Add image text to text pipeline

docs/source/en/main_classes/pipelines.md

@@ -478,6 +478,12 @@ Pipelines available for multimodal tasks include the following.
     - __call__
     - all
 
+### ImageTextToTextPipeline
+
+[[autodoc]] ImageTextToTextPipeline
+    - __call__
+    - all
+
 ### MaskGenerationPipeline
 
 [[autodoc]] MaskGenerationPipeline

docs/source/ja/main_classes/pipelines.md

@@ -481,6 +481,12 @@ my_pipeline = pipeline(model="xxxx", pipeline_class=MyPipeline)
     - __call__
     - all
 
+### ImageTextToTextPipeline
+
+[[autodoc]] ImageTextToTextPipeline
+    - __call__
+    - all
+
 ### VisualQuestionAnsweringPipeline
 
 [[autodoc]] VisualQuestionAnsweringPipeline

docs/source/zh/main_classes/pipelines.md

@@ -455,6 +455,12 @@ See [`TokenClassificationPipeline`] for all details.
     - __call__
     - all
 
+### ImageTextToTextPipeline
+
+[[autodoc]] ImageTextToTextPipeline
+    - __call__
+    - all
+
 ### MaskGenerationPipeline
 
 [[autodoc]] MaskGenerationPipeline

src/transformers/__init__.py

@@ -868,6 +868,7 @@
         "ImageClassificationPipeline",
         "ImageFeatureExtractionPipeline",
         "ImageSegmentationPipeline",
+        "ImageTextToTextPipeline",
         "ImageToImagePipeline",
         "ImageToTextPipeline",
         "JsonPipelineDataFormat",
@@ -5794,6 +5795,7 @@
         ImageClassificationPipeline,
         ImageFeatureExtractionPipeline,
         ImageSegmentationPipeline,
+        ImageTextToTextPipeline,
         ImageToImagePipeline,
         ImageToTextPipeline,
         JsonPipelineDataFormat,

src/transformers/image_utils.py

@@ -385,6 +385,27 @@ def load_image(image: Union[str, "PIL.Image.Image"], timeout: Optional[float] =
     return image
 
 
+def load_images(
+    images: Union[List, Tuple, str, "PIL.Image.Image"], timeout: Optional[float] = None
+) -> Union["PIL.Image.Image", List["PIL.Image.Image"], List[List["PIL.Image.Image"]]]:
+    """Loads images, handling different levels of nesting.
+
+    Args:
+      images: A single image, a list of images, or a list of lists of images to load.
+      timeout: Timeout for loading images.
+
+    Returns:
+      A single image, a list of images, a list of lists of images.
+    """
+    if isinstance(images, (list, tuple)):
+        if len(images) and isinstance(images[0], (list, tuple)):
+            return [[load_image(image, timeout=timeout) for image in image_group] for image_group in images]
+        else:
+            return [load_image(image, timeout=timeout) for image in images]
+    else:
+        return load_image(images, timeout=timeout)
+
+
 def validate_preprocess_arguments(
     do_rescale: Optional[bool] = None,
     rescale_factor: Optional[float] = None,

src/transformers/models/auto/image_processing_auto.py

@@ -114,6 +114,7 @@
             ("oneformer", ("OneFormerImageProcessor",)),
             ("owlv2", ("Owlv2ImageProcessor",)),
             ("owlvit", ("OwlViTImageProcessor",)),
+            ("paligemma", ("SiglipImageProcessor",)),
             ("perceiver", ("PerceiverImageProcessor",)),
             ("pix2struct", ("Pix2StructImageProcessor",)),
             ("pixtral", ("PixtralImageProcessor",)),

src/transformers/models/donut/processing_donut.py

@@ -16,6 +16,7 @@
 Processor class for Donut.
 """
 
