Merge pull request #3748 from sungchul2/enable-data-config-override

Enable to override data configurations

CHANGELOG.md

@@ -8,6 +8,9 @@ All notable changes to this project will be documented in this file.
 
 ### Enhancements
 
+- Enable to override data configurations
+  (<https://github.com/openvinotoolkit/training_extensions/pull/3748>)
+
 ### Bug fixes
 
 ## \[v2.1.0\]

src/otx/cli/utils/jsonargparse.py

@@ -1,6 +1,6 @@
 """Functions related to jsonargparse."""
 
-# Copyright (C) 2023 Intel Corporation
+# Copyright (C) 2023-2024 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
 from __future__ import annotations
@@ -136,22 +136,48 @@ def apply_config(self: ActionConfigFile, parser: ArgumentParser, cfg: Namespace,
         cfg.__dict__.update(cfg_merged.__dict__)
         overrides = cfg.__dict__.pop("overrides", None)
         if overrides is not None:
-            # This is a feature to handle the callbacks & logger override for user-convinience
-            list_override(configs=cfg, key="callbacks", overrides=overrides.pop("callbacks", []))
-            list_override(configs=cfg, key="logger", overrides=overrides.pop("logger", []))
+            apply_override(cfg, overrides)
             cfg.update(overrides)
         if cfg.get(dest) is None:
             cfg[dest] = []
         cfg[dest].append(cfg_path)
 
 
-def list_override(configs: Namespace, key: str, overrides: list) -> None:
+def namespace_override(
+    configs: Namespace,
+    key: str,
+    overrides: Namespace,
+    convert_dict_to_namespace: bool = True,
+) -> None:
+    """Overrides the nested namespace type in the given configs with the provided overrides.
+
+    Args:
+        configs (Namespace): The configuration object containing the key.
+        key (str): key of the configs want to override.
+        overrides (Namespace): The configuration object to override the existing ones.
+        convert_dict_to_namespace (bool): Whether to convert the dictionary to Namespace. Defaults to True.
+    """
+    for sub_key, sub_value in overrides.items():
+        if isinstance(sub_value, list) and all(isinstance(sv, dict) for sv in sub_value):
+            # only enable list of dictionary items
+            list_override(
+                configs=configs[key],
+                key=sub_key,
+                overrides=sub_value,
+                convert_dict_to_namespace=convert_dict_to_namespace,
+            )
+        else:
+            configs[key].update(sub_value, sub_key)
+
+
+def list_override(configs: Namespace, key: str, overrides: list, convert_dict_to_namespace: bool = True) -> None:
     """Overrides the nested list type in the given configs with the provided override_list.
 
     Args:
         configs (Namespace): The configuration object containing the key.
         key (str): key of the configs want to override.
         overrides (list): The list of dictionary item to override the existing ones.
+        convert_dict_to_namespace (bool): Whether to convert the dictionary to Namespace. Defaults to True.
 
