[Refactor] Refactor transform's compose and arranges (#111)

* Remove arrang op load

* Update trans pipline

* Update tests

* Update tutorials

* Bugfix

* -

* Fix and update indices

* Fix bugs and typo

* Update requirements.txt

---------

Co-authored-by: Bobholamovic <mhlin425@whu.edu.cn>

paddlers/datasets/base.py

@@ -18,7 +18,7 @@ from paddle.io import Dataset
 from paddle.fluid.dataloader.collate import default_collate_fn
 
 from paddlers.utils import get_num_workers
-from paddlers.transforms import construct_sample_from_dict
+from paddlers.transforms import construct_sample_from_dict, Compose, DecodeImg
 
 
 class BaseDataset(Dataset):
@@ -30,6 +30,8 @@ class BaseDataset(Dataset):
         self.data_dir = data_dir
         self.label_list = label_list
         self.transforms = deepcopy(transforms)
+        self.normal_transforms(self.transforms)
+
         self.num_workers = get_num_workers(num_workers)
         self.shuffle = shuffle
 
@@ -46,3 +48,11 @@ class BaseDataset(Dataset):
                 [s[0] for s in batch]), [s[1] for s in batch]
         else:
             return default_collate_fn([s[0] for s in batch])
+
+    def normal_transforms(self, trans):
+        # NOTE: add `DecodeImg` and convert to `Compose`
+        if isinstance(trans, list):
+            trans = Compose(trans)
+        if not isinstance(trans.transforms[0], DecodeImg):
+            trans.transforms.insert(0, DecodeImg())
+        self.transforms = trans

paddlers/datasets/cd_dataset.py

@@ -31,7 +31,7 @@ class CDDataset(BaseDataset):
             the change mask. When `with_seg_labels` is True, each line in the file contains the paths of the
             bi-temporal images, the path of the change mask, and the paths of the segmentation masks in both
             temporal phases.
-        transforms (paddlers.transforms.Compose): Data preprocessing and data augmentation operators to apply.
+        transforms (paddlers.transforms.Compose|list): Data preprocessing and data augmentation operators to apply.
         label_list (str|None, optional): Path of the file that contains the category names. Defaults to None.
         num_workers (int|str, optional): Number of processes used for data loading. If `num_workers` is 'auto',
             the number of workers will be automatically determined according to the number of CPU cores: If 

paddlers/datasets/clas_dataset.py

@@ -25,7 +25,7 @@ class ClasDataset(BaseDataset):
     Args:
         data_dir (str): Root directory of the dataset.
         file_list (str): Path of the file that contains relative paths of images and labels.
-        transforms (paddlers.transforms.Compose): Data preprocessing and data augmentation operators to apply.
+        transforms (paddlers.transforms.Compose|list): Data preprocessing and data augmentation operators to apply.
         label_list (str|None, optional): Path of the file that contains the category names. Defaults to None.
         num_workers (int|str, optional): Number of processes used for data loading. If `num_workers` is 'auto',
             the number of workers will be automatically determined according to the number of CPU cores: If 

paddlers/datasets/coco.py

@@ -35,7 +35,7 @@ class COCODetDataset(BaseDataset):
         data_dir (str): Root directory of the dataset.
         image_dir (str): Directory that contains the images.
         ann_path (str): Path to COCO annotations.
-        transforms (paddlers.transforms.Compose): Data preprocessing and data augmentation operators to apply.
+        transforms (paddlers.transforms.Compose|list): Data preprocessing and data augmentation operators to apply.
         label_list (str|None, optional): Path of the file that contains the category names. Defaults to None.
         num_workers (int|str, optional): Number of processes used for data loading. If `num_workers` is 'auto',
             the number of workers will be automatically determined according to the number of CPU cores: If 

paddlers/datasets/res_dataset.py

@@ -26,7 +26,7 @@ class ResDataset(BaseDataset):
     Args:
         data_dir (str): Root directory of the dataset.
         file_list (str): Path of the file that contains relative paths of source and target image files.
-        transforms (paddlers.transforms.Compose): Data preprocessing and data augmentation operators to apply.
+        transforms (paddlers.transforms.Compose|list): Data preprocessing and data augmentation operators to apply.
         num_workers (int|str, optional): Number of processes used for data loading. If `num_workers` is 'auto',
             the number of workers will be automatically determined according to the number of CPU cores: If 
             there are more than 16 cores, 8 workers will be used. Otherwise, the number of workers will be half 

paddlers/datasets/seg_dataset.py

@@ -26,7 +26,7 @@ class SegDataset(BaseDataset):
     Args:
         data_dir (str): Root directory of the dataset.
         file_list (str): Path of the file that contains relative paths of images and annotation files.
-        transforms (paddlers.transforms.Compose): Data preprocessing and data augmentation operators to apply.
+        transforms (paddlers.transforms.Compose|list): Data preprocessing and data augmentation operators to apply.
         label_list (str|None, optional): Path of the file that contains the category names. Defaults to None.
         num_workers (int|str, optional): Number of processes used for data loading. If `num_workers` is 'auto',
             the number of workers will be automatically determined according to the number of CPU cores: If 

paddlers/datasets/voc.py

@@ -36,11 +36,11 @@ class VOCDetDataset(BaseDataset):
     Args:
         data_dir (str): Root directory of the dataset.
         file_list (str): Path of the file that contains relative paths of images and annotation files.
-        transforms (paddlers.transforms.Compose): Data preprocessing and data augmentation operators to apply.
+        transforms (paddlers.transforms.Compose|list): Data preprocessing and data augmentation operators to apply.
         label_list (str|None, optional): Path of the file that contains the category names. Defaults to None.
         num_workers (int|str, optional): Number of processes used for data loading. If `num_workers` is 'auto',
             the number of workers will be automatically determined according to the number of CPU cores: If 
-            there are more than 16 cores，8 workers will be used. Otherwise, the number of workers will be half 
+            there are more than 16 cores, 8 workers will be used. Otherwise, the number of workers will be half 
             the number of CPU cores. Defaults: 'auto'.
         shuffle (bool, optional): Whether to shuffle the samples. Defaults to False.
         allow_empty (bool, optional): Whether to add negative samples. Defaults to False.

paddlers/tasks/base.py

@@ -20,6 +20,7 @@ import math
 import json
 from functools import partial, wraps
 from inspect import signature
+from typing import Optional
 
 import yaml
 import paddle
@@ -29,6 +30,7 @@ from paddleslim.analysis import flops
 from paddleslim import L1NormFilterPruner, FPGMFilterPruner
 
 import paddlers
+from paddlers.transforms.operators import Compose, DecodeImg, Arrange
 import paddlers.utils.logging as logging
 from paddlers.utils import (
     seconds_to_hms, get_single_card_bs, dict2str, get_pretrain_weights,
@@ -61,7 +63,7 @@ class ModelMeta(type):
 
 
 class BaseModel(metaclass=ModelMeta):
-
+    _arrange: Optional[Arrange] = None
     find_unused_parameters = False
 
     def __init__(self, model_type):
@@ -204,13 +206,11 @@ class BaseModel(metaclass=ModelMeta):
                     else:
                         attr = op.__dict__
                     info['Transforms'].append({name: attr})
-                arrange = self.test_transforms.arrange
-                if arrange is not None:
-                    info['Transforms'].append({
-                        arrange.__class__.__name__: {
-                            'mode': 'test'
-                        }
-                    })
+                info['Transforms'].append({
+                    self._arrange.__name__: {
+                        'mode': 'test'
+                    }
+                })
         info['completed_epochs'] = self.completed_epochs
         return info
 
@@ -267,11 +267,28 @@ class BaseModel(metaclass=ModelMeta):
         open(osp.join(save_dir, '.success'), 'w').close()
         logging.info("Model saved in {}.".format(save_dir))
 
+    def _build_transforms(self, trans, mode):
+        if isinstance(trans, list):
+            trans = Compose(trans)
+        if not isinstance(trans.transforms[0], DecodeImg):
+            trans.transforms.insert(0, DecodeImg())
+        if self._arrange is Arrange or not issubclass(self._arrange, Arrange):
+            raise ValueError(
+                "`self._arrange` must be set to a concrete Arrange type.")
+        if trans.arrange is None:
+            # For backward compatibility, we only set `trans.arrange`
+            # when it is not set by user.
+            trans.arrange = self._arrange(mode)
+        return trans
+
     def build_data_loader(self,
                           dataset,
                           batch_size,
                           mode='train',
                           collate_fn=None):
+        # NOTE: Append `Arrange` to transforms
+        dataset.transforms = self._build_transforms(dataset.transforms, mode)
+
         if dataset.num_samples < batch_size:
             raise ValueError(
                 'The volume of dataset({}) must be larger than batch size({}).'
@@ -315,7 +332,7 @@ class BaseModel(metaclass=ModelMeta):
                    early_stop=False,
                    early_stop_patience=5,
                    use_vdl=True):
-        self._check_transforms(train_dataset.transforms, 'train')
+        self._check_transforms(train_dataset.transforms)
 
         # XXX: Hard-coding
         if self.model_type == 'detector' and 'RCNN' in self.__class__.__name__ and train_dataset.pos_num < len(
@@ -351,6 +368,7 @@ class BaseModel(metaclass=ModelMeta):
 
         self.train_data_loader = self.build_data_loader(
             train_dataset, batch_size=train_batch_size, mode='train')
+        self._check_arrange(self.train_data_loader.dataset.transforms, 'train')
 
         if eval_dataset is not None:
             self.test_transforms = copy.deepcopy(eval_dataset.transforms)
@@ -493,7 +511,7 @@ class BaseModel(metaclass=ModelMeta):
 
         assert criterion in {'l1_norm', 'fpgm'}, \
             "Pruning criterion {} is not supported. Please choose from {'l1_norm', 'fpgm'}."
-        self._check_transforms(dataset.transforms, 'eval')
+        self._check_transforms(dataset.transforms)
         # XXX: Hard-coding
         if self.model_type == 'detector':
             self.net.eval()
@@ -681,16 +699,19 @@ class BaseModel(metaclass=ModelMeta):
 
         return outputs
 
-    def _check_transforms(self, transforms, mode):
-        # NOTE: Check transforms and transforms.arrange and give user-friendly error messages.
-        if not isinstance(transforms, paddlers.transforms.Compose):
-            raise TypeError("`transforms` must be paddlers.transforms.Compose.")
+    def _check_transforms(self, transforms):
+        # NOTE: Check transforms
+        if not isinstance(transforms, Compose):
+            raise TypeError(
+                "`transforms` must be `paddlers.transforms.Compose`.")
+
+    def _check_arrange(self, transforms, mode):
         arrange_obj = transforms.arrange
-        if not isinstance(arrange_obj, paddlers.transforms.operators.Arrange):
+        if not isinstance(arrange_obj, Arrange):
             raise TypeError("`transforms.arrange` must be an Arrange object.")
         if arrange_obj.mode != mode:
             raise ValueError(
-                f"Incorrect arrange mode! Expected {mode} but got {arrange_obj.mode}."
+                f"Incorrect arrange mode! Expected {repr(mode)} but got {repr(arrange_obj.mode)}."
             )
 
     def run(self, net, inputs, mode):

paddlers/tasks/change_detector.py

@@ -29,6 +29,7 @@ import paddlers.rs_models.cd as cmcd
 import paddlers.utils.logging as logging
 from paddlers.models import seg_losses
 from paddlers.transforms import Resize, decode_image, construct_sample
+from paddlers.transforms.operators import ArrangeChangeDetector
 from paddlers.utils import get_single_card_bs
 from paddlers.utils.checkpoint import cd_pretrain_weights_dict
 from .base import BaseModel
@@ -43,6 +44,8 @@ __all__ = [
 
 
 class BaseChangeDetector(BaseModel):
+    _arrange = ArrangeChangeDetector
+
     def __init__(self,
                  model_name,
                  num_classes=2,
@@ -426,7 +429,7 @@ class BaseChangeDetector(BaseModel):
                     "category_F1-score": F1 score}.
         """
 
-        self._check_transforms(eval_dataset.transforms, 'eval')
+        self._check_transforms(eval_dataset.transforms)
 
         self.net.eval()
         nranks = paddle.distributed.get_world_size()
@@ -448,6 +451,7 @@ class BaseChangeDetector(BaseModel):
             )
         self.eval_data_loader = self.build_data_loader(
             eval_dataset, batch_size=batch_size, mode='eval')
+        self._check_arrange(self.eval_data_loader.dataset.transforms, 'eval')
 
         intersect_area_all = 0
         pred_area_all = 0
@@ -531,7 +535,7 @@ class BaseChangeDetector(BaseModel):
             img_file (list[tuple] | tuple[str|np.ndarray]): Tuple of image paths or 
                 decoded image data for bi-temporal images, which also could constitute
                 a list, meaning all image pairs to be predicted as a mini-batch.
-            transforms (paddlers.transforms.Compose|None, optional): Transforms for 
+            transforms (paddlers.transforms.Compose|list|None, optional): Transforms for 
                 inputs. If None, the transforms for evaluation process will be used. 
                 Defaults to None.
 
@@ -557,6 +561,8 @@ class BaseChangeDetector(BaseModel):
             images = [img_file]
         else:
             images = img_file
+        transforms = self._build_transforms(transforms, "test")
+        self._check_arrange(transforms, "test")
         data = self.preprocess(images, transforms, self.model_type)
         self.net.eval()
         outputs = self.run(self.net, data, 'test')
@@ -597,7 +603,7 @@ class BaseChangeDetector(BaseModel):
             overlap (list[int] | tuple[int] | int, optional):
                 Overlap between two blocks. If `overlap` is a list or tuple, it should
                 be in (W, H) format. Defaults to 36.
-            transforms (paddlers.transforms.Compose|None, optional): Transforms for 
+            transforms (paddlers.transforms.Compose|list|None, optional): Transforms for 
                 inputs. If None, the transforms for evaluation process will be used. 
                 Defaults to None.
             invalid_value (int, optional): Value that marks invalid pixels in output 
@@ -618,7 +624,7 @@ class BaseChangeDetector(BaseModel):
                        eager_load, not quiet)
 
     def preprocess(self, images, transforms, to_tensor=True):
-        self._check_transforms(transforms, 'test')
+        self._check_transforms(transforms)
         batch_im1, batch_im2 = list(), list()
         batch_trans_info = list()
         for im1, im2 in images:
@@ -713,12 +719,12 @@ class BaseChangeDetector(BaseModel):
             score_maps.append(score_map.squeeze())
         return label_maps, score_maps
 
-    def _check_transforms(self, transforms, mode):
-        super()._check_transforms(transforms, mode)
-        if not isinstance(transforms.arrange,
-                          paddlers.transforms.ArrangeChangeDetector):
+    def _check_arrange(self, transforms, mode):
+        super()._check_arrange(transforms, mode)
+        if not isinstance(transforms.arrange, ArrangeChangeDetector):
             raise TypeError(
-                "`transforms.arrange` must be an ArrangeChangeDetector object.")
+                "`transforms.arrange` must be an `ArrangeChangeDetector` object."
+            )
 
     def set_losses(self, losses, weights=None):
         if weights is None:

paddlers/tasks/classifier.py

@@ -29,12 +29,16 @@ from paddlers.models import clas_losses
 from paddlers.models.ppcls.data.postprocess import build_postprocess
 from paddlers.utils.checkpoint import cls_pretrain_weights_dict
 from paddlers.transforms import Resize, decode_image, construct_sample
+from paddlers.transforms.operators import ArrangeClassifier
+
 from .base import BaseModel
 
 __all__ = ["ResNet50_vd", "MobileNetV3", "HRNet", "CondenseNetV2"]
 
 
 class BaseClassifier(BaseModel):
+    _arrange = ArrangeClassifier
+
     def __init__(self,
                  model_name,
                  in_channels=3,
@@ -380,7 +384,7 @@ class BaseClassifier(BaseModel):
                  "top5": acc of top5}.
         """
 
-        self._check_transforms(eval_dataset.transforms, 'eval')
+        self._check_transforms(eval_dataset.transforms)
 
         self.net.eval()
         nranks = paddle.distributed.get_world_size()
@@ -400,6 +404,8 @@ class BaseClassifier(BaseModel):
         if nranks < 2 or local_rank == 0:
             self.eval_data_loader = self.build_data_loader(
                 eval_dataset, batch_size=batch_size, mode='eval')
+            self._check_arrange(self.eval_data_loader.dataset.transforms,
+                                'eval')
             logging.info(
                 "Start to evaluate(total_samples={}, total_steps={})...".format(
                     eval_dataset.num_samples, eval_dataset.num_samples))
@@ -455,6 +461,8 @@ class BaseClassifier(BaseModel):
             images = [img_file]
         else:
             images = img_file
+        transforms = self._build_transforms(transforms, "test")
+        self._check_arrange(transforms, "test")
         data, _ = self.preprocess(images, transforms, self.model_type)
         self.net.eval()
 
@@ -480,7 +488,7 @@ class BaseClassifier(BaseModel):
         return prediction
 
     def preprocess(self, images, transforms, to_tensor=True):
-        self._check_transforms(transforms, 'test')
+        self._check_transforms(transforms)
         batch_im = list()
         for im in images:
             if isinstance(im, str):
@@ -496,12 +504,11 @@ class BaseClassifier(BaseModel):
 
         return batch_im, None
 
-    def _check_transforms(self, transforms, mode):
-        super()._check_transforms(transforms, mode)
-        if not isinstance(transforms.arrange,
-                          paddlers.transforms.ArrangeClassifier):
+    def _check_arrange(self, transforms, mode):
+        super()._check_arrange(transforms, mode)
+        if not isinstance(transforms.arrange, ArrangeClassifier):
             raise TypeError(
-                "`transforms.arrange` must be an ArrangeClassifier object.")
+                "`transforms.arrange` must be an `ArrangeClassifier` object.")
 
