Support tiling

application/backend/app/execution/training/getitune_trainer.py

@@ -22,7 +22,7 @@
 
 if TYPE_CHECKING:
     from getitune import TaskType
-    from getitune.config.data import SubsetConfig
+    from getitune.config.data import SubsetConfig, TileConfig
     from getitune.data.dataset.base import VisionDataset
     from getitune.engine.engine import Engine
 
@@ -90,6 +90,7 @@ class DatasetInfo:
     getitune_validation_subset_config: SubsetConfig
     getitune_testing_subset_config: SubsetConfig
     revision_id: UUID
+    tile_config: TileConfig
 
 
 @dataclass(frozen=True)
@@ -243,7 +244,7 @@ def prepare_training_configuration(
             return training_config, getitune_training_config
 
     @step("Prepare Training Dataset", 10)
-    def prepare_training_dataset(
+    def prepare_training_dataset(  # noqa: PLR0915
         self,
         project_id: UUID,
         task: Task,
@@ -258,7 +259,9 @@ def prepare_training_dataset(
         Otherwise, it creates a new dataset from the current items in the database with user-verified annotations.
         """
 
-        from getitune.config.data import SamplerConfig, SubsetConfig
+        from getitune import TaskType
+        from getitune.config.data import SamplerConfig, SubsetConfig, TileConfig
+        from getitune.data.dataset.tile import TileDatasetFactory
         from getitune.data.entity.utils import detect_storage_dtype
         from getitune.data.factory import TransformLibFactory
 
@@ -349,6 +352,21 @@ def build_subset_config(subset_name: str) -> SubsetConfig:
                 transforms=test_subset_config.transforms,  # pyrefly: ignore[missing-attribute,bad-argument-type]
             )
 
+            # Build a TileConfig, and wrap each subset dataset with the tiling factory (if tiling is enabled).
+            tile_cfg_data = getitune_training_config["data"].get("tile_config", {})
+            tile_config = TileConfig(**tile_cfg_data)
+            if tile_config.enable_tiler and getitune_task_type in (TaskType.DETECTION, TaskType.INSTANCE_SEGMENTATION):
+                logger.info("Tiling is enabled - wrapping subset datasets with TileDatasetFactory")
+                getitune_training_dataset = TileDatasetFactory.create(
+                    dataset=getitune_training_dataset, tile_config=tile_config
+                )
+                getitune_validation_dataset = TileDatasetFactory.create(
+                    dataset=getitune_validation_dataset, tile_config=tile_config
+                )
+                getitune_testing_dataset = TileDatasetFactory.create(
+                    dataset=getitune_testing_dataset, tile_config=tile_config
+                )
+
             return DatasetInfo(
                 getitune_training_dataset=getitune_training_dataset,
                 getitune_validation_dataset=getitune_validation_dataset,
@@ -357,6 +375,7 @@ def build_subset_config(subset_name: str) -> SubsetConfig:
                 getitune_validation_subset_config=val_subset_config,
                 getitune_testing_subset_config=test_subset_config,
                 revision_id=dataset_revision_id,
+                tile_config=tile_config,
             )
 
     @step("Prepare Model")
@@ -416,6 +435,8 @@ def train_model(
             val_subset=dataset_info.getitune_validation_subset_config,
             test_subset=dataset_info.getitune_testing_subset_config,
         )
+        # Ensure the datamodule (and downstream model) uses the resolved tiling configuration.
+        datamodule.tile_config = dataset_info.tile_config
 
         logger.info("Instantiating model for training (model_id={})", model_id)
         model_cfg = training_config["model"]

application/backend/app/supported_models/manifests/detection/dfine_l.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 54.7
     coco_map_50: 72.4
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0005

application/backend/app/supported_models/manifests/detection/dfine_m.yaml

@@ -13,9 +13,6 @@ stats:
   benchmark_metrics:
     coco_map_50_95: 52.7
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0004

