Agml

This release adds three new datasets:
- **Object Detection**: `apple_detection_drone_brazil` and `apple_detection_spain`.
- **Image Classification**: `guava_disease_pakistan`.

It also fixes a bug which was preventing data serialization and splitting.

This release adds three new object detection datasets: `grape_detection_syntheticday`, `grape_detection_californiaday`, and `grape_detection_californianight`, as well as general bugfixes.

This release adds a new dataset, `autonomous_greenhouse_regression`, as well as general support for image regression datasets.

This release provides some small updates and bugfixes.

Changes

`agml.data`

- Added calculated mean and standard deviation for all public datasets.
- Added a new `normalize_images()` method for the `AgMLDataLoader` which automatically scales images to the 0-1 range and applies normalization.
- Updated data splitting and fixed a number of bugs including creating copies of the original managers, correctly instantiating new `DataObject`s, and applying batching.

`agml.backend`

- Updated an instantiation bug which sometimes occurred when newly installing the module and generating the `config.json`.

This release reworks the `AgMLDataLoader` internals and updates its compatibility with TensorFlow/PyTorch training pipelines.

Changes

`agml.data`

- Reworks the `AgMLDataLoader` architecture, removing the task-type-based subclasses and adding internal data management classes.
- The `batch`, `split`, and `shuffle` methods can be used to perform different operations on the data within the `AgMLDataLoader` itself.
- The `transform` and `resize_images` methods can be used to apply transforms to the data or resize the images to a specific size (or auto-resize the images smartly).
- The new method `labels_to_one_hot` convert labels to one-hot-vectors for image classification tasks.
- The `as_keras_sequence` and `as_torch_dataset` methods (alongside the `reset_preprocessing`, `disable_preprocessing`, and `eval` methods) enable the `AgMLDataLoader` to be used for training and evaluation modes independently, and directly in TensorFlow/PyTorch pipelines.
- The `export_contents` method exports the raw data mapping for the user to use outside of `agml`.
- The `export_torch` and `export_tensorflow` methods convert the loader to a `torch.utils.data.DataLoader` or `tf.data.Dataset`, respectively.

`agml.backend`

- Added the capability to set a global path to save datasets, using `agml.backend.set_data_save_path`. Once run, this will change the path all datasets are downloaded to until it is either reset or changed back.

`agml.viz`

- Bugfixes and improved visualization for different types of input images (both normalized and unnormalized).