     def build_data_loader(self,
                           dataset,

paddlers/tasks/object_detector.py

@@ -26,6 +26,7 @@ import paddlers.models.ppdet as ppdet
 from paddlers.models.ppdet.modeling.proposal_generator.target_layer import BBoxAssigner, MaskAssigner
 from paddlers.transforms import decode_image, construct_sample
 from paddlers.transforms.operators import _NormalizeBox, _PadBox, _BboxXYXY2XYWH, Resize, Pad
+from paddlers.transforms.operators import ArrangeDetector
 from paddlers.transforms.batch_operators import BatchCompose, BatchRandomResize, BatchRandomResizeByShort, \
     _BatchPad, _Gt2YoloTarget
 from paddlers.models.ppdet.optimizer import ModelEMA
@@ -42,6 +43,8 @@ __all__ = [
 
 
 class BaseDetector(BaseModel):
+    _arrange = ArrangeDetector
+
     def __init__(self, model_name, num_classes=80, **params):
         self.init_params.update(locals())
         if 'with_net' in self.init_params:
@@ -530,8 +533,8 @@ class BaseDetector(BaseModel):
                 }
         eval_dataset.batch_transforms = self._compose_batch_transform(
             eval_dataset.transforms, mode='eval')
-        self._check_transforms(eval_dataset.transforms, 'eval')
 
+        self._check_transforms(eval_dataset.transforms)
         self.net.eval()
         nranks = paddle.distributed.get_world_size()
         local_rank = paddle.distributed.get_rank()
@@ -570,6 +573,8 @@ class BaseDetector(BaseModel):
                     coco_gt=copy.deepcopy(eval_dataset.coco_gt),
                     classwise=False)
             scores = collections.OrderedDict()
+
+            self._check_arrange(eval_dataset.transforms, 'eval')
             logging.info(
                 "Start to evaluate(total_samples={}, total_steps={})...".format(
                     eval_dataset.num_samples, eval_dataset.num_samples))
@@ -623,6 +628,8 @@ class BaseDetector(BaseModel):
         else:
             images = img_file
 
+        transforms = self._build_transforms(transforms, "test")
+        self._check_arrange(transforms, "test")
         batch_samples, _ = self.preprocess(images, transforms)
         self.net.eval()
         outputs = self.run(self.net, batch_samples, 'test')
@@ -633,7 +640,7 @@ class BaseDetector(BaseModel):
         return prediction
 
     def preprocess(self, images, transforms, to_tensor=True):
-        self._check_transforms(transforms, 'test')
+        self._check_transforms(transforms)
         batch_samples = list()
         for im in images:
             if isinstance(im, str):
@@ -718,12 +725,11 @@ class BaseDetector(BaseModel):
 
         return results
 
-    def _check_transforms(self, transforms, mode):
-        super()._check_transforms(transforms, mode)
-        if not isinstance(transforms.arrange,
-                          paddlers.transforms.ArrangeDetector):
+    def _check_arrange(self, transforms, mode):
+        super()._check_arrange(transforms, mode)
+        if not isinstance(transforms.arrange, ArrangeDetector):
             raise TypeError(
-                "`transforms.arrange` must be an ArrangeDetector object.")
+                "`transforms.arrange` must be an `ArrangeDetector` object.")
 
     def get_pruning_info(self):
         info = super().get_pruning_info()

paddlers/tasks/restorer.py

@@ -29,6 +29,7 @@ import paddlers.utils.logging as logging
 from paddlers.models import res_losses
 from paddlers.models.ppgan.modules.init import init_weights
 from paddlers.transforms import Resize, decode_image, construct_sample
+from paddlers.transforms.operators import ArrangeRestorer
 from paddlers.transforms.functions import calc_hr_shape
 from paddlers.utils.checkpoint import res_pretrain_weights_dict
 from .base import BaseModel
@@ -39,6 +40,7 @@ __all__ = ["DRN", "LESRCNN", "ESRGAN"]
 
 
 class BaseRestorer(BaseModel):
+    _arrange = ArrangeRestorer
     MIN_MAX = (0., 1.)
     TEST_OUT_KEY = None
 
@@ -390,7 +392,7 @@ class BaseRestorer(BaseModel):
 
         """
 
-        self._check_transforms(eval_dataset.transforms, 'eval')
+        self._check_transforms(eval_dataset.transforms)
 
         self.net.eval()
         nranks = paddle.distributed.get_world_size()
@@ -411,6 +413,7 @@ class BaseRestorer(BaseModel):
         if nranks < 2 or local_rank == 0:
             self.eval_data_loader = self.build_data_loader(
                 eval_dataset, batch_size=batch_size, mode='eval')
+            self._check_arrange(eval_dataset.transforms, 'eval')
             # XXX: Hard-code crop_border and test_y_channel
             psnr = metrics.PSNR(crop_border=4, test_y_channel=True)
             ssim = metrics.SSIM(crop_border=4, test_y_channel=True)
@@ -466,6 +469,8 @@ class BaseRestorer(BaseModel):
             images = [img_file]
         else:
             images = img_file
+        transforms = self._build_transforms(transforms, "test")
+        self._check_arrange(transforms, "test")
         data = self.preprocess(images, transforms, self.model_type)
         self.net.eval()
         outputs = self.run(self.net, data, 'test')
@@ -477,7 +482,7 @@ class BaseRestorer(BaseModel):
         return prediction
 
     def preprocess(self, images, transforms, to_tensor=True):
-        self._check_transforms(transforms, 'test')
+        self._check_transforms(transforms)
         batch_im = list()
         batch_trans_info = list()
         for im in images:
@@ -559,12 +564,11 @@ class BaseRestorer(BaseModel):
             res_maps.append(res_map.squeeze())
         return res_maps
 
-    def _check_transforms(self, transforms, mode):
-        super()._check_transforms(transforms, mode)
-        if not isinstance(transforms.arrange,
-                          paddlers.transforms.ArrangeRestorer):
+    def _check_arrange(self, transforms, mode):
+        super()._check_arrange(transforms, mode)
+        if not isinstance(transforms.arrange, ArrangeRestorer):
             raise TypeError(
-                "`transforms.arrange` must be an ArrangeRestorer object.")
+                "`transforms.arrange` must be an `ArrangeRestorer` object.")
 