+import logging
 import re
 import warnings
 from contextlib import contextmanager
@@ -24,12 +25,16 @@
 from ...image_utils import ImageInput
 from ...processing_utils import ProcessingKwargs, ProcessorMixin, Unpack
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
+from ...utils.deprecation import deprecate_kwarg
 
 
 class DonutProcessorKwargs(ProcessingKwargs, total=False):
     _defaults = {}
 
 
+logger = logging.getLogger(__name__)
+
+
 class DonutProcessor(ProcessorMixin):
     r"""
     Constructs a Donut processor which wraps a Donut image processor and an XLMRoBERTa tokenizer into a single
@@ -70,6 +75,7 @@ def __init__(self, image_processor=None, tokenizer=None, **kwargs):
         self.current_processor = self.image_processor
         self._in_target_context_manager = False
 
+    @deprecate_kwarg(old_name="legacy", version="5.0.0")
     def __call__(
         self,
         images: ImageInput = None,
@@ -85,6 +91,15 @@ def __call__(
         [`~DonutTokenizer.__call__`]. Please refer to the doctsring of the above two methods for more information.
         """
         # For backward compatibility
+        legacy = kwargs.pop("legacy", True)
+        if legacy:
+            # With `add_special_tokens=True`, the performance of donut are degraded when working with both images and text.
+            logger.warning_once(
+                "Legacy behavior is being used. The new behavior with legacy=False will be enabled in the future."
+                "In the new behavior, if both images and text are provided, the default value of `add_special_tokens` "
+                "will be changed to `False` when calling the tokenizer if `add_special_tokens` is unset."
+            )
+
         if self._in_target_context_manager:
             return self.current_processor(images, text, **kwargs)
 
@@ -100,6 +115,8 @@ def __call__(
         if images is not None:
             inputs = self.image_processor(images, **output_kwargs["images_kwargs"])
         if text is not None:
+            if not legacy and images is not None:
+                output_kwargs["text_kwargs"].setdefault("add_special_tokens", False)
             encodings = self.tokenizer(text, **output_kwargs["text_kwargs"])
 
         if text is None:

src/transformers/models/fuyu/image_processing_fuyu.py

@@ -19,7 +19,7 @@
 
 import numpy as np
 
-from ...image_processing_utils import BaseImageProcessor, BatchFeature
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
 from ...image_transforms import (
     pad,
     resize,
@@ -475,6 +475,7 @@ def preprocess(
             input_data_format = infer_channel_dimension_format(batch_images[0][0])
 
         original_image_sizes = [get_image_size(images[0], channel_dim=input_data_format) for images in batch_images]
+        size = get_size_dict(size)  # for BC
 
         if do_resize:
             batch_images = [

src/transformers/models/fuyu/processing_fuyu.py

@@ -264,10 +264,10 @@ def _tokenize_prompts_with_image_and_batch(
         bos_token = tokenizer.vocab["|ENDOFTEXT|"]
     prompts_tokens = [[[bos_token] + x for x in prompt_seq] for prompt_seq in prompts_tokens]
     if add_beginning_of_answer_token:
-        boa = tokenizer.vocab[BEGINNING_OF_ANSWER_STRING]
+        beginning_of_answer = tokenizer.vocab[BEGINNING_OF_ANSWER_STRING]
         # Only add bbox open token to the last subsequence since that is what will be completed
         for token_seq in prompts_tokens:
-            token_seq[-1].append(boa)
+            token_seq[-1].append(beginning_of_answer)
 
     # Now we have a list of list of tokens which each list has a different
     # size. We want to extend this list to:
@@ -682,6 +682,32 @@ def tokens_to_points(tokens, original_size):
 
         return results
 
+    def post_process_image_text_to_text(self, generated_outputs):
+        """
+        Post-processes the output of `FuyuForConditionalGeneration` to only return the text output.
+
+        Args:
+            generated_outputs (`torch.Tensor` or `np.ndarray`):
+                The output of the model. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
+                containing the token ids of the generated sequences.
+
+        Returns:
+            `List[str]`: The decoded text output.
+        """
+        beginning_of_answer = self.tokenizer.convert_tokens_to_ids(BEGINNING_OF_ANSWER_STRING)
+        # get boa index for each outputted sequence tensor
+        # start all generated sequences from the beginning of the answer token, pad to have consistent length
+        unpadded_output_sequences = [
+            seq[(seq == beginning_of_answer).nonzero(as_tuple=True)[0] + 1 :] for seq in generated_outputs
+        ]
+        max_len = max(len(seq) for seq in unpadded_output_sequences)
+        # convert to torch and pad sequences
+        padded_output_sequences = torch.full((len(unpadded_output_sequences), max_len), self.pad_token_id)
+        for i, seq in enumerate(unpadded_output_sequences):
+            padded_output_sequences[i, : len(seq)] = torch.tensor(seq)
+
+        return self.batch_decode(padded_output_sequences, skip_special_tokens=True)
+
     def batch_decode(self, *args, **kwargs):
         """
         This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please

src/transformers/models/git/processing_git.py

@@ -16,18 +16,23 @@
 Image/Text processor class for GIT
 """
 
+import logging
 from typing import List, Optional, Union
 
 from ...feature_extraction_utils import BatchFeature
 from ...image_utils import ImageInput
 from ...processing_utils import ProcessingKwargs, ProcessorMixin, Unpack, _validate_images_text_input_order
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
+from ...utils.deprecation import deprecate_kwarg
 
 
 class GitProcessorKwargs(ProcessingKwargs, total=False):
     _defaults = {}
 
 
+logger = logging.getLogger(__name__)
+
+
 class GitProcessor(ProcessorMixin):
     r"""
     Constructs a GIT processor which wraps a CLIP image processor and a BERT tokenizer into a single processor.
@@ -50,6 +55,7 @@ def __init__(self, image_processor, tokenizer):
         super().__init__(image_processor, tokenizer)
         self.current_processor = self.image_processor
 
+    @deprecate_kwarg(old_name="legacy", version="5.0.0")
     def __call__(
         self,
         images: Optional[ImageInput] = None,
@@ -91,6 +97,14 @@ def __call__(
               `None`).
             - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
         """
+        legacy = kwargs.pop("legacy", True)
+        if legacy:
+            logger.warning(
+                "Legacy behavior is being used. The new behavior with legacy=False will be enabled in the future."
+                "In the new behavior, if both images and text are provided, the last token (EOS token) "
+                "of the input_ids and attention_mask tensors will be removed."
+            )
+
         if text is None and images is None:
             raise ValueError("You have to specify either text or images. Both cannot be none.")
 
@@ -110,6 +124,10 @@ def __call__(
         if images is not None:
             image_features = self.image_processor(images, **output_kwargs["images_kwargs"])
             data.update(image_features)
+            if not legacy:
+                data["input_ids"] = data["input_ids"][:, :-1]
+                data["attention_mask"] = data["attention_mask"][:, :-1]
+
         return BatchFeature(data=data, tensor_type=output_kwargs["common_kwargs"].get("return_tensors"))
 
     def batch_decode(self, *args, **kwargs):

src/transformers/models/kosmos2/processing_kosmos2.py

@@ -428,6 +428,21 @@ def post_process_generation(self, text, cleanup_and_extract=True):
             return clean_text_and_extract_entities_with_bboxes(caption)
         return caption
 
+    def post_process_image_text_to_text(self, generated_outputs):
+        """
+        Post-process the output of the model to decode the text.
+
+        Args:
+            generated_outputs (`torch.Tensor` or `np.ndarray`):
+                The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
+                or `(sequence_length,)`.
+
+        Returns:
+            `List[str]`: The decoded text.
+        """
+        generated_texts = self.batch_decode(generated_outputs, skip_special_tokens=True)
+        return [self.post_process_generation(text, cleanup_and_extract=False) for text in generated_texts]
+
     @property
     # Copied from transformers.models.blip.processing_blip.BlipProcessor.model_input_names
     def model_input_names(self):

src/transformers/models/mllama/processing_mllama.py

@@ -342,6 +342,22 @@ def decode(self, *args, **kwargs):
         """
         return self.tokenizer.decode(*args, **kwargs)
 
+    def post_process_image_text_to_text(self, generated_outputs):
+        """
+        Post-process the output of the model to decode the text.
+
+        Args:
+            generated_outputs (`torch.Tensor` or `np.ndarray`):
+                The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
+                or `(sequence_length,)`.
+
+        Returns:
+            `List[str]`: The decoded text.
+        """
+        return self.tokenizer.batch_decode(
+            generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False
+        )
+
     @property
     def model_input_names(self):
         tokenizer_input_names = self.tokenizer.model_input_names

src/transformers/models/pix2struct/processing_pix2struct.py

@@ -16,11 +16,13 @@
 Processor class for Pix2Struct.
 """
 
+import logging
 from typing import List, Optional, Union
 
 from ...feature_extraction_utils import BatchFeature
 from ...processing_utils import ImagesKwargs, ProcessingKwargs, ProcessorMixin, Unpack
 from ...tokenization_utils_base import BatchEncoding, PreTokenizedInput, TextInput
+from ...utils.deprecation import deprecate_kwarg
 
 
 class Pix2StructImagesKwargs(ImagesKwargs, total=False):
@@ -48,6 +50,9 @@ class Pix2StructProcessorKwargs(ProcessingKwargs, total=False):
     }
 
 
+logger = logging.getLogger(__name__)
+
+
 class Pix2StructProcessor(ProcessorMixin):
     r"""
     Constructs a PIX2STRUCT processor which wraps a BERT tokenizer and PIX2STRUCT image processor into a single
@@ -71,6 +76,7 @@ def __init__(self, image_processor, tokenizer):
         tokenizer.return_token_type_ids = False
         super().__init__(image_processor, tokenizer)
 
+    @deprecate_kwarg(old_name="legacy", version="5.0.0")
     def __call__(
         self,
         images=None,
@@ -85,6 +91,14 @@ def __call__(
 
         Please refer to the docstring of the above two methods for more information.
         """
+        legacy = kwargs.pop("legacy", True)
+        if legacy:
+            logger.warning(
+                "Legacy behavior is being used. The new behavior with legacy=False will be enabled in the future."
+                "In the new behavior, If both images and text are provided, image_processor is not a VQA processor, and `add_special_tokens` is unset, "
+                "the default value of `add_special_tokens` will be changed to `False` when calling the tokenizer."
+            )
+
         if images is None and text is None:
             raise ValueError("You have to specify either images or text.")
 
@@ -93,8 +107,12 @@ def __call__(
             tokenizer_init_kwargs=self.tokenizer.init_kwargs,
             **kwargs,
         )
+        add_special_tokens = output_kwargs["text_kwargs"].pop("add_special_tokens", None)
         # Get only text
         if images is None and not self.image_processor.is_vqa:
+            output_kwargs["text_kwargs"]["add_special_tokens"] = (
+                add_special_tokens if add_special_tokens is not None else True
+            )
             self.current_processor = self.tokenizer
             text_encoding = self.tokenizer(text=text, **output_kwargs["text_kwargs"])
             return text_encoding
@@ -108,6 +126,9 @@ def __call__(
             encoding_image_processor = self.image_processor(images, **output_kwargs["images_kwargs"])
 
         if text is not None and not self.image_processor.is_vqa:
+            output_kwargs["text_kwargs"]["add_special_tokens"] = (
+                add_special_tokens if add_special_tokens is not None else legacy
+            )
             text_encoding = self.tokenizer(text=text, **output_kwargs["text_kwargs"])
 
             if "attention_mask" in text_encoding: