Automatically set a seed for data pipeline

PiperOrigin-RevId: 715752916

kauldron/data/pipelines.py

@@ -17,6 +17,7 @@
 import abc
 import dataclasses
 import functools
+import random
 from typing import Any, Optional, TypeAlias
 
 from etils import edc
@@ -76,3 +77,12 @@ def __iter__(self) -> iterators.Iterator:
     raise NotImplementedError()
 
   __repr__ = edc.repr
+
+  # Fox convenience, it can be annoying to have to manually set the seed when
+  # debugging on Colab.
+  def _assert_root_cfg_resolved(self) -> None:
+    # If the seed is not set when the pipeline is created, we create an
+    # arbitrary random seed.
+    if isinstance(self.seed, config_util._FakeRootCfg):  # pylint: disable=protected-access
+      object.__setattr__(self, 'seed', random.randint(0, 1000000000))
+    super()._assert_root_cfg_resolved()

kauldron/utils/config_util.py

@@ -386,7 +386,5 @@ def _assert_root_cfg_resolved_value(self) -> None:
         raise ValueError(
             f'{type(self).__qualname__}.{f.name} is an unresolved'
             f' `ROOT_CFG_REF` value ({value}).\nTo resolve the value, either'
-            f' explicitly set `{f.name}` in `__init__`, or call'
-            ' `obj.update_from_root_cfg(root_cfg)` to copy the value from the'
-            ' root `kd.train.Trainer` object.'
+            f' explicitly set `{f.name}` in `__init__`.'
         )