[Fix] Honour val_begin=0 in EpochBasedTrainLoop

mmengine/runner/loops.py

@@ -26,7 +26,9 @@ class EpochBasedTrainLoop(BaseLoop):
         dataloader (Dataloader or dict): A dataloader object or a dict to
             build a dataloader.
         max_epochs (int): Total training epochs.
-        val_begin (int): The epoch that begins validating.
+        val_begin (int): The epoch that begins validating. If it is set to
+            0, an additional validation is performed before the first epoch
+            is trained, which allows the model to be evaluated at epoch 0.
             Defaults to 1.
         val_interval (int): Validation interval. Defaults to 1.
         dynamic_intervals (List[Tuple[int, int]], optional): The
@@ -94,6 +96,12 @@ def run(self) -> torch.nn.Module:
         """Launch training."""
         self.runner.call_hook('before_train')
 
+        # `val_begin` is 1-based for the validation performed after an epoch.
+        # Allowing it to be 0 additionally validates at epoch 0, i.e. before
+        # any training step has been taken.
+        if self.val_begin <= 0 and self.runner.val_loop is not None:
+            self.runner.val_loop.run()
+
         while self._epoch < self._max_epochs and not self.stop_training:
             self.run_epoch()
 

tests/test_runner/test_runner.py

@@ -1835,6 +1835,57 @@ def train_step(self, *args, **kwargs):
         runner.train()
         self.assertEqual(runner.iter, 3 * 2)
 
+    def test_val_begin_zero(self):
+        # `val_begin=0` should additionally validate before the first epoch is
+        # trained, so that the model can be evaluated at epoch 0. See #1448.
+        val_epochs = []
+        val_iters = []
+
+        @HOOKS.register_module(force=True)
+        class TestValBeginZeroHook(Hook):
+
+            def before_val_epoch(self, runner):
+                val_epochs.append(runner.epoch)
+                val_iters.append(runner.iter)
+
+        cfg = copy.deepcopy(self.epoch_based_cfg)
+        cfg.experiment_name = 'test_val_begin_zero'
+        cfg.custom_hooks = [dict(type='TestValBeginZeroHook', priority=50)]
+        cfg.train_cfg = dict(
+            by_epoch=True, max_epochs=2, val_interval=1, val_begin=0)
+        runner = Runner.from_cfg(cfg)
+        runner.train()
+
+        # epoch 0 is validated before any training step has been taken, and
+        # the regular per-epoch validation is unaffected
+        self.assertEqual(val_epochs, [0, 1, 2])
+        self.assertEqual(val_iters, [0, 4, 8])
+        HOOKS.module_dict.pop('TestValBeginZeroHook')
+
+        # the default behaviour must stay unchanged: validation only happens
+        # after an epoch has been trained
+        val_epochs = []
+        val_iters = []
+
+        @HOOKS.register_module(force=True)
+        class TestValBeginDefaultHook(Hook):
+
+            def before_val_epoch(self, runner):
+                val_epochs.append(runner.epoch)
+                val_iters.append(runner.iter)
+
+        cfg = copy.deepcopy(self.epoch_based_cfg)
+        cfg.experiment_name = 'test_val_begin_default'
+        cfg.custom_hooks = [dict(type='TestValBeginDefaultHook', priority=50)]
+        cfg.train_cfg = dict(
+            by_epoch=True, max_epochs=2, val_interval=1, val_begin=1)
+        runner = Runner.from_cfg(cfg)
+        runner.train()
+
+        self.assertEqual(val_epochs, [1, 2])
+        self.assertEqual(val_iters, [4, 8])
+        HOOKS.module_dict.pop('TestValBeginDefaultHook')
+
     @skipIf(
         SKIP_TEST_COMPILE,
         reason='torch.compile is not valid, please install PyTorch>=2.0.0')