     def build_data_loader(self,
                           dataset,

paddlers/tasks/segmenter.py

@@ -28,6 +28,7 @@ import paddlers.rs_models.seg as cmseg
 import paddlers.utils.logging as logging
 from paddlers.models import seg_losses
 from paddlers.transforms import Resize, decode_image, construct_sample
+from paddlers.transforms.operators import ArrangeSegmenter
 from paddlers.utils import get_single_card_bs, DisablePrint
 from paddlers.utils.checkpoint import seg_pretrain_weights_dict
 from .base import BaseModel
@@ -42,6 +43,8 @@ __all__ = [
 
 
 class BaseSegmenter(BaseModel):
+    _arrange = ArrangeSegmenter
+
     def __init__(self,
                  model_name,
                  num_classes=2,
@@ -406,7 +409,7 @@ class BaseSegmenter(BaseModel):
 
         """
 
-        self._check_transforms(eval_dataset.transforms, 'eval')
+        self._check_transforms(eval_dataset.transforms)
         self.net.eval()
         nranks = paddle.distributed.get_world_size()
         local_rank = paddle.distributed.get_rank()
@@ -426,6 +429,7 @@ class BaseSegmenter(BaseModel):
                 "is forcibly set to {}.".format(batch_size))
         self.eval_data_loader = self.build_data_loader(
             eval_dataset, batch_size=batch_size, mode='eval')
+        self._check_arrange(eval_dataset.transforms, 'eval')
 
         intersect_area_all = 0
         pred_area_all = 0
@@ -524,6 +528,8 @@ class BaseSegmenter(BaseModel):
             images = [img_file]
         else:
             images = img_file
+        transforms = self._build_transforms(transforms, "test")
+        self._check_arrange(transforms, "test")
         data = self.preprocess(images, transforms, self.model_type)
         self.net.eval()
         outputs = self.run(self.net, data, 'test')
@@ -585,7 +591,7 @@ class BaseSegmenter(BaseModel):
                        eager_load, not quiet)
 
     def preprocess(self, images, transforms, to_tensor=True):
-        self._check_transforms(transforms, 'test')
+        self._check_transforms(transforms)
         batch_im = list()
         batch_trans_info = list()
         for im in images:
@@ -675,12 +681,11 @@ class BaseSegmenter(BaseModel):
             score_maps.append(score_map.squeeze())
         return label_maps, score_maps
 
-    def _check_transforms(self, transforms, mode):
-        super()._check_transforms(transforms, mode)
-        if not isinstance(transforms.arrange,
-                          paddlers.transforms.ArrangeSegmenter):
+    def _check_arrange(self, transforms, mode):
+        super()._check_arrange(transforms, mode)
+        if not isinstance(transforms.arrange, ArrangeSegmenter):
             raise TypeError(
-                "`transforms.arrange` must be an ArrangeSegmenter object.")
+                "`transforms.arrange` must be an `ArrangeSegmenter` object.")
 
     def set_losses(self, losses, weights=None):
         if weights is None:

paddlers/tasks/utils/slider_predict.py

@@ -330,7 +330,7 @@ def slider_predict(predict_func,
         overlap (list[int] | tuple[int] | int):
             Overlap between two blocks. If `overlap` is list or tuple, it should
             be in (W, H) format.
-        transforms (paddlers.transforms.Compose|None): Transforms for inputs. If 
+        transforms (paddlers.transforms.Compose|list|None): Transforms for inputs. If 
             None, the transforms for evaluation process will be used. 
         invalid_value (int): Value that marks invalid pixels in output image. 
             Defaults to 255.

paddlers/transforms/indices.py

@@ -27,21 +27,24 @@ __all__ = [
 ]
 
 EPS = 1e-32
-BAND_NAMES = ["b", "g", "r", "re1", "re2", "re3", "n", "s1", "s2", "t1", "t2"]
-
-# | Band name | Description | Wavelength (μm) | Satellite |
-# |-----------|-------------|-----------------|-----------|
-# |     b     | Blue        |   0.450-0.515   | Landsat8  |
-# |     g     | Green       |   0.525-0.600   | Landsat8  |
-# |     r     | Red         |   0.630-0.680   | Landsat8  |
-# |    re1    | Red Edge 1  |   0.698-0.713   | Sentinel2 |
-# |    re2    | Red Edge 2  |   0.733-0.748   | Sentinel2 |
-# |    re3    | Red Edge 3  |   0.773-0.793   | Sentinel2 |
-# |     n     | NIR         |   0.845-0.885   | Landsat8  |
-# |    s1     | SWIR 1      |   1.560-1.660   | Landsat8  |
-# |    s2     | SWIR 2      |   2.100-2.300   | Landsat8  |
-# |    t1     | Thermal 1   |   10.60-11.19   | Landsat8  |
-# |    t2     | Thermal 2   |   11.50-12.51   | Landsat8  |
+BAND_NAMES = [
+    "b", "g", "r", "re1", "re2", "re3", "n", "s1", "s2", "t", "t1", "t2"
+]
+
+# | Band name |   Description    | Wavelength (μm) | Satellite |
+# |-----------|------------------|-----------------|-----------|
+# |     b     | Blue             |   0.450-0.515   | Landsat8  |
+# |     g     | Green            |   0.525-0.600   | Landsat8  |
+# |     r     | Red              |   0.630-0.680   | Landsat8  |
+# |    re1    | Red Edge 1       |   0.698-0.713   | Sentinel2 |
+# |    re2    | Red Edge 2       |   0.733-0.748   | Sentinel2 |
+# |    re3    | Red Edge 3       |   0.773-0.793   | Sentinel2 |
+# |     n     | NIR              |   0.845-0.885   | Landsat8  |
+# |    s1     | SWIR 1           |   1.560-1.660   | Landsat8  |
+# |    s2     | SWIR 2           |   2.100-2.300   | Landsat8  |
+# |     t     | Thermal Infrared |   10.40-12.50   | Landsat7  |
+# |    t1     | Thermal 1        |   10.60-11.19   | Landsat8  |
+# |    t2     | Thermal 2        |   11.50-12.51   | Landsat8  |
 