     Example:
         >>> configs = [
@@ -179,6 +205,28 @@ def list_override(configs: Namespace, key: str, overrides: list) -> None:
         ...     ),
         ...     ...
         ... ]
+        >>> append_callbacks = [
+        ...     {
+        ...         'class_path': 'new_callbacks',
+        ...     },
+        ... ]
+        >>> list_override(configs=configs, key="callbacks", overrides=append_callbacks)
+        >>> configs = [
+        ...     ...
+        ...     Namespace(class_path='new_callbacks'),
+        ... ]
+        >>> append_callbacks_as_dict = [
+        ...     {
+        ...         'class_path': 'new_callbacks1',
+        ...     },
+        ... ]
+        >>> list_override(
+        ...     configs=configs, key="callbacks", overrides=append_callbacks_as_dict, convert_dict_to_namespace=False
+        ... )
+        >>> configs = [
+        ...     ...
+        ...     {'class_path': 'new_callbacks1'},
+        ... ]
     """
     if key not in configs or configs[key] is None:
         return
@@ -192,7 +240,40 @@ def list_override(configs: Namespace, key: str, overrides: list) -> None:
         if item is not None:
             Namespace(item).update(target)
         else:
-            configs[key].append(dict_to_namespace(target))
+            converted_target = dict_to_namespace(target) if convert_dict_to_namespace else target
+            configs[key].append(converted_target)
+
+
+def apply_override(cfg: Namespace, overrides: Namespace) -> None:
+    """Overrides the provided overrides in the given configs.
+
+    Args:
+        configs (Namespace): The configuration object containing the key.
+        overrides (Namespace): The configuration object to override the existing ones.
+    """
+    # replace the config with the overrides for keys in reset list
+    reset = overrides.pop("reset", [])
+    if isinstance(reset, str):
+        reset = [reset]
+    for key in reset:
+        if key in overrides:
+            # callbacks, logger -> update to namespace
+            # rest -> use dict as is
+            cfg[key] = (
+                [dict_to_namespace(o) for o in overrides.pop(key)]
+                if key in ("callbacks", "logger")
+                else overrides.pop(key)
+            )
+
+    # This is a feature to handle the callbacks, logger, and data override for user-convinience
+    list_override(configs=cfg, key="callbacks", overrides=overrides.pop("callbacks", []))
+    list_override(configs=cfg, key="logger", overrides=overrides.pop("logger", []))
+    namespace_override(
+        configs=cfg,
+        key="data",
+        overrides=overrides.pop("data", Namespace()),
+        convert_dict_to_namespace=False,
+    )
 
 
 # [FIXME] harimkang: have to see if there's a better way to do it. (For now, Added 2 lines to existing function)
@@ -235,8 +316,7 @@ def get_defaults_with_overrides(self: ArgumentParser, skip_check: bool = False)
             cfg_file = self._load_config_parser_mode(default_config_file.get_content(), key=key)
             cfg = self.merge_config(cfg_file, cfg)
             overrides = cfg.__dict__.pop("overrides", {})
-            list_override(configs=cfg, key="callbacks", overrides=overrides.pop("callbacks", []))
-            list_override(configs=cfg, key="logger", overrides=overrides.pop("logger", []))
+            apply_override(cfg, overrides)
             if overrides is not None:
                 cfg.update(overrides)
             try:

src/otx/recipe/_base_/data/detection.yaml

@@ -12,7 +12,24 @@ train_subset:
   num_workers: 2
   to_tv_image: false
   transforms:
-    - class_path: torchvision.transforms.v2.ToImage
+    - class_path: otx.core.data.transform_libs.torchvision.MinIoURandomCrop
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        scale:
+          - 800
+          - 992
+        transform_bbox: true
+    - class_path: otx.core.data.transform_libs.torchvision.RandomFlip
+      init_args:
+        prob: 0.5
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [0.0, 0.0, 0.0]
+        std: [255.0, 255.0, 255.0]
   sampler:
     class_path: torch.utils.data.RandomSampler
 
@@ -23,7 +40,19 @@ val_subset:
   num_workers: 2
   to_tv_image: false
   transforms:
-    - class_path: torchvision.transforms.v2.ToImage
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        scale:
+          - 800
+          - 992
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [0.0, 0.0, 0.0]
+        std: [255.0, 255.0, 255.0]
   sampler:
     class_path: torch.utils.data.RandomSampler
 
@@ -34,6 +63,18 @@ test_subset:
   num_workers: 2
   to_tv_image: false
   transforms:
-    - class_path: torchvision.transforms.v2.ToImage
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        scale:
+          - 800
+          - 992
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [0.0, 0.0, 0.0]
+        std: [255.0, 255.0, 255.0]
   sampler:
     class_path: torch.utils.data.RandomSampler

src/otx/recipe/_base_/data/instance_segmentation.yaml

@@ -9,32 +9,86 @@ unannotated_items_ratio: 0.0
 train_subset:
   subset_name: train
   transform_lib_type: TORCHVISION
-  to_tv_image: true
-  transforms:
-    - class_path: torchvision.transforms.v2.ToImage
   batch_size: 1
   num_workers: 2
+  to_tv_image: true
+  transforms:
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        keep_ratio: true
+        transform_bbox: true
+        transform_mask: true
+        scale:
+          - 1024
+          - 1024
+    - class_path: otx.core.data.transform_libs.torchvision.Pad
+      init_args:
+        pad_to_square: true
+        transform_mask: true
+    - class_path: otx.core.data.transform_libs.torchvision.RandomFlip
+      init_args:
+        prob: 0.5
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [123.675, 116.28, 103.53]
+        std: [58.395, 57.12, 57.375]
   sampler:
     class_path: torch.utils.data.RandomSampler
 
 val_subset:
   subset_name: val
   transform_lib_type: TORCHVISION
-  to_tv_image: true
-  transforms:
-    - class_path: torchvision.transforms.v2.ToImage
   batch_size: 1
   num_workers: 2
+  to_tv_image: true
+  transforms:
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        keep_ratio: true
+        scale:
+          - 1024
+          - 1024
+    - class_path: otx.core.data.transform_libs.torchvision.Pad
+      init_args:
+        pad_to_square: true
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [123.675, 116.28, 103.53]
+        std: [58.395, 57.12, 57.375]
   sampler:
     class_path: torch.utils.data.RandomSampler
 
 test_subset:
   subset_name: test
   transform_lib_type: TORCHVISION
-  to_tv_image: true
-  transforms:
-    - class_path: torchvision.transforms.v2.ToImage
   batch_size: 1
   num_workers: 2
+  to_tv_image: true
+  transforms:
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        keep_ratio: true
+        scale:
+          - 1024
+          - 1024
+    - class_path: otx.core.data.transform_libs.torchvision.Pad
+      init_args:
+        pad_to_square: true
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [123.675, 116.28, 103.53]
+        std: [58.395, 57.12, 57.375]
   sampler:
     class_path: torch.utils.data.RandomSampler

src/otx/recipe/_base_/data/rotated_detection.yaml

@@ -11,7 +11,29 @@ train_subset:
   transform_lib_type: TORCHVISION
   to_tv_image: false
   transforms:
-    - class_path: torchvision.transforms.v2.ToImage
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        keep_ratio: true
+        transform_bbox: true
+        transform_mask: true
+        scale:
+          - 1024
+          - 1024
+    - class_path: otx.core.data.transform_libs.torchvision.Pad
+      init_args:
+        size_divisor: 32
+        transform_mask: true
+    - class_path: otx.core.data.transform_libs.torchvision.RandomFlip
+      init_args:
+        prob: 0.5
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [123.675, 116.28, 103.53]
+        std: [58.395, 57.12, 57.375]
   batch_size: 1
   num_workers: 2
   sampler:
@@ -22,7 +44,23 @@ val_subset:
   transform_lib_type: TORCHVISION
   to_tv_image: false
   transforms:
-    - class_path: torchvision.transforms.v2.ToImage
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        keep_ratio: true
+        scale:
+          - 1024
+          - 1024
+    - class_path: otx.core.data.transform_libs.torchvision.Pad
+      init_args:
+        size_divisor: 32
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [123.675, 116.28, 103.53]
+        std: [58.395, 57.12, 57.375]
   batch_size: 1
   num_workers: 2
   sampler:
@@ -33,7 +71,23 @@ test_subset:
   transform_lib_type: TORCHVISION
   to_tv_image: false
   transforms:
-    - class_path: torchvision.transforms.v2.ToImage
+    - class_path: otx.core.data.transform_libs.torchvision.Resize
+      init_args:
+        keep_ratio: true
+        scale:
+          - 1024
+          - 1024
+    - class_path: otx.core.data.transform_libs.torchvision.Pad
+      init_args:
+        size_divisor: 32
+        is_numpy_to_tvtensor: true
+    - class_path: torchvision.transforms.v2.ToDtype
+      init_args:
+        dtype: ${as_torch_dtype:torch.float32}
+    - class_path: torchvision.transforms.v2.Normalize
+      init_args:
+        mean: [123.675, 116.28, 103.53]
+        std: [58.395, 57.12, 57.375]
   batch_size: 1
   num_workers: 2
   sampler: