Load and save video-processor from separate folder (#33562)

* load and save from video-processor folder

* Update src/transformers/models/llava_onevision/processing_llava_onevision.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

---------

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

src/transformers/models/llava_onevision/image_processing_llava_onevision.py

@@ -621,6 +621,7 @@ class LlavaOnevisionImageProcessor(BaseImageProcessor):
         """
         do_resize = do_resize if do_resize is not None else self.do_resize
         size = size if size is not None else self.size
+        size = get_size_dict(size, default_to_square=False)
         image_grid_pinpoints = image_grid_pinpoints if image_grid_pinpoints is not None else self.image_grid_pinpoints
         resample = resample if resample is not None else self.resample
         do_rescale = do_rescale if do_rescale is not None else self.do_rescale

src/transformers/models/llava_onevision/processing_llava_onevision.py

@@ -17,6 +17,7 @@ Processor class for LLaVa-Onevision.
 """
 
 import math
+import os
 import sys
 from typing import Iterable, List, Union
 
@@ -34,6 +35,11 @@ from ...processing_utils import (
     ProcessorMixin,
 )
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
+from ...utils import logging
+from ..auto import AutoImageProcessor
+
+
+logger = logging.get_logger(__name__)
 
 
 class LlavaOnevisionProcessorKwargs(ProcessingKwargs, total=False):
@@ -96,7 +102,7 @@ class LlavaOnevisionProcessor(ProcessorMixin):
         chat_template=None,
         image_token="<image>",
         video_token="<video>",
-        **kwargs: Unpack[LlavaOnevisionProcessorKwargs],
+        **kwargs,
     ):
         self.num_image_tokens = num_image_tokens
         self.vision_feature_select_strategy = vision_feature_select_strategy
@@ -109,7 +115,7 @@ class LlavaOnevisionProcessor(ProcessorMixin):
         images: ImageInput = None,
         text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
         videos: VideoInput = None,
-        **kwargs,
+        **kwargs: Unpack[LlavaOnevisionProcessorKwargs],
     ) -> BatchFeature:
         """
         Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
@@ -272,3 +278,46 @@ class LlavaOnevisionProcessor(ProcessorMixin):
         tokenizer_input_names = self.tokenizer.model_input_names
         image_processor_input_names = self.image_processor.model_input_names
         return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+    # override to save video-config in a separate config file
+    def save_pretrained(self, save_directory, **kwargs):
+        if os.path.isfile(save_directory):
+            raise ValueError(f"Provided path ({save_directory}) should be a directory, not a file")
+        os.makedirs(save_directory, exist_ok=True)
+        video_processor_path = os.path.join(save_directory, "video_processor")
+        self.video_processor.save_pretrained(video_processor_path)
+
+        video_processor_present = "video_processor" in self.attributes
+        if video_processor_present:
+            self.attributes.remove("video_processor")
+
+        outputs = super().save_pretrained(save_directory, **kwargs)
+
+        if video_processor_present:
+            self.attributes += ["video_processor"]
+        return outputs
+
+    # override to load video-config from a separate config file
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path, **kwargs):
+        processor = super().from_pretrained(pretrained_model_name_or_path, **kwargs)
+
+        # if return_unused_kwargs a tuple is returned where the second element is 'unused_kwargs'
+        if isinstance(processor, tuple):
+            processor = processor[0]
+
+        try:
+            video_processor = AutoImageProcessor.from_pretrained(
+                pretrained_model_name_or_path, subfolder="video_processor"
+            )
+            processor.video_processor = video_processor
+        except EnvironmentError:
+            # this means users are using prev version of saved processor where we had only one preprocessor_config.json
+            # for loading back that should work and load a LlavaOnevisionVideoProcessor class
+            logger.info(
+                "You are loading `LlavaOnevisionProcessor` but the indicated `path` doesn't contain a folder called "
+                "`video_processor`. It is strongly recommended to load and save the processor again so the video processor is saved "
+                "in a separate config."
+            )
+
+        return processor

tests/models/llava_onevision/test_processing_llava_onevision.py

@@ -58,15 +58,16 @@ class LlavaOnevisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).video_processor
 
     def prepare_processor_dict(self):
-        return {"chat_template": "dummy_template"}
+        return {"chat_template": "dummy_template", "num_image_tokens": 6, "vision_feature_select_strategy": "default"}
 
-    @unittest.skip(
-        "Skip because the model has no processor kwargs except for chat template and"
-        "chat template is saved as a separate file. Stop skipping this test when the processor"
-        "has new kwargs saved in config file."
-    )
     def test_processor_to_json_string(self):
-        pass
+        processor = self.get_processor()
+        obj = json.loads(processor.to_json_string())
+        for key, value in self.prepare_processor_dict().items():
+            # chat_tempalate are tested as a separate test because they are saved in separate files
+            if key != "chat_template":
+                self.assertEqual(obj[key], value)
+                self.assertEqual(getattr(processor, key, None), value)
 
     # Copied from tests.models.llava.test_processor_llava.LlavaProcessorTest.test_chat_template_is_saved
     def test_chat_template_is_saved(self):
@@ -191,7 +192,7 @@ class LlavaOnevisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):
             max_length=76,
         )
         self.assertEqual(inputs["pixel_values"].shape[3], 214)
-        self.assertEqual(len(inputs["input_ids"][0]), 5)
+        self.assertEqual(len(inputs["input_ids"][0]), 4)
 
     @require_torch
     @require_vision
@@ -282,7 +283,7 @@ class LlavaOnevisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         image_input = self.prepare_image_inputs()
 
         inputs = processor(text=input_str, images=image_input, return_tensors="pt", max_length=112)
-        self.assertEqual(len(inputs["input_ids"][0]), 112)
+        self.assertEqual(len(inputs["input_ids"][0]), 2)
 
     @require_vision
     @require_torch
@@ -299,4 +300,4 @@ class LlavaOnevisionProcessorTest(ProcessorTesterMixin, unittest.TestCase):
         image_input = self.prepare_image_inputs()
 
         inputs = processor(text=input_str, images=image_input, return_tensors="pt")
-        self.assertEqual(len(inputs["input_ids"][0]), 117)
+        self.assertEqual(len(inputs["input_ids"][0]), 2)

tests/test_processing_common.py

@@ -112,6 +112,14 @@ class ProcessorTesterMixin:
 
                 self.assertEqual(processor_second.to_dict(), processor_first.to_dict())
 
+                for attribute in processor_first.attributes:
+                    attribute_first = getattr(processor_first, attribute)
+                    attribute_second = getattr(processor_second, attribute)
+
+                    # tokenizer repr contains model-path from where we loaded
+                    if "tokenizer" not in attribute:
+                        self.assertEqual(repr(attribute_first), repr(attribute_second))
+
     # These kwargs-related tests ensure that processors are correctly instantiated.
     # they need to be applied only if an image_processor exists.