36978 | Fast image processor for DPT model (#37481)

* chore: ran codegen script

* test: test_image_processor_properties

* test: test_image_processor_from_dict_with_kwargs

* test: wip - test_padding

* test: test_padding

* test: test_keep_aspect_ratio

* wip

* test

* test: wip

* test: wip

* test: test_call_segmentation_maps, wip

* chore: tidy up

* test: test_call_segmentation_maps

* fix: test_save_load_fast_slow

* test: reduce labels

* chore: make fixup

* chore: rm comment

* chore: tidy

* chore remove comment

* refactor: no need to infer channel dimesnion

* refactor: encapsulate logic for preparing segmentation maps

* refactor: improve readability of segmentation_map preparation

* improvement: batched version of pad_image

* chore: fixup

* docs

* chore: make quality

* chore: remove unecessary comment

* fix: add SemanticSegmentationMixin

* feat: add post_process_depth_estimation to fast dpt image processor

* chore: fix formatting

* remove max_height, max_width

* fix: better way of processin segmentation maps
- copied from Beit Fast processor

* chore: formatting + remove TODO

* chore: fixup styles

* chore: remove unecessary line break

* chore: core review suggestion to remove autodocstring

* fix: add do_reduce_labels logic + refactor
- refactor preprocess logic to make it consistent with other processors
- add missing reduce labels logic

* refactor: remove deprecated mixin

* chore: fixup

* use modular for dpt + final nit changes

* fix style

---------

Co-authored-by: Samuel Rae <[email protected]>
Co-authored-by: yonigozlan <[email protected]>

Author: 3 people

docs/source/en/model_doc/dpt.md

@@ -78,7 +78,13 @@ If you're interested in submitting a resource to be included here, please feel f
 
 [[autodoc]] DPTImageProcessor
     - preprocess
+
+## DPTImageProcessorFast
+
+[[autodoc]] DPTImageProcessorFast
+    - preprocess
     - post_process_semantic_segmentation
+    - post_process_depth_estimation
 
 ## DPTModel
 

src/transformers/models/auto/image_processing_auto.py

@@ -74,14 +74,14 @@
             ("data2vec-vision", ("BeitImageProcessor", "BeitImageProcessorFast")),
             ("deformable_detr", ("DeformableDetrImageProcessor", "DeformableDetrImageProcessorFast")),
             ("deit", ("DeiTImageProcessor", "DeiTImageProcessorFast")),
-            ("depth_anything", ("DPTImageProcessor",)),
+            ("depth_anything", ("DPTImageProcessor", "DPTImageProcessorFast")),
             ("depth_pro", ("DepthProImageProcessor", "DepthProImageProcessorFast")),
             ("deta", ("DetaImageProcessor",)),
             ("detr", ("DetrImageProcessor", "DetrImageProcessorFast")),
             ("dinat", ("ViTImageProcessor", "ViTImageProcessorFast")),
             ("dinov2", ("BitImageProcessor", "BitImageProcessorFast")),
             ("donut-swin", ("DonutImageProcessor", "DonutImageProcessorFast")),
-            ("dpt", ("DPTImageProcessor",)),
+            ("dpt", ("DPTImageProcessor", "DPTImageProcessorFast")),
             ("efficientformer", ("EfficientFormerImageProcessor",)),
             ("efficientnet", ("EfficientNetImageProcessor", "EfficientNetImageProcessorFast")),
             ("flava", ("FlavaImageProcessor", "FlavaImageProcessorFast")),

src/transformers/models/beit/image_processing_beit_fast.py

@@ -174,11 +174,6 @@ def _preprocess_segmentation_maps(
         processed_segmentation_maps = processed_segmentation_maps.to(torch.int64)
         return processed_segmentation_maps
 
-    def __call__(self, images, segmentation_maps=None, **kwargs):
-        # Overrides the `__call__` method of the `Preprocessor` class such that the images and segmentation maps can both
-        # be passed in as positional arguments.
-        return super().__call__(images, segmentation_maps=segmentation_maps, **kwargs)
-
     @auto_docstring
     def preprocess(
         self,

src/transformers/models/dpt/__init__.py

@@ -21,6 +21,7 @@
     from .configuration_dpt import *
     from .feature_extraction_dpt import *
     from .image_processing_dpt import *
+    from .image_processing_dpt_fast import *
     from .modeling_dpt import *
 else:
     import sys