Error when providing superclass arguments as kwargs

[jlcsilva]

🐛 Bug

When some classes are instantiated with superclass arguments as kwargs, the superclass constructor receives multiple values for these arguments and instantiation fails with an error like:

TypeError: torchmetrics.classification.stat_scores.StatScores.__init__() got multiple values for keyword argument 'reduce'

This happens for the following classes:

• Accuracy
• Dice
• JaccardIndex
• Precision
• Recall
• Specificity

To Reproduce

Instantiate Accuracy, Dice, Precision, Recall or Specificity specifying the argument reduce or mdmc_reduce as kwarg, or the class JaccardIndex with the normalize argument as kwarg.

Code sample

Accuracy(reduce='micro')
Accuracy(mdmc_reduce='global')
Dice(reduce='micro')
Dice(mdmc_reduce='global')
Precision(reduce='micro')
Precision(mdmc_reduce='global')
Recall(reduce='micro')
Recall(mdmc_reduce='global')
Specificity(reduce='micro')
Specificity(mdmc_reduce='global')
JaccardIndex(num_classes=1, normalize='true')

Expected behavior

The class should be instantiated prioritizing values provided as kwargs, instead of default values generated inside the constructor.

Environment

• TorchMetrics 0.9.0 built from source
• Python 3.10.4 & PyTorch 1.11.0
• WSL Ubuntu 20.04

[github-actions]

Hi! thanks for your contribution!, great first issue!