application/backend/app/supported_models/manifests/detection/dfine_x.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 56.5
     coco_map_50: 74.0
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0005

application/backend/app/supported_models/manifests/detection/dinov3_detr_l.yaml

@@ -13,9 +13,6 @@ stats:
   benchmark_metrics:
     coco_map_50_95: 56.0
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0005

application/backend/app/supported_models/manifests/detection/dinov3_detr_m.yaml

@@ -13,9 +13,6 @@ stats:
   benchmark_metrics:
     coco_map_50_95: 53.0
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0004

application/backend/app/supported_models/manifests/detection/dinov3_detr_s.yaml

@@ -13,9 +13,6 @@ stats:
   benchmark_metrics:
     coco_map_50_95: 50.9
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0004

application/backend/app/supported_models/manifests/detection/rfdetr_l.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 56.5
     coco_map_50: 75.1
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0001

application/backend/app/supported_models/manifests/detection/rfdetr_m.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 54.7
     coco_map_50: 73.6
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.00021

application/backend/app/supported_models/manifests/detection/rfdetr_n.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 48.4
     coco_map_50: 67.6
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.00021

application/backend/app/supported_models/manifests/detection/rfdetr_s.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 53.0
     coco_map_50: 72.1
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.00021

application/backend/app/supported_models/manifests/instance_segmentation/rfdetr_2xl.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 49.9
     coco_map_50: 73.1
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0001

application/backend/app/supported_models/manifests/instance_segmentation/rfdetr_l.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 47.1
     coco_map_50: 70.5
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0001

application/backend/app/supported_models/manifests/instance_segmentation/rfdetr_m.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 45.3
     coco_map_50: 68.4
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0001

application/backend/app/supported_models/manifests/instance_segmentation/rfdetr_n.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 40.3
     coco_map_50: 63.0
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0001

application/backend/app/supported_models/manifests/instance_segmentation/rfdetr_s.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 43.1
     coco_map_50: 66.2
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0001

application/backend/app/supported_models/manifests/instance_segmentation/rfdetr_xl.yaml

@@ -14,9 +14,6 @@ stats:
     coco_map_50_95: 48.8
     coco_map_50: 72.2
 
-capabilities:
-  tiling: false
-
 hyperparameters:
   training:
     learning_rate: 0.0001

application/backend/tests/integration/services/test_training_configuration_service.py

@@ -148,9 +148,9 @@ def test_get_default_by_model_architecture_strips_tiling_when_unsupported(
         """Tiling parameters should be removed from the default config when the architecture doesn't support tiling."""
         TrainingConfigurationService.get_default_by_model_architecture.cache_clear()
 
-        # D-FINE-M does not support tiling (capabilities.tiling: false in its manifest)
+        # YOLO11-N does not support tiling (capabilities.tiling: false in its manifest)
         result = TrainingConfigurationService.get_default_by_model_architecture(
-            model_architecture_id="object-detection-dfine-m",
+            model_architecture_id="object-detection-yolo11-n",
         )
 
         assert result.algo_level_parameters.dataset_preparation.augmentation.tiling is None
@@ -183,15 +183,15 @@ def test_get_by_model_architecture_strips_tiling_when_unsupported(
         algo_level_config = TrainingConfigurationDB(
             id=str(uuid4()),
             project_id=fxt_project.id,
-            model_architecture_id="object-detection-dfine-m",
+            model_architecture_id="object-detection-yolo11-n",
             configuration_data=algo_level_data,
         )
         db_session.add(algo_level_config)
         db_session.flush()
 
         result = fxt_training_configuration_service.get_by_model_architecture(
             project_id=UUID(fxt_project.id),
-            model_architecture_id="object-detection-dfine-m",
+            model_architecture_id="object-detection-yolo11-n",
         )
 
         assert result.algo_level_parameters.dataset_preparation.augmentation.tiling is None