 
 class RSIndex(metaclass=abc.ABCMeta):
@@ -189,8 +192,8 @@ class CSI(RSIndex):
 
 
 class CSIT(RSIndex):
-    def _compute(self, n, s2, t1):
-        return n / ((s2 * t1) / 10000.0 + EPS)
+    def _compute(self, n, s2, t):
+        return n / ((s2 * t) / 10000.0 + EPS)
 
 
 class DBI(RSIndex):
@@ -213,9 +216,9 @@ class DVI(RSIndex):
 
 
 class EBBI(RSIndex):
-    def _compute(self, n, s1, t1):
+    def _compute(self, n, s1, t):
         num = s1 - n
-        denom = (10.0 * ((s1 + t1)**0.5))
+        denom = (10.0 * ((s1 + t)**0.5))
         return num / (denom + EPS)
 
 
@@ -334,8 +337,8 @@ class MSI(RSIndex):
 
 
 class NBLI(RSIndex):
-    def _compute(self, r, t1):
-        return compute_normalized_difference_index(r, t1)
+    def _compute(self, r, t):
+        return compute_normalized_difference_index(r, t)
 
 
 class NDSI(RSIndex):

paddlers/transforms/operators.py

@@ -61,11 +61,6 @@ __all__ = [
     "ReloadMask",
     "AppendIndex",
     "MatchRadiance",
-    "ArrangeRestorer",
-    "ArrangeSegmenter",
-    "ArrangeChangeDetector",
-    "ArrangeClassifier",
-    "ArrangeDetector",
 ]
 
 interp_dict = {
@@ -115,7 +110,8 @@ class Compose(object):
                 "Length of transforms must not be less than 1, but received is {}."
                 .format(len(transforms)))
         transforms = copy.deepcopy(transforms)
-        self.arrange = self._pick_arrange(transforms)
+        # We will have to do a late binding of `self.arrange`
+        self.arrange = None
         self.transforms = transforms
 
     def __call__(self, sample):
@@ -143,17 +139,6 @@ class Compose(object):
             sample = self.arrange(sample)
         return sample
 
-    def _pick_arrange(self, transforms):
-        arrange = None
-        for idx, op in enumerate(transforms):
-            if isinstance(op, Arrange):
-                if idx != len(transforms) - 1:
-                    raise ValueError(
-                        "Arrange operator must be placed at the end of the list."
-                    )
-                arrange = transforms.pop(idx)
-        return arrange
-
 
 class Transform(object):
     """

requirements.txt

@@ -24,5 +24,5 @@ scikit-learn
 scikit-image >= 0.14.0
 scipy
 shapely
-spyndex
-visualdl >= 2.1.1
+spyndex >= 0.4.0
+visualdl >= 2.1.1

tests/deploy/test_predictor.py

@@ -126,10 +126,7 @@ class TestCDPredictor(TestPredictor):
         t2_path = "data/ssmt/optical_t2.bmp"
         single_input = (t1_path, t2_path)
         num_inputs = 2
-        transforms = pdrs.transforms.Compose([
-            pdrs.transforms.DecodeImg(), pdrs.transforms.Normalize(),
-            pdrs.transforms.ArrangeChangeDetector('test')
-        ])
+        transforms = [pdrs.transforms.Normalize()]
 
         # Expected failure
         with self.assertRaises(ValueError):
@@ -194,10 +191,7 @@ class TestClasPredictor(TestPredictor):
     def check_predictor(self, predictor, trainer):
         single_input = "data/ssst/optical.bmp"
         num_inputs = 2
-        transforms = pdrs.transforms.Compose([
-            pdrs.transforms.DecodeImg(), pdrs.transforms.Normalize(),
-            pdrs.transforms.ArrangeClassifier('test')
-        ])
+        transforms = [pdrs.transforms.Normalize()]
         labels = list(range(2))
         trainer.labels = labels
         predictor._model.labels = labels
@@ -263,10 +257,7 @@ class TestDetPredictor(TestPredictor):
         # given that the network is (partially?) randomly initialized.
         single_input = "data/ssst/optical.bmp"
         num_inputs = 2
-        transforms = pdrs.transforms.Compose([
-            pdrs.transforms.DecodeImg(), pdrs.transforms.Normalize(),
-            pdrs.transforms.ArrangeDetector('test')
-        ])
+        transforms = [pdrs.transforms.Normalize()]
         labels = list(range(80))
         trainer.labels = labels
         predictor._model.labels = labels
@@ -328,10 +319,7 @@ class TestResPredictor(TestPredictor):
         # because the output is of uint8 type.
         single_input = "data/ssst/optical.bmp"
         num_inputs = 2
-        transforms = pdrs.transforms.Compose([
-            pdrs.transforms.DecodeImg(), pdrs.transforms.Normalize(),
-            pdrs.transforms.ArrangeRestorer('test')
-        ])
+        transforms = [pdrs.transforms.Normalize()]
 
         # Single input (file path)
         input_ = single_input
@@ -383,10 +371,7 @@ class TestSegPredictor(TestPredictor):
     def check_predictor(self, predictor, trainer):
         single_input = "data/ssst/optical.bmp"
         num_inputs = 2
-        transforms = pdrs.transforms.Compose([
-            pdrs.transforms.DecodeImg(), pdrs.transforms.Normalize(),
-            pdrs.transforms.ArrangeSegmenter('test')
-        ])
+        transforms = [pdrs.transforms.Normalize()]
 
