Remove config from torch inferencer

README.md

@@ -192,7 +192,6 @@ Example OpenVINO Inference:
 
 ```bash
 python tools/inference/openvino_inference.py \
-    --config src/anomalib/models/padim/config.yaml \
     --weights results/padim/mvtec/bottle/run/openvino/model.bin \
     --metadata results/padim/mvtec/bottle/run/openvino/metadata.json \
     --input datasets/MVTec/bottle/test/broken_large/000.png \
@@ -207,7 +206,6 @@ A quick example:
 
 ```bash
 python tools/inference/gradio_inference.py \
-        --config src/anomalib/models/padim/config.yaml \
         --weights results/padim/mvtec/bottle/run/weights/model.ckpt
 ```
 

notebooks/400_openvino/401_nncf.ipynb

@@ -152,7 +152,7 @@
     "\n",
     "```yaml\n",
     "optimization:\n",
-    "  export_mode: null #options: onnx, openvino\n",
+    "  export_mode: null # options: torch, onnx, openvino\n",
     "  nncf:\n",
     "    apply: true\n",
     "    input_info:\n",
@@ -282,7 +282,7 @@
     "\n",
     "```yaml\n",
     "optimization:\n",
-    "  export_mode: null #options: onnx, openvino\n",
+    "  export_mode: null # options: torch, onnx, openvino\n",
     "  nncf:\n",
     "    apply: true\n",
     "    input_info:\n",

src/anomalib/deploy/export.py

@@ -25,6 +25,7 @@ class ExportMode(str, Enum):
 
     ONNX = "onnx"
     OPENVINO = "openvino"
+    TORCH = "torch"
 
 
 def get_model_metadata(model: AnomalyModule) -> dict[str, Tensor]:
@@ -59,6 +60,7 @@ def get_metadata(task: TaskType, transform: dict[str, Any], model: AnomalyModule
         task (TaskType): Task type.
         transform (dict[str, Any]): Transform used for the model.
         model (AnomalyModule): Model to export.
+        export_mode (ExportMode): Mode to export the model. Torch, ONNX or OpenVINO.
 
     Returns:
         dict[str, Any]: Metadata for the exported model.
@@ -90,20 +92,44 @@ def export(
         transform (dict[str, Any]): Data transforms (augmentatiions) used for the model.
         input_size (tuple[int, int]): Input size of the model.
         model (AnomalyModule): Anomaly model to export.
-        export_mode (ExportMode): Mode to export the model. ONNX or OpenVINO.
-        export_root (str | Path): Path to exported ONNX/OpenVINO IR.
+        export_mode (ExportMode): Mode to export the model. Torch, ONNX or OpenVINO.
+        export_root (str | Path): Path to exported Torch, ONNX or OpenVINO IR.
     """
-    # Write metadata to json file. The file is written in the same directory as the target model.
-    export_path = Path(export_root) / export_mode.value
+    # Create export directory.
+    export_path = Path(export_root) / "weights" / export_mode.value
     export_path.mkdir(parents=True, exist_ok=True)
-    with (Path(export_path) / "metadata.json").open("w", encoding="utf-8") as metadata_file:
-        metadata = get_metadata(task, transform, model)
-        json.dump(metadata, metadata_file, ensure_ascii=False, indent=4)
-
-    # Export model to onnx and convert to OpenVINO IR if export mode is set to OpenVINO.
-    onnx_path = export_to_onnx(model, input_size, export_path)
-    if export_mode == ExportMode.OPENVINO:
-        export_to_openvino(export_path, onnx_path)
+
+    # Get metadata.
+    metadata = get_metadata(task, transform, model)
+
+    if export_mode == ExportMode.TORCH:
+        export_to_torch(model, metadata, export_path)
+
+    elif export_mode in (ExportMode.ONNX, ExportMode.OPENVINO):
+        # Write metadata to json file. The file is written in the same directory as the target model.
+        with (Path(export_path) / "metadata.json").open("w", encoding="utf-8") as metadata_file:
+            json.dump(metadata, metadata_file, ensure_ascii=False, indent=4)
+
+        # Export model to onnx and convert to OpenVINO IR if export mode is set to OpenVINO.
+        onnx_path = export_to_onnx(model, input_size, export_path)
+        if export_mode == ExportMode.OPENVINO:
+            export_to_openvino(export_path, onnx_path)
+
+    else:
+        raise ValueError(f"Unknown export mode {export_mode}")
+
+
+def export_to_torch(model: AnomalyModule, metadata: dict[str, Any], export_path: Path) -> None:
+    """Export AnomalibModel to torch.
+
+    Args:
+        model (AnomalyModule): Model to export.
+        export_path (Path): Path to the folder storing the exported model.
+    """
+    torch.save(
+        obj={"model": model.model, "metadata": metadata},
+        f=export_path / "model.pt",
+    )
 
 
 def export_to_onnx(model: AnomalyModule, input_size: tuple[int, int], export_path: Path) -> Path:

src/anomalib/deploy/inferencers/torch_inferencer.py

@@ -8,19 +8,15 @@
 from pathlib import Path
 from typing import Any
 
+import albumentations as A
 import cv2
 import numpy as np
 import torch
-from omegaconf import DictConfig, ListConfig
-from torch import Tensor
+from omegaconf import DictConfig
+from torch import Tensor, nn
 
-from anomalib.config import get_configurable_parameters
 from anomalib.data import TaskType
-from anomalib.data.utils import InputNormalizationMethod, get_transforms
 from anomalib.data.utils.boxes import masks_to_boxes
-from anomalib.deploy.export import get_model_metadata
-from anomalib.models import get_model
-from anomalib.models.components import AnomalyModule
 
 from .base_inferencer import Inferencer
 
@@ -29,38 +25,20 @@ class TorchInferencer(Inferencer):
     """PyTorch implementation for the inference.
 
     Args:
-        config (str | Path | DictConfig | ListConfig): Configurable parameters that are used
-            during the training stage.
-        model_source (str | Path | AnomalyModule): Path to the model ckpt file or the Anomaly model.
-        metadata_path (str | Path, optional): Path to metadata file. If none, it tries to load the params
-                from the model state_dict. Defaults to None.
+        path (str | Path | AnomalyModule): Path to the model ckpt file or the Anomaly model.
         device (str | None, optional): Device to use for inference. Options are auto, cpu, cuda. Defaults to "auto".
     """
 
     def __init__(
         self,
-        config: str | Path | DictConfig | ListConfig,
-        model_source: str | Path | AnomalyModule,
-        metadata_path: str | Path | None = None,
+        path: str | Path,
         device: str = "auto",
     ) -> None:
         self.device = self._get_device(device)
 
-        # Check and load the configuration
-        if isinstance(config, (str, Path)):
-            self.config = get_configurable_parameters(config_path=config)
-        elif isinstance(config, (DictConfig, ListConfig)):
-            self.config = config
-        else:
-            raise ValueError(f"Unknown config type {type(config)}")
-
-        # Check and load the model weights.
-        if isinstance(model_source, AnomalyModule):
-            self.model = model_source
-        else:
-            self.model = self.load_model(model_source)
-
-        self.metadata = self._load_metadata(metadata_path)
+        # Load the model weights.
+        self.model = self.load_model(path)
+        self.metadata = self._load_metadata(path)
 
     @staticmethod
     def _get_device(device: str) -> torch.device:
@@ -82,24 +60,18 @@ def _get_device(device: str) -> torch.device:
         return torch.device(device)
 
     def _load_metadata(self, path: str | Path | None = None) -> dict | DictConfig:
-        """Load metadata from file or from model state dict.
+        """Load metadata from file.
 
         Args:
-            path (str | Path | None, optional): Path to metadata file. If none, it tries to load the params
-                from the model state_dict. Defaults to None.
+            path (str | Path): Path to the model pt file.
 
         Returns:
             dict: Dictionary containing the metadata.
         """
-        metadata: dict[str, float | np.ndarray | Tensor] | DictConfig
-        if path is None:
-            # Torch inferencer still reads metadata from the model.
-            metadata = get_model_metadata(self.model)
-        else:
-            metadata = super()._load_metadata(path)
+        metadata = torch.load(path, map_location=self.device)["metadata"] if path else {}
         return metadata
 
-    def load_model(self, path: str | Path) -> AnomalyModule:
+    def load_model(self, path: str | Path) -> nn.Module:
         """Load the PyTorch model.
 
         Args:
@@ -108,8 +80,8 @@ def load_model(self, path: str | Path) -> AnomalyModule:
         Returns:
             (AnomalyModule): PyTorch Lightning model.
