The num_classes parameter in unimodal loss functions currently has a default value. However, it's important to note that specifying the correct number of classes for this parameter must be done for every case. Relying on a default value may lead to errors that are difficult to diagnose.

Recommendation:

• Remove the default value for the num_classes parameter in BetaCrossEntropyLoss, BinomialCrossEntropyLoss, ExponentialRegularisedCrossEntropyLoss, GeneralTriangularCrossEntropyLoss, PoissonCrossEntropyLoss, and TriangularCrossEntropyLoss

Describe the bug

The CLM layer with cloglog and logit link functions has a numerical instability in the computation of the z3 variable. It uses a torch.exp(-z3) so when z3 is aproximately above 15 it returns infinity.

Steps/Code to reproduce the bug

.

Expected results

.

Actual results

.

The tutorials require an update to work with the latest versions of torch and torchvision. Currently in the model creation section they return the following warnings

UserWarning: The parameter 'pretrained' is deprecated since 0.13 and may be removed in the future, please use 'weights' instead.
UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and may be removed in the future. The current behavior is equivalent to passing `weights=ResNet18_Weights.IMAGENET1K_V1`. You can also use `weights=ResNet18_Weights.DEFAULT` to get the most up-to-date weights.

Please, can you describe the required steps to generate the documentation using sphinx?
Thank you.

Describe the issue

label_smoothing parameter of CrossEntropyLoss applies label smoothing by mixing the one-hot targets with a uniform distribution. However, in soft labelling loss functions, it makes no sense to mix an already soft label encoding with a uniform distribution.

Suggest a potential alternative/fix

The label_smoothing parameter should be removed from

• PoissonCrossEntropyLoss
• BinomialCrossEntropyLoss
• ExponentialCrossEntropyLoss
• BetaCrossEntropyLoss
• TriangularCrossEntropyLoss
• GeneralTriangularCrossEntropyLoss

Then, the value passed to the CrossEntropyLoss when initialising the ce_loss attribute should be 0.

Additional context

Thank you!

Hello team!
Thank you for this very useful tool.
I've been working with it for a few days and I've detected that there is a problem with displaying the descriptions of the class attributes of the datasets module in the software documentation.

Could you please fix it?

Thank you again!

Describe the issue

PytorchEstimator currently offers a very basic classifier with the interface of scikit-learn. However, it lacks numerous essential functionalities. Some python packages like skorch provide implementations for these missing features. Since dlordinal elements seamlessly integrate with such packages, it seems unnecessary and beyond the package's intended scope to have an estimator class within it.

Suggest a potential alternative/fix

The PytorchEstimator class should be deprecated and subsequently removed from this package. Instead, users should be encouraged to utilize third-party packages that already incorporate a PyTorch estimator with a scikit-learn interface. To facilitate this transition, comprehensive tutorials should be provided, describing how to seamlessly integrate dlordinal with these third-party alternatives.

Additional context

No response

The predict and predict_proba methods within the PytorchEstimator class should be modified to return numpy arrays instead of Tensors. This adjustment is necessary to align with the interface conventions of scikit-learn estimators, ensuring seamless integration and consistency across frameworks.

Also, a verbose parameter should be included to enable or disable the messages which are printed in the current version.

Currently, the PytorchEstimator lacks flexibility in managing verbosity during the training phase. It uniformly prints progress updates on each epoch, displaying only the current epoch and the total number of epochs. However, there are scenarios where users might prefer to customize this output. Some may seek to remove this message, while others might find it beneficial to include additional information such as the loss value per epoch.

Is it possible to add a verbose parameter to achieve this? Thank you!

distributions module do not implement probability distributions. Instead, it employs different probability distributions to determine soft labels for a given number of splits. Therefore, the whole module should be renamed to softlabelling and the functions that it contains should also be renamed as follows:

• get_beta_probabilities -> get_beta_softlabels
• get_binomial_probabilities -> get_binomial_softlabels
• get_exponential_probabilities -> get_exponential_softlabels
• get_triangular_probabilities -> get_triangular_softlabels
• get_general_triangular_probabilities -> get_general_triangular_softlabels

Absolute imports should be used in test files.

Some functions in the distributions module compute the soft labels for a single target (get_beta_probabilities) while others compute the soft labels for all J classes. Therefore, some functions return a vector while others return a 2d matrix. It should be standardised.

Better description of the functions implemented in the distributions module should be provided.