         # Single input (file path)
         input_ = single_input

tests/tasks/test_slider_predict.py

@@ -232,10 +232,7 @@ class _TestSliderPredictNamespace:
 class TestSegSliderPredict(_TestSliderPredictNamespace.TestSliderPredict):
     def setUp(self):
         self.model = pdrs.tasks.seg.UNet(in_channels=10)
-        self.transforms = T.Compose([
-            T.DecodeImg(), T.Normalize([0.5] * 10, [0.5] * 10),
-            T.ArrangeSegmenter('test')
-        ])
+        self.transforms = [T.Normalize([0.5] * 10, [0.5] * 10)]
         self.image_path = "data/ssst/multispectral.tif"
         self.ref_path = self.image_path
         self.basename = osp.basename(self.ref_path)
@@ -244,10 +241,7 @@ class TestSegSliderPredict(_TestSliderPredictNamespace.TestSliderPredict):
 class TestCDSliderPredict(_TestSliderPredictNamespace.TestSliderPredict):
     def setUp(self):
         self.model = pdrs.tasks.cd.BIT(in_channels=10)
-        self.transforms = T.Compose([
-            T.DecodeImg(), T.Normalize([0.5] * 10, [0.5] * 10),
-            T.ArrangeChangeDetector('test')
-        ])
+        self.transforms = [T.Normalize([0.5] * 10, [0.5] * 10)]
         self.image_path = ("data/ssmt/multispectral_t1.tif",
                            "data/ssmt/multispectral_t2.tif")
         self.ref_path = self.image_path[0]

tests/transforms/test_indices.py

@@ -30,6 +30,7 @@ NAME_MAPPING = {
     'n': 'N',
     's1': 'S1',
     's2': 'S2',
+    't': 'T',
     't1': 'T1',
     't2': 'T2'
 }

tutorials/train/change_detection/bit.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/cdnet.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/changeformer.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/transforms.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/dsamnet.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/dsifn.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/fc_ef.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/fc_siam_conc.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/fc_siam_diff.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/fccdn.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/transforms.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/p2v.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/snunet.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/change_detection/stanet.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪
     T.RandomCrop(
         # 裁剪区域将被缩放到256x256
@@ -37,18 +35,15 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeChangeDetector('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeChangeDetector('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.CDDataset(

tutorials/train/classification/condensenetv2.py

@@ -24,9 +24,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到256x256大小
     T.Resize(target_size=256),
     # 以50%的概率实施随机水平翻转
@@ -35,18 +33,15 @@ train_transforms = T.Compose([
     T.RandomVerticalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeClassifier('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=256),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeClassifier('eval')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.ClasDataset(

tutorials/train/classification/hrnet.py

@@ -24,9 +24,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到256x256大小
     T.Resize(target_size=256),
     # 以50%的概率实施随机水平翻转
@@ -35,18 +33,15 @@ train_transforms = T.Compose([
     T.RandomVerticalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeClassifier('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=256),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeClassifier('eval')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.ClasDataset(

tutorials/train/classification/mobilenetv3.py

@@ -24,9 +24,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到256x256大小
     T.Resize(target_size=256),
     # 以50%的概率实施随机水平翻转
@@ -35,18 +33,15 @@ train_transforms = T.Compose([
     T.RandomVerticalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeClassifier('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=256),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeClassifier('eval')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.ClasDataset(

tutorials/train/classification/resnet50_vd.py

@@ -24,9 +24,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到256x256大小
     T.Resize(target_size=256),
     # 以50%的概率实施随机水平翻转
@@ -35,18 +33,15 @@ train_transforms = T.Compose([
     T.RandomVerticalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeClassifier('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=256),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeClassifier('eval')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.ClasDataset(

tutorials/train/image_restoration/drn.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 从输入影像中裁剪96x96大小的影像块
     T.RandomCrop(crop_size=96),
     # 以50%的概率实施随机水平翻转
@@ -33,19 +31,16 @@ train_transforms = T.Compose([
     T.RandomVerticalFlip(prob=0.5),
     # 将数据归一化到[0,1]
     T.Normalize(
-        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0]),
-    T.ArrangeRestorer('train')
-])
+        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 将输入影像缩放到256x256大小
     T.Resize(target_size=256),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
-        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0]),
-    T.ArrangeRestorer('eval')
-])
+        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.ResDataset(

tutorials/train/image_restoration/esrgan.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 从输入影像中裁剪32x32大小的影像块
     T.RandomCrop(crop_size=32),
     # 以50%的概率实施随机水平翻转
@@ -33,19 +31,16 @@ train_transforms = T.Compose([
     T.RandomVerticalFlip(prob=0.5),
     # 将数据归一化到[0,1]
     T.Normalize(
-        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0]),
-    T.ArrangeRestorer('train')
-])
+        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 将输入影像缩放到256x256大小
     T.Resize(target_size=256),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
-        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0]),
-    T.ArrangeRestorer('eval')
-])
+        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.ResDataset(

tutorials/train/image_restoration/lesrcnn.py

@@ -22,9 +22,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 从输入影像中裁剪32x32大小的影像块
     T.RandomCrop(crop_size=32),
     # 以50%的概率实施随机水平翻转
@@ -33,19 +31,16 @@ train_transforms = T.Compose([
     T.RandomVerticalFlip(prob=0.5),
     # 将数据归一化到[0,1]
     T.Normalize(
-        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0]),
-    T.ArrangeRestorer('train')
-])
+        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 将输入影像缩放到256x256大小
     T.Resize(target_size=256),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
-        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0]),
-    T.ArrangeRestorer('eval')
-])
+        mean=[0.0, 0.0, 0.0], std=[1.0, 1.0, 1.0])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.ResDataset(

tutorials/train/object_detection/faster_rcnn.py

@@ -26,9 +26,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪，裁块大小在一定范围内变动
     T.RandomCrop(),
     # 随机水平翻转
@@ -38,20 +36,17 @@ train_transforms = T.Compose([
         target_sizes=[512, 544, 576, 608], interp='RANDOM'),
     # 影像归一化
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('train')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 使用双三次插值将输入影像缩放到固定大小
     T.Resize(
         target_size=608, interp='CUBIC'),
     # 验证阶段与训练阶段的归一化方式必须相同
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('eval')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.VOCDetDataset(