         """
-        model = get_model(self.config)
-        model.load_state_dict(torch.load(path, map_location=self.device)["state_dict"])
+
+        model = torch.load(path, map_location=self.device)["model"]
         model.eval()
         return model.to(self.device)
 
@@ -122,18 +94,7 @@ def pre_process(self, image: np.ndarray) -> Tensor:
         Returns:
             Tensor: pre-processed image.
         """
-        transform_config = (
-            self.config.dataset.transform_config.eval if "transform_config" in self.config.dataset.keys() else None
-        )
-
-        image_size = (self.config.dataset.image_size[0], self.config.dataset.image_size[1])
-        center_crop = self.config.dataset.get("center_crop")
-        if center_crop is not None:
-            center_crop = tuple(center_crop)
-        normalization = InputNormalizationMethod(self.config.dataset.normalization)
-        transform = get_transforms(
-            config=transform_config, image_size=image_size, center_crop=center_crop, normalization=normalization
-        )
+        transform = A.from_dict(self.metadata["transform"])
         processed_image = transform(image=image)["image"]
 
         if len(processed_image) == 3:
@@ -209,7 +170,7 @@ def post_process(self, predictions: Tensor, metadata: dict | DictConfig | None =
             if pred_mask is not None:
                 pred_mask = cv2.resize(pred_mask, (image_width, image_height))
 
-        if self.config.dataset.task == TaskType.DETECTION:
+        if self.metadata["task"] == TaskType.DETECTION:
             pred_boxes = masks_to_boxes(torch.from_numpy(pred_mask))[0][0].numpy()
             box_labels = np.ones(pred_boxes.shape[0])
         else:

src/anomalib/models/cfa/config.yaml

@@ -62,7 +62,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 
 # PL Trainer Args. Don't add extra parameter here.
 trainer:

src/anomalib/models/cflow/config.yaml

@@ -68,7 +68,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 
 # PL Trainer Args. Don't add extra parameter here.
 trainer:

src/anomalib/models/csflow/config.yaml

@@ -70,7 +70,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 # PL Trainer Args. Don't add extra parameter here.
 trainer:
   enable_checkpointing: true

src/anomalib/models/dfkde/config.yaml

@@ -56,7 +56,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 # PL Trainer Args. Don't add extra parameter here.
 trainer:
   enable_checkpointing: true

src/anomalib/models/dfm/config.yaml

@@ -57,7 +57,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 # PL Trainer Args. Don't add extra parameter here.
 trainer:
   enable_checkpointing: true

src/anomalib/models/draem/config.yaml

@@ -65,7 +65,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 # PL Trainer Args. Don't add extra parameter here.
 trainer:
   enable_checkpointing: true

src/anomalib/models/fastflow/config.yaml

@@ -68,7 +68,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 
 # PL Trainer Args. Don't add extra parameter here.
 trainer:

src/anomalib/models/padim/config.yaml

@@ -63,7 +63,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 
 # PL Trainer Args. Don't add extra parameter here.
 trainer:

src/anomalib/models/reverse_distillation/config.yaml

@@ -72,7 +72,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 
 # PL Trainer Args. Don't add extra parameter here.
 trainer:

src/anomalib/models/rkde/config.yaml

@@ -66,7 +66,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 # PL Trainer Args. Don't add extra parameter here.
 trainer:
   enable_checkpointing: true

src/anomalib/models/stfpm/config.yaml

@@ -70,7 +70,7 @@ logging:
   log_graph: false # Logs the model graph to respective logger.
 
 optimization:
-  export_mode: null #options: onnx, openvino
+  export_mode: null # options: torch, onnx, openvino
 # PL Trainer Args. Don't add extra parameter here.
 trainer:
   enable_checkpointing: true

src/anomalib/utils/callbacks/__init__.py

@@ -61,7 +61,7 @@ def get_callbacks(config: DictConfig | ListConfig) -> list[Callback]:
     monitor_mode = "max" if "early_stopping" not in config.model.keys() else config.model.early_stopping.mode
 
     checkpoint = ModelCheckpoint(
-        dirpath=os.path.join(config.project.path, "weights"),
+        dirpath=os.path.join(config.project.path, "weights", "lightning"),
         filename="model",
         monitor=monitor_metric,
         mode=monitor_mode,