Feat/remove ood dataset

[paulnovello]

This pull request includes significant changes to the oodeel library, focusing on refactoring and improving the data handling and dataset management functionalities. The major changes include the introduction of a new data handler loader, backend detection, and the deprecation of older dataset handling classes (OODDataset).

Before:

• An OODDataset object is instantiated
• Operations are performed on OODDatasets
• Then OODDatasets return a tf.dataset or DataLoader using the .prepare method

ds_fit = OODDataset("cifar10", load_kwargs={"split": "train"}, input_key="image", backend=X)
ds_in = OODDataset("cifar10", load_kwargs={"split": "test"}, input_key="image", backend=X)

# 1b- Load out-of-distribution dataset: SVHN
ds_out = OODDataset("svhn_cropped", load_kwargs={"split": "test"}, backend=X)


# 2- prepare data (preprocess, shuffle, batch)
def preprocess_fn(*inputs):
    x = inputs[0] / 255
    return tuple([x] + list(inputs[1:]))


ds_fit = ds_fit.prepare(batch_size, preprocess_fn)
ds_in = ds_in.prepare(batch_size, preprocess_fn)
ds_out = ds_out.prepare(batch_size, preprocess_fn)

Problems

• OODDataset is a misleading name since it can also designate ID dataset
• OODDataset nature is redundant with tf.dataset and DataLoader, so it is confusing (in the example, ds_fit is an OODDataset first, but at the end is a tf.dataset or a DataLoader

After:

data_handler = load_data_handler(backend=X) #backend=X is optional since  load_data_handler automatically detects backend if only tf or torch is installed

ds_fit = data_handler.load_dataset("cifar10", load_kwargs={"split": "train"})
ds_in = data_handler.load_dataset("cifar10", load_kwargs={"split": "test"})

# 1b- Load out-of-distribution dataset: SVHN
ds_out = data_handler.load_dataset("svhn_cropped", load_kwargs={"split": "test"})


# 2- prepare data (preprocess, shuffle, batch)
def preprocess_fn(*inputs):
    x = inputs[0] / 255
    return tuple([x] + list(inputs[1:]))


ds_fit = data_handler.prepare(ds_fit, batch_size, preprocess_fn)
ds_in = data_handler.prepare(ds_in, batch_size, preprocess_fn)
ds_out = data_handler.prepare(ds_out, batch_size, preprocess_fn)

---

Refactoring and new functionalities:

• oodeel/datasets/__init__.py: Replaced the import of OODDataset with load_data_handler and moved OODDataset to a deprecated module.
• oodeel/datasets/data_handler.py: Added functions for backend detection (get_backend) and dynamic data handler loading ( load_data_handler). Introduced several new methods for dataset preparation and handling, such as split_by_class, prepare, load_dataset_from_arrays, and load_custom_dataset. [1] [2] [3] [4]

Deprecation and reorganization:

• oodeel/datasets/deprecated/DEPRECATED_data_handler.py: Added a deprecated version of the DataHandler class with the original methods.
• oodeel/datasets/deprecated/DEPRECATED_ooddataset.py: Renamed and moved ooddataset.py to a deprecated module, updating internal imports to reflect the new location. [1] [2] [3]
• oodeel/datasets/deprecated/__init__.py: Added a deprecation warning for the OODDataset object.

Initialization and imports:

• oodeel/datasets/data_handler.py: Added import statements for importlib and TensorType. [1] [2]
• oodeel/datasets/tf_data_handler.py: Added __init__ method to TFDataHandler class for initialization. [1] [2]

[github-actions]

☂️ Python Coverage

current status: ✅

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
2921	2637	90%	70%	🟢

New Files

File	Coverage	Status
oodeel/datasets/deprecated/DEPRECATED_data_handler.py	100%	🟢
oodeel/datasets/deprecated/DEPRECATED_tf_data_handler.py	81%	🟢
oodeel/datasets/deprecated/DEPRECATED_torch_data_handler.py	84%	🟢
oodeel/datasets/deprecated/init.py	100%	🟢
TOTAL	91%	🟢

Modified Files

File	Coverage	Status
oodeel/datasets/init.py	100%	🟢
oodeel/datasets/data_handler.py	86%	🟢
oodeel/datasets/tf_data_handler.py	82%	🟢
oodeel/datasets/torch_data_handler.py	83%	🟢
oodeel/eval/metrics.py	99%	🟢
oodeel/utils/tf_training_tools.py	75%	🟢
TOTAL	88%	🟢

updated for commit: a075c63 by action🐍

[y-prudent]

2 remarks, regarding:

• Data handler loading => Check the comments below
• Deprecated code: Should we retain the deprecated code? We could release a new major version (with an appropriate tag) when merging this PR, allowing users who depend on the old API to continue using an earlier version of oodeel. Moreover, since the path to the deprecated code has been changed, any updates using the old code would fail anyway.

Otherwise, LGTM !

[y-prudent]

With Python 3.12, when I do:

import importlib

print(importlib.util.find_spec("torch"))

I get the error:

AttributeError: module 'importlib' has no attribute 'util'

Here is a Python issue disscussing this error for versions of python > 3.11: https://discuss.python.org/t/python3-11-importlib-no-longer-exposes-util/25641. The solution seem to do the explicit import import importlib.util. By the way, get_backend function is never tested in oodeel!

Suggested change

	import importlib
	import importlib.util

[y-prudent]

It should be renamed test_instanciate_torch_datahandler.

And the case where the backend is not specified (equal to None) is not tested. It should be tested in the case where only torch is installed.