Update CLU metrics to support dictionary returns for computed metric results to allow for additional flexibility

clu/metrics.py

@@ -92,6 +92,8 @@ def _assert_same_shape(a: jnp.ndarray, b: jnp.ndarray):
 
 
 M = TypeVar("M", bound="Metric")
+R = TypeVar("R", jnp.ndarray, dict[str, jnp.ndarray])
+V = TypeVar("V", clu.values.Value, dict[str, clu.values.Value])
 
 
 class Metric:
@@ -160,7 +162,7 @@ def merge(self: M, other: M) -> M:
   def _reduce_merge(self: M, other: M) -> M:
     return self.merge(other)
 
-  def compute(self) -> jnp.ndarray:
+  def compute(self) -> R:
     """Computes final metrics from intermediate values."""
     raise NotImplementedError("Must override compute()")
 
@@ -169,9 +171,13 @@ def empty(cls: type[M]) -> M:
     """Returns an empty instance (i.e. `.merge(Metric.empty())` is a no-op)."""
     raise NotImplementedError("Must override empty()")
 
-  def compute_value(self) -> clu.values.Value:
-    """Wraps compute() and returns a values.Value."""
-    return clu.values.Scalar(self.compute())
+  def compute_value(self) -> V:
+    """Wraps compute() and returns a values.Value or dict of values.Value."""
+    result = self.compute()
+    if isinstance(result, dict):
+      return {k: clu.values.Scalar(v) for k, v in result.items()}
+    else:
+      return clu.values.Scalar(result)
 
   def reduce(self: M) -> M:
     """Reduces the metric along it first axis by calling `_reduce_merge()`.
@@ -623,22 +629,34 @@ def reduce(self: C) -> C:
     })
 
   def compute(self) -> dict[str, jnp.ndarray]:
-    """Returns a dictionary mapping metric field name to `Metric.compute()`."""
+    """Returns a dictionary mapping metrics to their computed values."""
     _check_reduction_counter_ndim(self._reduction_counter)
-    return {
-        metric_name: metric.compute()
-        for metric_name, metric in vars(self).items()
-        if metric_name != "_reduction_counter"
-    }
+    metric_results = {}
+    for metric_name, metric in vars(self).items():
+      if metric_name != "_reduction_counter":
+        metric_result = metric.compute()
+        if isinstance(metric_result, dict):
+          metric_results.update(
+              {f"{metric_name}/{k}": v for k, v in metric_result.items()}
+          )
+        else:
+          metric_results[metric_name] = metric_result
+    return metric_results
 
   def compute_values(self) -> dict[str, clu.values.Value]:
-    """Computes metrics and returns them as clu.values.Value."""
+    """Computes metrics and returns them as clu_values.Value."""
     _check_reduction_counter_ndim(self._reduction_counter)
-    return {
-        metric_name: metric.compute_value()
-        for metric_name, metric in vars(self).items()
-        if metric_name != "_reduction_counter"
-    }
+    metric_results = {}
+    for metric_name, metric in vars(self).items():
+      if metric_name != "_reduction_counter":
+        metric_result = metric.compute_value()
+        if isinstance(metric_result, dict):
+          metric_results.update(
+              {f"{metric_name}/{k}": v for k, v in metric_result.items()}
+          )
+        else:
+          metric_results[metric_name] = metric_result
+    return metric_results
 
   def unreplicate(self: C) -> C:
     """Short-hand for `flax.jax_utils.unreplicate(self)`.