Catasta

New Features
This release brings classification support for Catasta,, aligning with the existing regression functionalities for a consistent API experience. This update introduces several key classes designed to facilitate the development and evaluation of classification models:
- `ClassificationTrainInfo`
- `ClassificationEvalInfo`
- `ImageClassificationDataset`
- `ClassificationScaffold`

Preimplemented Models
This release comes with a selection of preimplemented models for immediate use in classification tasks:
- Feedforward Neural Network
- Convolutional Neural Network (CNN)
- Transformer Model
- Mamba Model

Upcoming Features
- **Image Classification Inference:** While currently not supported, we are planning to introduce image classification inference capabilities in an upcoming update.

Features and Changes
* The device type for inference can now be set to Archways
* You can select in Gaussian regressors wether to use ARD

Features and Changes
* added new argument to transformers to disable layer normalization
* Changed attribute from AproximateGPRegressor from "n_inputs" to "context_length" for consistence between regressors
* Implemented Mamba regressor

Features and Changes
* Improved the training remaining time for real
* RegressionTrainInfo class:
* Changed the attribute name from "train_loss" to "train_losses"
* Changed the attribute name from "eval_loss" to "val_losses"
* Added attributes "best_val_loss", "best_train_loss", and "lr_values"
* Deleted the "verbose" argument from training as it does not have a significant impact in performance

Features and Changes
* Improved evaluation time with batch inference
* Deleted predict method from Scaffold
* Changed the name in RegressionEvalInfo from "stds" to "predicted_std"
* Changed the name n RegressionPrediction from "prediction" to "value" and "stds" to "std"
* Implemented Kalman filter to transformations

Features and Changes
* Now the model can be saved to the specified path
* A new class has been added to make the predictions given a model
* Improved the error message of the regression dataset when the path is not found
* Now "/" is added at then end of the path if not specified
* The learning rate is now displayed in scientific notation
* Deleted ms/batch metric

Bug fixes
* Now the model state manager saves the first model state so it is not empty when saved
* The time remaining is displayed more accurately