improve: update InfoLM class to dynamically set higher_is_better

src/torchmetrics/text/infolm.py

@@ -38,6 +38,21 @@
     __doctest_skip__ = ["InfoLM", "InfoLM.plot"]
 
 
+_information_measure_higher_is_better = {
+    # following values are <0
+    "kl_divergence": True,
+    "alpha_divergence": True,
+    # following values are >0
+    "beta_divergence": False,
+    "ab_divergence": False,
+    "renyi_divergence": False,
+    "l1_distance": False,
+    "l2_distance": False,
+    "l_infinity_distance": False,
+    "fisher_rao_distance": False,
+}
+
+
 class InfoLM(Metric):
     """Calculate `InfoLM`_.
 
@@ -111,7 +126,6 @@ class InfoLM(Metric):
     """
 
     is_differentiable = False
-    higher_is_better = True
     preds_input_ids: List[Tensor]
     preds_attention_mask: List[Tensor]
     target_input_ids: List[Tensor]
@@ -156,6 +170,15 @@ def __init__(
         self.add_state("target_input_ids", [], dist_reduce_fx="cat")
         self.add_state("target_attention_mask", [], dist_reduce_fx="cat")
 
+    @property
+    def higher_is_better(self) -> bool:
+        """Returns a bool indicating whether a higher value of the information measure is better.
+
+        Done this way as depends on if the information measure is positive or negative.
+
+        """
+        return _information_measure_higher_is_better[self.information_measure]
+
     def update(self, preds: Union[str, Sequence[str]], target: Union[str, Sequence[str]]) -> None:
         """Update state with predictions and targets."""
         preds_input_ids, preds_attention_mask, target_input_ids, target_attention_mask = _infolm_update(

tests/unittests/text/test_infolm.py

@@ -16,7 +16,7 @@
 import pytest
 import torch
 from torchmetrics.functional.text.infolm import infolm
-from torchmetrics.text.infolm import InfoLM
+from torchmetrics.text.infolm import InfoLM, _information_measure_higher_is_better
 from torchmetrics.utilities.imports import _TRANSFORMERS_GREATER_EQUAL_4_4
 
 from unittests._helpers import skip_on_connection_issues
@@ -182,3 +182,18 @@ def test_infolm_differentiability(self, preds, targets, information_measure, idf
             metric_functional=infolm,
             metric_args=metric_args,
         )
+
+    @skip_on_connection_issues()
+    def test_infolm_higher_is_better_property(self, preds, targets, information_measure, idf, alpha, beta):
+        """Test the `higher_is_better` property of the metric."""
+        metric_args = {
+            "model_name_or_path": MODEL_NAME,
+            "information_measure": information_measure,
+            "idf": idf,
+            "alpha": alpha,
+            "beta": beta,
+            "max_length": MAX_LENGTH,
+        }
+
+        metric = InfoLM(**metric_args)
+        assert metric.higher_is_better == _information_measure_higher_is_better[information_measure]