tutorials/train/object_detection/ppyolo.py

@@ -26,9 +26,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪，裁块大小在一定范围内变动
     T.RandomCrop(),
     # 随机水平翻转
@@ -38,20 +36,17 @@ train_transforms = T.Compose([
         target_sizes=[512, 544, 576, 608], interp='RANDOM'),
     # 影像归一化
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('train')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 使用双三次插值将输入影像缩放到固定大小
     T.Resize(
         target_size=608, interp='CUBIC'),
     # 验证阶段与训练阶段的归一化方式必须相同
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('eval')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.VOCDetDataset(

tutorials/train/object_detection/ppyolo_tiny.py

@@ -26,9 +26,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪，裁块大小在一定范围内变动
     T.RandomCrop(),
     # 随机水平翻转
@@ -38,20 +36,17 @@ train_transforms = T.Compose([
         target_sizes=[512, 544, 576, 608], interp='RANDOM'),
     # 影像归一化
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('train')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 使用双三次插值将输入影像缩放到固定大小
     T.Resize(
         target_size=608, interp='CUBIC'),
     # 验证阶段与训练阶段的归一化方式必须相同
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('eval')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.VOCDetDataset(

tutorials/train/object_detection/ppyolov2.py

@@ -26,9 +26,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪，裁块大小在一定范围内变动
     T.RandomCrop(),
     # 随机水平翻转
@@ -38,20 +36,17 @@ train_transforms = T.Compose([
         target_sizes=[512, 544, 576, 608], interp='RANDOM'),
     # 影像归一化
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('train')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 使用双三次插值将输入影像缩放到固定大小
     T.Resize(
         target_size=608, interp='CUBIC'),
     # 验证阶段与训练阶段的归一化方式必须相同
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('eval')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.VOCDetDataset(

tutorials/train/object_detection/yolov3.py

@@ -26,9 +26,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 随机裁剪，裁块大小在一定范围内变动
     T.RandomCrop(),
     # 随机水平翻转
@@ -38,20 +36,17 @@ train_transforms = T.Compose([
         target_sizes=[512, 544, 576, 608], interp='RANDOM'),
     # 影像归一化
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('train')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 使用双三次插值将输入影像缩放到固定大小
     T.Resize(
         target_size=608, interp='CUBIC'),
     # 验证阶段与训练阶段的归一化方式必须相同
     T.Normalize(
-        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
-    T.ArrangeDetector('eval')
-])
+        mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.VOCDetDataset(

tutorials/train/semantic_segmentation/bisenetv2.py

@@ -27,28 +27,23 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到512x512大小
     T.Resize(target_size=512),
     # 以50%的概率实施随机水平翻转
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ArrangeSegmenter('train')
-])
+        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS)
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=512),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ReloadMask(),
-    T.ArrangeSegmenter('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.SegDataset(

tutorials/train/semantic_segmentation/deeplabv3p.py

@@ -27,28 +27,23 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到512x512大小
     T.Resize(target_size=512),
     # 以50%的概率实施随机水平翻转
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ArrangeSegmenter('train')
-])
+        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS)
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=512),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ReloadMask(),
-    T.ArrangeSegmenter('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.SegDataset(

tutorials/train/semantic_segmentation/factseg.py

@@ -24,9 +24,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 选择前三个波段
     T.SelectBand([1, 2, 3]),
     # 将影像缩放到512x512大小
@@ -35,21 +33,18 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeSegmenter('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段应当选择相同的波段
     T.SelectBand([1, 2, 3]),
     T.Resize(target_size=512),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeSegmenter('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.SegDataset(

tutorials/train/semantic_segmentation/farseg.py

@@ -24,9 +24,7 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 选择前三个波段
     T.SelectBand([1, 2, 3]),
     # 将影像缩放到512x512大小
@@ -35,21 +33,18 @@ train_transforms = T.Compose([
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ArrangeSegmenter('train')
-])
+        mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     # 验证阶段与训练阶段应当选择相同的波段
     T.SelectBand([1, 2, 3]),
     T.Resize(target_size=512),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
-    T.ReloadMask(),
-    T.ArrangeSegmenter('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.SegDataset(

tutorials/train/semantic_segmentation/fast_scnn.py

@@ -27,28 +27,23 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到512x512大小
     T.Resize(target_size=512),
     # 以50%的概率实施随机水平翻转
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ArrangeSegmenter('train')
-])
+        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS)
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=512),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ReloadMask(),
-    T.ArrangeSegmenter('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.SegDataset(

tutorials/train/semantic_segmentation/hrnet.py

@@ -27,28 +27,23 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到512x512大小
     T.Resize(target_size=512),
     # 以50%的概率实施随机水平翻转
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ArrangeSegmenter('train')
-])
+        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS)
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=512),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ReloadMask(),
-    T.ArrangeSegmenter('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.SegDataset(

tutorials/train/semantic_segmentation/unet.py

@@ -27,28 +27,23 @@ pdrs.utils.download_and_decompress(
 # 定义训练和验证时使用的数据变换（数据增强、预处理等）
 # 使用Compose组合多种变换方式。Compose中包含的变换将按顺序串行执行
 # API说明：https://github.com/PaddlePaddle/PaddleRS/blob/develop/docs/apis/data.md
-train_transforms = T.Compose([
-    # 读取影像
-    T.DecodeImg(),
+train_transforms = [
     # 将影像缩放到512x512大小
     T.Resize(target_size=512),
     # 以50%的概率实施随机水平翻转
     T.RandomHorizontalFlip(prob=0.5),
     # 将数据归一化到[-1,1]
     T.Normalize(
-        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ArrangeSegmenter('train')
-])
+        mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS)
+]
 
-eval_transforms = T.Compose([
-    T.DecodeImg(),
+eval_transforms = [
     T.Resize(target_size=512),
     # 验证阶段与训练阶段的数据归一化方式必须相同
     T.Normalize(
         mean=[0.5] * NUM_BANDS, std=[0.5] * NUM_BANDS),
-    T.ReloadMask(),
-    T.ArrangeSegmenter('eval')
-])
+    T.ReloadMask()
+]
 
 # 分别构建训练和验证所用的数据集
 train_dataset = pdrs.datasets.SegDataset(