warn and disable training resize-and-crop branch when aug_config={}

docs/learn/train/advanced.md

@@ -285,6 +285,10 @@ To disable all augmentations, pass an empty dict:
 model.train(dataset_dir="path/to/dataset", aug_config={})
 ```
 
+Passing `{}` also drops the training resize-and-crop branch, so images are resized
+directly to the target scale without random cropping. To keep the default resize
+pipeline, omit `aug_config` or pass `aug_config=None`.
+
 ---
 
 ## Memory Optimization

docs/learn/train/augmentations.md

@@ -32,7 +32,7 @@ model.train(
 )
 ```
 
-To disable augmentations: `aug_config={}`. Omitting it uses the default (horizontal flip at 50%).
+To disable augmentations: `aug_config={}`. This also drops the training resize-and-crop branch, so images are resized directly to the target scale without random cropping. Omitting it (or passing `aug_config=None`) uses the default (horizontal flip at 50%) and keeps the resize-and-crop branch.
 
 ## Built-in Presets
 

src/rfdetr/datasets/aug_config.py

@@ -14,7 +14,7 @@
 model.train(dataset_dir="...", aug_config=AUG_CONSERVATIVE)
 model.train(dataset_dir="...", aug_config=AUG_AGGRESSIVE)
 
-# Disable all augmentations
+# Disable all augmentations (also drops the training resize-and-crop branch)
 model.train(dataset_dir="...", aug_config={})
 
 # Fully custom

src/rfdetr/datasets/coco.py

@@ -312,6 +312,7 @@ def _build_train_resize_config(
     *,
     square: bool,
     max_size: Optional[int] = None,
+    include_crop_branch: bool = True,
 ) -> List[Dict[str, Any]]:
     """Build the training resize pipeline as an Albumentations config list.
 
@@ -336,9 +337,15 @@ def _build_train_resize_config(
             ratio using ``A.SmallestMaxSize`` with an optional long-side cap.
         max_size: Maximum long-side size for non-square resizes.  Defaults to
             ``1333`` when *square* is ``False``.
+        include_crop_branch: If ``False``, omit the resize-and-crop branch so the
+            pipeline always uses Option A (direct resize). Useful when objects
+            of interest can be partially or fully cropped out of frame.
+            Defaults to ``True`` (preserves the original two-branch behavior).
 
     Returns:
-        A single-element list containing a ``OneOf`` config entry.
+        A single-element list. When ``include_crop_branch`` is ``True`` (default)
+        the entry wraps a ``OneOf`` over both branches; when ``False`` the
+        entry is Option A directly.
     """
     if square:
         option_a: Dict[str, Any] = {
@@ -384,9 +391,29 @@ def _build_train_resize_config(
             }
         }
 
+    if not include_crop_branch:
+        return [option_a]
+
     return [{"OneOf": {"transforms": [option_a, option_b]}}]
 
 
+def _crop_branch_enabled(aug_config: Optional[Dict[str, Any]]) -> bool:
+    """Decide whether the training resize pipeline keeps its resize-and-crop branch.
+
+    ``aug_config={}`` is an explicit request to disable augmentations; it also
+    drops the resize-and-crop branch. ``aug_config=None`` (the default) and any
+    non-empty config keep it.
+    """
+    if aug_config == {}:
+        logger.warning(
+            "aug_config={} disables the training resize-and-crop branch in addition to all "
+            "augmentations; images will not be randomly cropped. Pass aug_config=None to keep "
+            "the default resize pipeline."
+        )
+        return False
+    return True
+
+
 def make_coco_transforms(
     image_set: str,
     resolution: int,
@@ -470,8 +497,9 @@ def make_coco_transforms(
 
     if image_set == "train":
         resolved_aug_config = aug_config if aug_config is not None else AUG_CONFIG
+        include_crop_branch = _crop_branch_enabled(aug_config)
         resize_wrappers = AlbumentationsWrapper.from_config(
-            _build_train_resize_config(scales, square=False, max_size=1333)
+            _build_train_resize_config(scales, square=False, max_size=1333, include_crop_branch=include_crop_branch)
         )
         pipeline = [*resize_wrappers]
         if not gpu_postprocess:
@@ -564,7 +592,10 @@ def make_coco_transforms_square_div_64(
 
