0.3
- Using `TimeDiffusionProjector` as base for `TD` framework with option to choose `TimeDiffusionAttention`. - Model `eval()` fix in TD framework methods - Additional time series forecasting examples added for bitcoin historical prices
0.2.2
- Early stopping epochs patience mechanism added in TD.fit, alowing to interrupt learning process at the epoch end, if model hasn't learned anything new in latest epochs (validation quality hasn't improved)
0.2.1
- Nan support added to: - - `DimUniversalStandardScaler` - - `TD.restore` method (`TD.fit` method already supported nans in mask positions) - Removed redundant `fill_mask` parameter of `TD.fit` method
0.2
- Few TD framework improvements and bug fixes related to scaling, typing and mask operations. - Scaler tests and nan checks in model's and framework's tests added.