Do not re-randomize when shuffle=False

kauldron/data/kmix/mixture.py

@@ -37,14 +37,15 @@ class SampleFromDatasets(base.TFDataPipeline):
     stop_on_empty_dataset: If True, the iteration will stop on the first empty
       dataset.
     rerandomize_each_iteration: If True, the mixture will reshuffle the datasets
-      each time it is iterated over.
+      each time it is iterated over. By default (`None`), only reshuffle when
+      all the sub-datasets are also shuffled.
   """
 
   datasets: list[base.TFDataPipeline]
   _: dataclasses.KW_ONLY
   weights: None | list[float | int] = None
   stop_on_empty_dataset: bool = False
-  rerandomize_each_iteration: bool = True
+  rerandomize_each_iteration: bool | None = None
 
   def __post_init__(self):
     if not all(isinstance(ds, base.TFDataPipeline) for ds in self.datasets):
@@ -60,12 +61,19 @@ def ds_for_current_process(self, rng: random.PRNGKey) -> tf.data.Dataset:
         for i, ds in enumerate(self.datasets)
     ]
 
+    # Do not re-randomize if all datasets are in eval mode (shuffle=False)
+    # This is so iterating twice on evaluation yield examples in deterministic
+    # order (due to an issue with `tf.data`: b/362450807).
+    rerandomize_each_iteration = any(
+        not hasattr(ds, 'shuffle') or ds.shuffle for ds in self.datasets
+    )
+
     ds = tf.data.Dataset.sample_from_datasets(
         datasets,
         weights=self.weights,
         seed=int(rng.fold_in('sample_from_datasets').bits()),
         stop_on_empty_dataset=self.stop_on_empty_dataset,
-        rerandomize_each_iteration=self.rerandomize_each_iteration,
+        rerandomize_each_iteration=rerandomize_each_iteration,
     )
     return ds