     if image_set == "train":
         resolved_aug_config = aug_config if aug_config is not None else AUG_CONFIG
-        resize_wrappers = AlbumentationsWrapper.from_config(_build_train_resize_config(scales, square=True))
+        include_crop_branch = _crop_branch_enabled(aug_config)
+        resize_wrappers = AlbumentationsWrapper.from_config(
+            _build_train_resize_config(scales, square=True, include_crop_branch=include_crop_branch)
+        )
         pipeline = [*resize_wrappers]
         if not gpu_postprocess:
             aug_wrappers = AlbumentationsWrapper.from_config(resolved_aug_config)

tests/datasets/test_coco.py

@@ -449,3 +449,58 @@ def test_gpu_postprocess_flag(
 
         call_kwargs = mock_transforms.call_args.kwargs if mock_transforms.call_args else mock_transforms.call_args[1]
         assert call_kwargs["gpu_postprocess"] is expected_gpu_postprocess
+
+
+class TestAugConfigDisablesCrop:
+    """``aug_config={}`` disables the training resize-and-crop branch.
+
+    ``aug_config=None`` (the default) and any non-empty config keep it.
+    """
+
+    @pytest.mark.parametrize(
+        "aug_config,expected",
+        [
+            pytest.param(None, True, id="none_keeps_crop"),
+            pytest.param({}, False, id="empty_disables_crop"),
+            pytest.param({"HorizontalFlip": {"p": 0.5}}, True, id="nonempty_keeps_crop"),
+        ],
+    )
+    def test_crop_branch_enabled(self, aug_config, expected):
+        """_crop_branch_enabled maps aug_config to the crop-branch decision."""
+        from rfdetr.datasets.coco import _crop_branch_enabled
+
+        assert _crop_branch_enabled(aug_config) is expected
+
+    def test_empty_aug_config_warns(self):
+        """Passing aug_config={} logs a warning about the dropped resize-and-crop branch."""
+        from unittest.mock import patch
+
+        from rfdetr.datasets.coco import _crop_branch_enabled
+
+        with patch("rfdetr.datasets.coco.logger") as mock_logger:
+            _crop_branch_enabled({})
+        mock_logger.warning.assert_called_once()
+
+    @pytest.mark.parametrize(
+        "aug_config,expected",
+        [
+            pytest.param(None, True, id="none_keeps_crop"),
+            pytest.param({}, False, id="empty_disables_crop"),
+        ],
+    )
+    def test_make_coco_transforms_forwards_crop_decision(self, aug_config, expected):
+        """make_coco_transforms passes the aug_config-derived value to _build_train_resize_config."""
+        from unittest.mock import patch
+
+        from rfdetr.datasets.coco import make_coco_transforms
+
+        # Patch AlbumentationsWrapper.from_config so the test exercises only argument
+        # forwarding; without it the mocked empty configs trigger unrelated
+        # "Empty augmentation config provided" warnings.
+        with (
+            patch("rfdetr.datasets.coco._build_train_resize_config", return_value=[]) as mock_build,
+            patch("rfdetr.datasets.coco.AlbumentationsWrapper.from_config", return_value=[]),
+        ):
+            make_coco_transforms("train", 640, aug_config=aug_config)
+
+        assert mock_build.call_args.kwargs["include_crop_branch"] is expected

tests/datasets/test_coco_resize_config.py

@@ -195,3 +195,18 @@ def test_custom_max_size_propagates_to_both_options(self):
         option_b = result[0]["OneOf"]["transforms"][1]
         assert option_a["Sequential"]["transforms"][1] == {"LongestMaxSize": {"max_size": 1000}}
         assert option_b["Sequential"]["transforms"][3] == {"LongestMaxSize": {"max_size": 1000}}
+
+
+class TestBuildTrainResizeConfigCropBranch:
+    """include_crop_branch=False drops the resize-and-crop branch so only Option A runs."""
+
+    @pytest.mark.parametrize(
+        "square",
+        [pytest.param(True, id="square"), pytest.param(False, id="nonsquare")],
+    )
+    def test_include_crop_branch_false_drops_crop_branch(self, square):
+        """No RandomCrop/RandomSizedCrop appears in either square or non-square pipelines."""
+        result = _build_train_resize_config([480, 640], square=square, include_crop_branch=False)
+        flat = str(result)
+        assert "RandomSizedCrop" not in flat
+        assert "RandomCrop" not in flat