library/src/getitune/recipe/detection/deim_dfine_l_tile.yaml

@@ -0,0 +1,115 @@
+task: DETECTION
+model:
+  class_path: getitune.backend.lightning.models.detection.DEIMDFine
+  init_args:
+    model_name: deim_dfine_hgnetv2_l
+    label_info: 80
+    multi_scale: false
+
+    optimizer:
+      class_path: torch.optim.AdamW
+      init_args:
+        lr: 0.0005
+        betas: [0.9, 0.999]
+        weight_decay: 0.000125
+
+    scheduler:
+      class_path: getitune.backend.lightning.schedulers.LinearWarmupSchedulerCallable
+      init_args:
+        num_warmup_steps: 30
+        main_scheduler_callable:
+          class_path: lightning.pytorch.cli.ReduceLROnPlateau
+          init_args:
+            mode: max
+            factor: 0.5
+            patience: 10
+            monitor: val/map_50
+
+engine:
+  device: auto
+
+callback_monitor: val/map_50
+
+# NOTE: Tiling is incompatible with mosaic/mixup based augmentation scheduling,
+# so the AugmentationSchedulerCallback used by the non-tiling recipe is omitted here
+# and a static tiling-friendly augmentation pipeline is defined in overrides.
+data: ../_base_/data/detection_tile.yaml
+
+callbacks:
+  - class_path: getitune.backend.lightning.callbacks.adaptive_train_scheduling.AdaptiveTrainScheduling
+    init_args:
+      max_interval: 1
+      min_lrschedule_patience: 3
+  - class_path: getitune.backend.lightning.callbacks.adaptive_early_stopping.EarlyStoppingWithWarmup
+    init_args:
+      mode: max
+      patience: 15
+      min_delta: 0.001
+      warmup_iters: 30
+      warmup_epochs: 3
+  - class_path: lightning.pytorch.callbacks.ModelCheckpoint
+    init_args:
+      dirpath: "" # use engine.work_dir
+      monitor: val/map_50
+      mode: max
+      save_top_k: 1
+      save_last: true
+      auto_insert_metric_name: false
+      filename: "checkpoints/epoch_{epoch:03d}"
+
+overrides:
+  callbacks:
+    - class_path: getitune.backend.lightning.callbacks.adaptive_train_scheduling.AdaptiveTrainScheduling
+      init_args:
+        max_interval: 1
+        min_lrschedule_patience: 3
+    - class_path: getitune.backend.lightning.callbacks.adaptive_early_stopping.EarlyStoppingWithWarmup
+      init_args:
+        warmup_iters: 100
+        warmup_epochs: 7
+
+  data:
+    input_size:
+      - 640
+      - 640
+    task: DETECTION
+    train_subset:
+      batch_size: 8
+      augmentations_cpu:
+        - class_path: torchvision.transforms.v2.RandomZoomOut
+          init_args:
+            fill: 0
+            p: 0.5
+        - class_path: getitune.data.augmentation.transforms.RandomIoUCrop
+        - class_path: torchvision.transforms.v2.SanitizeBoundingBoxes
+          init_args:
+            min_size: 1
+        - class_path: getitune.data.augmentation.transforms.Resize
+          init_args:
+            size: $(input_size)
+            keep_aspect_ratio: false
+      augmentations_gpu:
+        - class_path: kornia.augmentation.RandomHorizontalFlip
+          init_args:
+            p: 0.5
+      sampler:
+        class_path: getitune.data.samplers.balanced_sampler.BalancedSampler
+
+    val_subset:
+      batch_size: 8
+      augmentations_cpu:
+        - class_path: getitune.data.augmentation.transforms.Resize
+          init_args:
+            size: $(input_size)
+            keep_aspect_ratio: false
+            resize_targets: false
+      augmentations_gpu: []
+    test_subset:
+      batch_size: 8
+      augmentations_cpu:
+        - class_path: getitune.data.augmentation.transforms.Resize
+          init_args:
+            size: $(input_size)
+            keep_aspect_ratio: false
+            resize_targets: false
+      augmentations_gpu: []