Upload processor

preprocessor_config.json

@@ -1,6 +1,6 @@
 {
   "auto_map": {
-    "AutoProcessor": "microsoft/Florence-2-base--processing_florence2.Florence2Processor"
+    "AutoProcessor": "processing_florence2.Florence2Processor"
   },
   "crop_size": {
     "height": 768,

processing_florence2.py

@@ -0,0 +1,1088 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Processor class for Florence-2.
+"""
+
+import re
+import logging
+from typing import List, Optional, Union
+import numpy as np
+
+import torch
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput, is_valid_image
+from transformers.processing_utils import ProcessorMixin
+from transformers.tokenization_utils_base import (
+    PaddingStrategy,
+    PreTokenizedInput,
+    TextInput,
+    TruncationStrategy,
+)
+from transformers.utils import TensorType
+
+
+logger = logging.getLogger(__name__)
+
+# Copied from transformers.models.idefics2.processing_idefics2.is_url
+def is_url(val) -> bool:
+    return isinstance(val, str) and val.startswith("http")
+
+# Copied from transformers.models.idefics2.processing_idefics2.is_image_or_image_url
+def is_image_or_image_url(elem):
+    return is_url(elem) or is_valid_image(elem)
+
+
+def _is_str_or_image(elem):
+    return isinstance(elem, (str)) or is_image_or_image_url(elem)
+
+
+class Florence2Processor(ProcessorMixin):
+    r"""
+    Constructs a Florence2 processor which wraps a Florence2 image processor and a Florence2 tokenizer into a single processor.
+
+    [`Florence2Processor`] offers all the functionalities of [`CLIPImageProcessor`] and [`BartTokenizerFast`]. See the
+    [`~Florence2Processor.__call__`] and [`~Florence2Processor.decode`] for more information.
+
+    Args:
+        image_processor ([`CLIPImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`BartTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "CLIPImageProcessor"
+    tokenizer_class = ("BartTokenizer", "BartTokenizerFast")
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+    ):
+        if image_processor is None:
+            raise ValueError("You need to specify an `image_processor`.")
+        if tokenizer is None:
+            raise ValueError("You need to specify a `tokenizer`.")
+        if not hasattr(image_processor, "image_seq_length"):
+            raise ValueError("Image processor is missing an `image_seq_length` attribute.")
+
+        self.image_seq_length = image_processor.image_seq_length
+
+        tokens_to_add = {
+                'additional_special_tokens': \
+                    tokenizer.additional_special_tokens + \
+                    ['<od>', '</od>', '<ocr>', '</ocr>'] + \
+                    [f'<loc_{x}>' for x in range(1000)] + \
+                    ['<cap>', '</cap>', '<ncap>', '</ncap>','<dcap>', '</dcap>', '<grounding>', '</grounding>', '<seg>', '</seg>', '<sep>', '<region_cap>', '</region_cap>', '<region_to_desciption>', '</region_to_desciption>', '<proposal>', '</proposal>', '<poly>', '</poly>', '<and>']
+            }
+        tokenizer.add_special_tokens(tokens_to_add)
+
+        self.tasks_answer_post_processing_type = {
+            '<OCR>': 'pure_text',
+            '<OCR_WITH_REGION>': 'ocr',
+            '<CAPTION>': 'pure_text',
+            '<DETAILED_CAPTION>': 'pure_text',
+            '<MORE_DETAILED_CAPTION>': 'pure_text',
+            '<OD>': 'description_with_bboxes',
+            '<DENSE_REGION_CAPTION>': 'description_with_bboxes',
+            '<CAPTION_TO_PHRASE_GROUNDING>': "phrase_grounding",
+            '<REFERRING_EXPRESSION_SEGMENTATION>': 'polygons',
+            '<REGION_TO_SEGMENTATION>': 'polygons',
+            '<OPEN_VOCABULARY_DETECTION>': 'description_with_bboxes_or_polygons',
+            '<REGION_TO_CATEGORY>': 'pure_text',
+            '<REGION_TO_DESCRIPTION>': 'pure_text',
+            '<REGION_TO_OCR>': 'pure_text',
+            '<REGION_PROPOSAL>': 'bboxes'
+        }
+
+        self.task_prompts_without_inputs = {
+            '<OCR>': 'What is the text in the image?',
+            '<OCR_WITH_REGION>': 'What is the text in the image, with regions?',
+            '<CAPTION>': 'What does the image describe?',
+            '<DETAILED_CAPTION>': 'Describe in detail what is shown in the image.',
+            '<MORE_DETAILED_CAPTION>': 'Describe with a paragraph what is shown in the image.',
+            '<OD>': 'Locate the objects with category name in the image.',
+            '<DENSE_REGION_CAPTION>': 'Locate the objects in the image, with their descriptions.',
+            '<REGION_PROPOSAL>': 'Locate the region proposals in the image.'
+        }
+
+        self.task_prompts_with_input = {
+            '<CAPTION_TO_PHRASE_GROUNDING>': "Locate the phrases in the caption: {input}",
+            '<REFERRING_EXPRESSION_SEGMENTATION>': 'Locate {input} in the image with mask',
+            '<REGION_TO_SEGMENTATION>': 'What is the polygon mask of region {input}',
+            '<OPEN_VOCABULARY_DETECTION>': 'Locate {input} in the image.',
+            '<REGION_TO_CATEGORY>': 'What is the region {input}?',
+            '<REGION_TO_DESCRIPTION>': 'What does the region {input} describe?',
+            '<REGION_TO_OCR>': 'What text is in the region {input}?',
+        }
+
+        self.post_processor = Florence2PostProcesser(tokenizer=tokenizer)
+
+
+        super().__init__(image_processor, tokenizer)
+    
+    def _construct_prompts(self, text):
+        # replace the task tokens with the task prompts if task token is in the text
+        prompts = []
+        for _text in text:
+            # 1. fixed task prompts without additional inputs
+            for task_token, task_prompt in self.task_prompts_without_inputs.items():
+                if task_token in _text:
+                    assert _text == task_token, f"Task token {task_token} should be the only token in the text."
+                    _text = task_prompt
+                    break
+            # 2. task prompts with additional inputs 
+            for task_token, task_prompt in self.task_prompts_with_input.items():
+                if task_token in _text:
+                    _text = task_prompt.format(input=_text.replace(task_token, ''))
+                    break
+            prompts.append(_text)
+        return prompts
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        images: ImageInput = None,
+        tokenize_newline_separately: bool = True,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length=None,
+        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
+        do_resize: bool = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        data_format: Optional["ChannelDimension"] = "channels_first",  # noqa: F821
+        input_data_format: Optional[
+            Union[str, "ChannelDimension"]  # noqa: F821
+        ] = None,
+        resample: "PILImageResampling" = None,  # noqa: F821
+        do_convert_rgb: bool = None,
+        do_thumbnail: bool = None,
+        do_align_long_axis: bool = None,
+        do_rescale: bool = None,
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to BartTokenizerFast's [`~BartTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        CLIPImageProcessor's [`~CLIPImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
+        of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
+                number of channels, H and W are image height and width.
+            tokenize_newline_separately (`bool`, defaults to `True`):
+                Adds a separately tokenized '\n' at the end of the prompt.
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
+                Select a strategy to pad the returned sequences (according to the model's padding side and padding
+                index) among:
+                - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+                  sequence if provided).
+                - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
+                  acceptable input length for the model if that argument is not provided.
+                - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
+                  lengths).
+            max_length (`int`, *optional*):
+                Maximum length of the returned list and optionally padding length (see above).
+            truncation (`bool`, *optional*):
+                Activates truncation to cut input sequences longer than `max_length` to `max_length`.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`. If `suffix`
+              is provided, the `input_ids` will also contain the suffix input ids.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **labels** -- Labels compatible with training if `suffix` is not None
+        """
+
+        return_token_type_ids = False
+
+        if images is None:
+            raise ValueError("`images` are expected as arguments to a `Florence2Processor` instance.")
+        if text is None:
+            logger.warning_once(
+                "You are using Florence-2 without a text prompt."
+            )
+            text = ""
+
+        if isinstance(text, List) and isinstance(images, List):
+            if len(images) < len(text):
+                raise ValueError(
+                    f"Received {len(images)} images for {len(text)} prompts. Each prompt should be associated with an image."
+                )
+        if _is_str_or_image(text):
+            text = [text]
+        elif isinstance(text, list) and _is_str_or_image(text[0]):
+            pass
+
+        pixel_values = self.image_processor(
+            images,
+            do_resize=do_resize,
+            do_normalize=do_normalize,
+            return_tensors=return_tensors,
+            image_mean=image_mean,
+            image_std=image_std,
+            input_data_format=input_data_format,
+            data_format=data_format,
+            resample=resample,
+            do_convert_rgb=do_convert_rgb,
+        )["pixel_values"]
+
+        if max_length is not None:
+            max_length -= self.image_seq_length  # max_length has to account for the image tokens
+
+        text = self._construct_prompts(text)
+
+        inputs = self.tokenizer(
+            text,
+            return_tensors=return_tensors,
+            padding=padding,
+            max_length=max_length,
+            truncation=truncation,
+            return_token_type_ids=return_token_type_ids,
+        )
+
+        return_data = {**inputs, "pixel_values": pixel_values}
+
+        if return_token_type_ids:
+            labels = inputs["input_ids"].masked_fill(inputs["token_type_ids"] == 0, -100)
+            return_data.update({"labels": labels})
+        return BatchFeature(data=return_data)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Florence2
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BartTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Florence2
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BartTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names with CLIP->Florence2
+    def model_input_names(self):
+        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))
+
+    def post_process_generation(self, text, task, image_size):
+        """
+        Post-process the output of the model to each of the task outputs.
+
+        Args:
+            text (`str`): The text to post-process.
+            task (`str`): The task to post-process the text for.
+            image_size (`Tuple[int, int]`): The size of the image. height x width.
+        """
+
+        task_answer_post_processing_type = self.tasks_answer_post_processing_type.get(task, 'pure_text')
+        task_answer = self.post_processor(
+            text=text,
+            image_size=image_size,
+            parse_tasks=task_answer_post_processing_type,
+        )[task_answer_post_processing_type]
+
+        if task_answer_post_processing_type == 'pure_text':
+            final_answer = task_answer
+            # remove the special tokens
+            final_answer = final_answer.replace('<s>', '').replace('</s>', '')
+        elif task_answer_post_processing_type in ['od', 'description_with_bboxes', 'bboxes']:
+            od_instances = task_answer
+            bboxes_od = [_od_instance['bbox'] for _od_instance in od_instances]
+            labels_od = [str(_od_instance['cat_name']) for _od_instance in od_instances]
+            final_answer = {'bboxes': bboxes_od, 'labels': labels_od}
+        elif task_answer_post_processing_type in ['ocr']:
+            bboxes = [_od_instance['quad_box'] for _od_instance in task_answer]
+            labels = [str(_od_instance['text']) for _od_instance in task_answer]
+            final_answer = {'quad_boxes': bboxes, 'labels': labels}
+        elif task_answer_post_processing_type in ['phrase_grounding']:
+            bboxes = []
+            labels = []
+            for _grounded_phrase in task_answer:
+                for _bbox in _grounded_phrase['bbox']:
+                    bboxes.append(_bbox)
+                    labels.append(_grounded_phrase['cat_name'])
+            final_answer = {'bboxes': bboxes, 'labels': labels}
+        elif task_answer_post_processing_type in ['description_with_polygons', 'polygons']:
+            labels = []
+            polygons = []
+            for result in task_answer:
+                label = result['cat_name']
+                _polygons = result['polygons']
+                labels.append(label)
+                polygons.append(_polygons)
+            final_answer = {'polygons': polygons, 'labels': labels}
+        elif task_answer_post_processing_type in ['description_with_bboxes_or_polygons']:
+            bboxes = []
+            bboxes_labels = []
+            polygons = []
+            polygons_labels = []
+            for result in task_answer:
+                label = result['cat_name']
+                if 'polygons' in result:
+                    _polygons = result['polygons']
+                    polygons.append(_polygons)
+                    polygons_labels.append(label)
+                else:
+                    _bbox = result['bbox']
+                    bboxes.append(_bbox)
+                    bboxes_labels.append(label)
+            final_answer = {'bboxes': bboxes, 'bboxes_labels': bboxes_labels, 'polygons': polygons, 'polygons_labels': polygons_labels}
+        else:
+            raise ValueError('Unknown task answer post processing type: {}'.format(task_answer_post_processing_type))
+
+        final_answer = {
+            task: final_answer}
+        return final_answer 
+
+class BoxQuantizer(object):
+    def __init__(self, mode, bins):
+        self.mode = mode
+        self.bins = bins
+
+    def quantize(self, boxes: torch.Tensor, size):
+        bins_w, bins_h = self.bins  # Quantization bins.
+        size_w, size_h = size       # Original image size.
+        size_per_bin_w = size_w / bins_w
+        size_per_bin_h = size_h / bins_h
+        xmin, ymin, xmax, ymax = boxes.split(1, dim=-1)  # Shape: 4 * [N, 1].
+
+        if self.mode == 'floor':
+            quantized_xmin = (
+                xmin / size_per_bin_w).floor().clamp(0, bins_w - 1)
+            quantized_ymin = (
+                ymin / size_per_bin_h).floor().clamp(0, bins_h - 1)
+            quantized_xmax = (
+                xmax / size_per_bin_w).floor().clamp(0, bins_w - 1)
+            quantized_ymax = (
+                ymax / size_per_bin_h).floor().clamp(0, bins_h - 1)
+
+        elif self.mode == 'round':
+            raise NotImplementedError()
+
+        else:
+            raise ValueError('Incorrect quantization type.')
+
+        quantized_boxes = torch.cat(
+            (quantized_xmin, quantized_ymin, quantized_xmax, quantized_ymax), dim=-1
+        ).int()
+
+        return quantized_boxes
+
+    def dequantize(self, boxes: torch.Tensor, size):
+        bins_w, bins_h = self.bins  # Quantization bins.
+        size_w, size_h = size       # Original image size.
+        size_per_bin_w = size_w / bins_w
+        size_per_bin_h = size_h / bins_h
+        xmin, ymin, xmax, ymax = boxes.split(1, dim=-1)  # Shape: 4 * [N, 1].
+
+        if self.mode == 'floor':
+            # Add 0.5 to use the center position of the bin as the coordinate.
+            dequantized_xmin = (xmin + 0.5) * size_per_bin_w
+            dequantized_ymin = (ymin + 0.5) * size_per_bin_h
+            dequantized_xmax = (xmax + 0.5) * size_per_bin_w
+            dequantized_ymax = (ymax + 0.5) * size_per_bin_h
+
+        elif self.mode == 'round':
+            raise NotImplementedError()
+
+        else:
+            raise ValueError('Incorrect quantization type.')
+
+        dequantized_boxes = torch.cat(
+            (dequantized_xmin, dequantized_ymin,
+             dequantized_xmax, dequantized_ymax), dim=-1
+        )
+
+        return dequantized_boxes
+
+
+class CoordinatesQuantizer(object):
+    """
+    Quantize coornidates (Nx2)
+    """
+
+    def __init__(self, mode, bins):
+        self.mode = mode
+        self.bins = bins
+
+    def quantize(self, coordinates: torch.Tensor, size):
+        bins_w, bins_h = self.bins  # Quantization bins.
+        size_w, size_h = size       # Original image size.
+        size_per_bin_w = size_w / bins_w
+        size_per_bin_h = size_h / bins_h
+        assert coordinates.shape[-1] == 2, 'coordinates should be shape (N, 2)'
+        x, y = coordinates.split(1, dim=-1)  # Shape: 4 * [N, 1].
+
+        if self.mode == 'floor':
+            quantized_x = (x / size_per_bin_w).floor().clamp(0, bins_w - 1)
+            quantized_y = (y / size_per_bin_h).floor().clamp(0, bins_h - 1)
+
+        elif self.mode == 'round':
+            raise NotImplementedError()
+
+        else:
+            raise ValueError('Incorrect quantization type.')
+
+        quantized_coordinates = torch.cat(
+            (quantized_x, quantized_y), dim=-1
+        ).int()
+
+        return quantized_coordinates
+
+    def dequantize(self, coordinates: torch.Tensor, size):
+        bins_w, bins_h = self.bins  # Quantization bins.
+        size_w, size_h = size       # Original image size.
+        size_per_bin_w = size_w / bins_w
+        size_per_bin_h = size_h / bins_h
+        assert coordinates.shape[-1] == 2, 'coordinates should be shape (N, 2)'
+        x, y = coordinates.split(1, dim=-1)  # Shape: 4 * [N, 1].
+
+        if self.mode == 'floor':
+            # Add 0.5 to use the center position of the bin as the coordinate.
+            dequantized_x = (x + 0.5) * size_per_bin_w
+            dequantized_y = (y + 0.5) * size_per_bin_h
+
+        elif self.mode == 'round':
+            raise NotImplementedError()
+
+        else:
+            raise ValueError('Incorrect quantization type.')
+
+        dequantized_coordinates = torch.cat(
+            (dequantized_x, dequantized_y), dim=-1
+        )
+
+        return dequantized_coordinates
+
+
+class Florence2PostProcesser(object):
+    """
+    Florence-2 post process for converting text prediction to various tasks results. 
+
+    Args:
+        config: A dict of configs.
+        tokenizer: A tokenizer for decoding text to spans.
+        sample config:
+            UNIFIED_POST_PROCESS:
+                # commom configs
+                NUM_BBOX_HEIGHT_BINS: 1000
+                NUM_BBOX_WIDTH_BINS: 1000
+                COORDINATES_HEIGHT_BINS: 1000
+                COORDINATES_WIDTH_BINS: 1000
+                # task specific configs, override the common configs
+                PRASE_TASKS:
+                    - TASK_NAME: 'video_dense_caption'
+                      PATTERN: 'r<time_(\d+)><time_(\d+)>([a-zA-Z0-9 ]+)'
+                      SCORE_MODE: 'avg_cat_name_scores'
+                      NUM_BINS: 100
+                    - TASK_NAME: 'od'
+                      PATTERN: 'r<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>([a-zA-Z0-9 ]+)'
+                      SCORE_MODE: 'avg_cat_name_scores'
+
+    Returns:
+        parsed_dict (dict): A dict of parsed results.
+    """
+    def __init__(
+        self,
+        tokenizer=None
+    ):
+        parse_tasks = []
+        parse_task_configs = {}
+        config = self._create_default_config()
+        for task in config['PARSE_TASKS']:
+            parse_tasks.append(task['TASK_NAME'])
+            parse_task_configs[task['TASK_NAME']] = task
+
+        self.config = config
+        self.parse_tasks = parse_tasks
+        self.parse_tasks_configs = parse_task_configs
+
+        self.tokenizer =  tokenizer
+        if self.tokenizer is not None:
+            self.all_special_tokens = set(self.tokenizer.all_special_tokens)
+
+        self.init_quantizers()
+        self.black_list_of_phrase_grounding = self._create_black_list_of_phrase_grounding()
+
+    def _create_black_list_of_phrase_grounding(self):
+        black_list = {}
+
+        if 'phrase_grounding' in self.parse_tasks and self.parse_tasks_configs['phrase_grounding']['FILTER_BY_BLACK_LIST']:
+            black_list =  set(
+                ['it', 'I', 'me', 'mine',
+                 'you', 'your', 'yours',
+                 'he', 'him', 'his',
+                 'she', 'her', 'hers',
+                 'they', 'them', 'their', 'theirs',
+                 'one', 'oneself',
+                 'we', 'us', 'our', 'ours',
+                 'you', 'your', 'yours',
+                 'they', 'them', 'their', 'theirs',
+                 'mine', 'yours', 'his', 'hers', 'its',
+                 'ours', 'yours', 'theirs',
+                 'myself', 'yourself', 'himself', 'herself', 'itself',
+                 'ourselves', 'yourselves', 'themselves',
+                 'this', 'that',
+                 'these', 'those',
+                 'who', 'whom', 'whose', 'which', 'what',
+                 'who', 'whom', 'whose', 'which', 'that',
+                 'all', 'another', 'any', 'anybody', 'anyone', 'anything',
+                 'each', 'everybody', 'everyone', 'everything',
+                 'few', 'many', 'nobody', 'none', 'one', 'several',
+                 'some', 'somebody', 'someone', 'something',
+                 'each other', 'one another',
+                 'myself', 'yourself', 'himself', 'herself', 'itself',
+                 'ourselves', 'yourselves', 'themselves',
+                 'the image', 'image', 'images', 'the', 'a', 'an', 'a group',
+                 'other objects', 'lots', 'a set',
+                 ]
+            )
+
+        return black_list
+    
+    def _create_default_config(self):
+        config = {
+            'NUM_BBOX_HEIGHT_BINS': 1000,
+            'NUM_BBOX_WIDTH_BINS': 1000,
+            'BOX_QUANTIZATION_MODE': 'floor',
+            'COORDINATES_HEIGHT_BINS': 1000,
+            'COORDINATES_WIDTH_BINS': 1000,
+            'COORDINATES_QUANTIZATION_MODE': 'floor',
+            'PARSE_TASKS': [
+                {
+                    'TASK_NAME': 'od',
+                    'PATTERN': r'([a-zA-Z0-9 ]+)<loc_(\\d+)><loc_(\\d+)><loc_(\\d+)><loc_(\\d+)>'
+                },
+                {
+                    'TASK_NAME': 'ocr',
+                    'PATTERN':  r'(.+?)<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>',
+                    'AREA_THRESHOLD': 0.00
+                },
+                {
+                    'TASK_NAME': 'phrase_grounding',
+                    'FILTER_BY_BLACK_LIST': True
+                },
+                {
+                    'TASK_NAME': 'pure_text',
+                },
+                {
+                    'TASK_NAME': 'description_with_bboxes',
+                },
+                {
+                    'TASK_NAME': 'description_with_polygons',
+                },
+                {
+                    'TASK_NAME': 'polygons',
+                },
+                {
+                    'TASK_NAME': 'bboxes',
+                },
+                {
+                    'TASK_NAME': 'description_with_bboxes_or_polygons',
+                }
+            ]
+        }
+
+        return config
+
+    def init_quantizers(self):
+        # we have box_quantizer (od, grounding) and coordinates_quantizer (ocr, referring_segmentation)
+        num_bbox_height_bins = self.config.get('NUM_BBOX_HEIGHT_BINS', 1000)
+        num_bbox_width_bins = self.config.get('NUM_BBOX_WIDTH_BINS', 1000)
+        box_quantization_mode = self.config.get('BOX_QUANTIZATION_MODE', 'floor')
+        self.box_quantizer = BoxQuantizer(
+            box_quantization_mode,
+            (num_bbox_width_bins, num_bbox_height_bins),
+        )
+        
+        num_bbox_height_bins = self.config['COORDINATES_HEIGHT_BINS'] if 'COORDINATES_HEIGHT_BINS' in self.config else self.config.get('NUM_BBOX_HEIGHT_BINS', 1000)
+        num_bbox_width_bins = self.config['COORDINATES_WIDTH_BINS'] if 'COORDINATES_WIDTH_BINS' in self.config else self.config.get('NUM_BBOX_WIDTH_BINS', 1000)
+        box_quantization_mode = self.config.get('COORDINATES_QUANTIZATION_MODE') if 'COORDINATES_QUANTIZATION_MODE' in self.config else self.config.get('BOX_QUANTIZATION_MODE', 'floor')
+        self.coordinates_quantizer = CoordinatesQuantizer(
+            box_quantization_mode,
+            (num_bbox_width_bins, num_bbox_height_bins),
+        )
+
+    def decode_with_spans(self, tokenizer, token_ids):
+        filtered_tokens = tokenizer.convert_ids_to_tokens(
+            token_ids, skip_special_tokens=False)
+        assert len(filtered_tokens) == len(token_ids)
+
+        # To avoid mixing byte-level and unicode for byte-level BPT
+        # we need to build string separately for added tokens and byte-level tokens
+        # cf. https://github.com/huggingface/transformers/issues/1133
+        sub_texts = []
+        for token in filtered_tokens:
+            if token in self.all_special_tokens:
+                sub_texts.append(token)
+            else:
+                if isinstance(tokenizer, (BartTokenizer, BartTokenizerFast)):
+                    sub_text = tokenizer.convert_tokens_to_string([token])
+                elif isinstance(tokenizer, (T5Tokenizer, T5TokenizerFast)):
+                    # Ref: https://github.com/google/sentencepiece#whitespace-is-treated-as-a-basic-symbol
+                    # Note: Do not strip sub_text as it may have functional whitespace
+                    sub_text = token.replace('▁', ' ')
+                else:
+                    raise ValueError(f'type {type(tokenizer)} not supported')
+                sub_texts.append(sub_text)
+
+        text = ''
+        spans = []
+        for sub_text in sub_texts:
+            span = (len(text), len(text) + len(sub_text))  # [start index, end index).
+            text += sub_text
+            spans.append(span)
+
+        # Text format:
+        # 1. T5Tokenizer/T5TokenizerFast: 
+        #      "<loc_1><loc_2><loc_3><loc_4> transplanting dog<loc_1><loc_2><loc_3><loc_4> cat</s>"
+        #    Equivalent to t5_tokenizer.decode(input_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False, spaces_between_special_tokens=False)
+        # 2. BartTokenizer (need to double check):
+        #      "<s><loc_1><loc_2><loc_3><loc_4>transplanting dog<loc_1><loc_2><loc_3><loc_4>cat</s>"
+        #    Equivalent to bart_tokenizer.decode(input_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False, spaces_between_special_tokens=False)
+        return text, spans
+
+    def parse_od_from_text_and_spans(
+        self,
+        text,
+        pattern,
+        image_size,
+        phrase_centric=False
+    ):
+        parsed = list(re.finditer(pattern, text))
+
+        instances = []
+        for i in range(len(parsed)):
+            # Prepare instance.
+            instance = {}
+
+            if phrase_centric:
+                bbox_bins = [int(parsed[i].group(j)) for j in range(2, 6)]
+            else:
+                bbox_bins = [int(parsed[i].group(j)) for j in range(1, 5)]
+            instance['bbox'] = self.box_quantizer.dequantize(
+                boxes=torch.tensor(bbox_bins),
+                size=image_size
+            ).tolist()  
+
+            if phrase_centric:
+                instance['cat_name'] = parsed[i].group(1).lower().strip()
+            else:
+                instance['cat_name'] = parsed[i].group(5).lower().strip()
+            instances.append(instance)
+
+        return instances
+
+    def parse_ocr_from_text_and_spans(self, 
+                                    text, 
+                                     pattern, 
+                                     image_size,
+                                     area_threshold=-1.0,
+        ):
+        bboxes = []
+        labels = []
+        text = text.replace('<s>', '')
+        # ocr with regions
+        parsed = re.findall(pattern, text)
+        instances = []
+        image_width, image_height = image_size
+
+        for ocr_line in parsed:
+            ocr_content = ocr_line[0]
+            quad_box = ocr_line[1:]
+            quad_box = [int(i) for i in quad_box]
+            quad_box = self.coordinates_quantizer.dequantize(
+                torch.tensor(np.array(quad_box).reshape(-1, 2)),
+                size=image_size
+            ).reshape(-1).tolist()
+
+            if area_threshold > 0:
+                x_coords = [i for i in quad_box[0::2]]
+                y_coords = [i for i in quad_box[1::2]]
+
+                # apply the Shoelace formula
+                area = 0.5 * abs(sum(x_coords[i] * y_coords[i + 1] - x_coords[i + 1] * y_coords[i] for i in range(4 - 1)))
+
+                if area < (image_width * image_height) * area_threshold:
+                    continue
+
+            bboxes.append(quad_box)
+            labels.append(ocr_content)
+            instances.append({
+                'quad_box': quad_box,
+                'text': ocr_content,
+            })
+        return instances
+
+    def parse_phrase_grounding_from_text_and_spans(self, text, pattern, image_size):
+        # ignore <s> </s> and <pad>
+        cur_span = 0
+        if text.startswith('<s>'):   
+            cur_span += 3
+
+        text = text.replace('<s>', '')
+        text = text.replace('</s>', '')
+        text = text.replace('<pad>', '')
+
+        pattern = r"([^<]+(?:<loc_\d+>){4,})"
+        phrases = re.findall(pattern, text)
+    
+        # pattern should be text pattern and od pattern
+        pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_)'
+        box_pattern = r'<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>'
+
+        instances = []
+        for pharse_text in phrases:
+            phrase_text_strip = pharse_text.replace('<ground>', '', 1)
+            phrase_text_strip = pharse_text.replace('<obj>', '', 1)
+
+            if phrase_text_strip == '':
+                cur_span += len(pharse_text)
+                continue
+
+            # Prepare instance.
+            instance = {}
+
+            # parse phrase, get string 
+            phrase = re.search(pattern, phrase_text_strip)
+            if phrase is None:
+                cur_span += len(pharse_text)
+                continue
+
+            # parse bboxes by box_pattern
+            bboxes_parsed = list(re.finditer(box_pattern, pharse_text))
+            if len(bboxes_parsed) == 0:
+                cur_span += len(pharse_text)
+                continue
+
+            phrase = phrase.group()
+            # remove leading and trailing spaces
+            phrase = phrase.strip()
+
+            if phrase in self.black_list_of_phrase_grounding:
+                cur_span += len(pharse_text)
+                continue
+
+            # a list of list 
+            bbox_bins = [[int(_bboxes_parsed.group(j)) for j in range(1, 5)] for _bboxes_parsed in bboxes_parsed]
+            instance['bbox'] = self.box_quantizer.dequantize(
+                boxes=torch.tensor(bbox_bins),
+                size=image_size
+            ).tolist()  
+
+            # exclude non-ascii characters
+            phrase = phrase.encode('ascii',errors='ignore').decode('ascii')
+            instance['cat_name'] = phrase
+
+            instances.append(instance)
+
+        return instances
+
+    def parse_description_with_bboxes_from_text_and_spans(self, text, pattern, image_size, allow_empty_phrase=False):
+        # temporary parse solution, split by '.'
+        # ignore <s> </s> and <pad>
+
+        text = text.replace('<s>', '')
+        text = text.replace('</s>', '')
+        text = text.replace('<pad>', '')
+
+        if allow_empty_phrase:
+            pattern = rf"(?:(?:<loc_\d+>){{4,}})"
+        else:
+            pattern = r"([^<]+(?:<loc_\d+>){4,})"
+        phrases = re.findall(pattern, text)
+    
+        # pattern should be text pattern and od pattern
+        pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_)'
+        box_pattern = r'<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>'
+
+        instances = []
+        for pharse_text in phrases:
+            phrase_text_strip = pharse_text.replace('<ground>', '', 1)
+            phrase_text_strip = pharse_text.replace('<obj>', '', 1)
+
+            if phrase_text_strip == '' and not allow_empty_phrase:
+                continue
+
+            # parse phrase, get string 
+            phrase = re.search(pattern, phrase_text_strip)
+            if phrase is None:
+                continue
+
+            phrase = phrase.group()
+            # remove leading and trailing spaces
+            phrase = phrase.strip()
+
+            # parse bboxes by box_pattern
+            bboxes_parsed = list(re.finditer(box_pattern, pharse_text))
+            if len(bboxes_parsed) == 0:
+                continue
+
+            # a list of list 
+            bbox_bins = [[int(_bboxes_parsed.group(j)) for j in range(1, 5)] for _bboxes_parsed in bboxes_parsed]
+
+            bboxes = self.box_quantizer.dequantize(
+                boxes=torch.tensor(bbox_bins),
+                size=image_size
+            ).tolist()  
+
+            phrase = phrase.encode('ascii',errors='ignore').decode('ascii')
+            for _bboxes in bboxes:
+                # Prepare instance.
+                instance = {}
+                instance['bbox'] = _bboxes
+                # exclude non-ascii characters
+                instance['cat_name'] = phrase
+                instances.append(instance)
+
+        return instances
+
+    def parse_description_with_polygons_from_text_and_spans(self, text, pattern, image_size, 
+                                                            allow_empty_phrase=False,
+                                                            polygon_sep_token='<sep>',
+                                                            polygon_start_token='<poly>',
+                                                            polygon_end_token='</poly>',
+                                                            with_box_at_start=False,
+                                                            ):
+        
+        # ref_seg format: '<expression><x1><y1><x2><y2><><><sep><><><><>'
+        # ignore <s> </s> and <pad>
+
+        text = text.replace('<s>', '')
+        text = text.replace('</s>', '')
+        text = text.replace('<pad>', '')
+
+        if allow_empty_phrase:
+            pattern = rf"(?:(?:<loc_\d+>|{re.escape(polygon_sep_token)}|{re.escape(polygon_start_token)}|{re.escape(polygon_end_token)}){{4,}})"
+        else:
+            # [^<]+: This part matches one or more characters that are not the < symbol. 
+            # The ^ inside the square brackets [] is a negation, meaning it matches anything except <.
+            #
+            pattern = rf"([^<]+(?:<loc_\d+>|{re.escape(polygon_sep_token)}|{re.escape(polygon_start_token)}|{re.escape(polygon_end_token)}){{4,}})"
+        phrases = re.findall(pattern, text)
+
+        phrase_string_pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_|<poly>)'
+        box_pattern =  rf'((?:<loc_\d+>)+)(?:{re.escape(polygon_sep_token)}|$)'
+
+        # one polygons instance is separated by polygon_start_token and polygon_end_token
+        polygons_instance_pattern = rf'{re.escape(polygon_start_token)}(.*?){re.escape(polygon_end_token)}'
+    
+        instances = []
+        for phrase_text in phrases:
+
+            # exclude loc_\d+>
+            # need to get span if want to include category score
+            phrase_text_strip = re.sub(r'^loc_\d+>', '', phrase_text, count=1)
+
+            # phrase = phrase.replace('<poly>', '')
+            # phrase = phrase.replace('poly>', '')
+
+            if phrase_text_strip == '' and not allow_empty_phrase:
+                continue
+
+
+            # parse phrase, get string 
+            phrase = re.search(phrase_string_pattern, phrase_text_strip)
+            if phrase is None:
+                continue
+            phrase = phrase.group()
+            # remove leading and trailing spaces
+            phrase = phrase.strip()
+
+            # parse bboxes by box_pattern
+
+            # split by polygon_start_token and polygon_end_token first using polygons_instance_pattern
+            if polygon_start_token in phrase_text and polygon_end_token in phrase_text:
+                polygons_instances_parsed = list(re.finditer(polygons_instance_pattern, phrase_text))
+            else:
+                polygons_instances_parsed = [phrase_text]
+
+            for _polygons_instances_parsed in polygons_instances_parsed:
+                # Prepare instance.
+                instance = {}
+
+                # polygons_parsed= list(re.finditer(box_pattern, phrase_text))
+                if isinstance(_polygons_instances_parsed, str): 
+                    polygons_parsed= list(re.finditer(box_pattern, _polygons_instances_parsed))
+                else:
+                    polygons_parsed= list(re.finditer(box_pattern, _polygons_instances_parsed.group(1)))
+                if len(polygons_parsed) == 0:
+                    continue
+
+                # a list of list (polygon)
+                bbox = []
+                polygons = []
+                for _polygon_parsed in polygons_parsed:
+                    # group 1: whole <loc_\d+>...</loc_\d+>
+                    _polygon = _polygon_parsed.group(1)
+                    # parse into list of int
+                    _polygon = [int(_loc_parsed.group(1)) for _loc_parsed in re.finditer(r'<loc_(\d+)>', _polygon)]
+                    if with_box_at_start and len(bbox) == 0:
+                        if len(_polygon) > 4:
+                            # no valid bbox prediction
+                            bbox = _polygon[:4]
+                            _polygon = _polygon[4:]
+                        else:
+                            bbox = [0, 0, 0, 0]
+                    # abandon last element if is not paired 
+                    if len(_polygon) % 2 == 1:
+                        _polygon = _polygon[:-1]
+                    
+                    # reshape into (n, 2)
+                    _polygon = self.coordinates_quantizer.dequantize(
+                        torch.tensor(np.array(_polygon).reshape(-1, 2)),
+                        size=image_size
+                    ).reshape(-1).tolist()
+                    # reshape back
+                    polygons.append(_polygon)
+
+                instance['cat_name'] = phrase
+                instance['polygons'] = polygons
+                if len(bbox) != 0:
+                    instance['bbox'] = self.box_quantizer.dequantize(
+                        boxes=torch.tensor([bbox]),
+                        size=image_size
+                    ).tolist()[0]  
+
+                instances.append(instance)
+
+        return instances
+
+    def __call__(
+        self,
+        text=None,
+        image_size=None,
+        parse_tasks=None,
+    ):
+        """
+        Args:
+            text: model outputs
+            image_size: (width, height)
+            parse_tasks: a list of tasks to parse, if None, parse all tasks.
+
+        """
+        if parse_tasks is not None:
+            if isinstance(parse_tasks, str):
+                parse_tasks = [parse_tasks]
+            for _parse_task in parse_tasks:
+                assert _parse_task in self.parse_tasks, f'parse task {_parse_task} not supported'
+        
+        # sequence or text should be provided 
+        assert text is not None, 'text should be provided'
+
+        parsed_dict = {
+            'text': text
+        }
+
+        for task in self.parse_tasks:
+            if parse_tasks is not None and task not in parse_tasks:
+                continue
+
+            pattern = self.parse_tasks_configs[task].get('PATTERN', None)
+
+            if task == 'ocr':
+                instances = self.parse_ocr_from_text_and_spans(
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                    area_threshold=self.parse_tasks_configs[task].get('AREA_THRESHOLD', 0.0),
+                )
+                parsed_dict['ocr'] = instances
+            elif task == 'phrase_grounding':
+                instances = self.parse_phrase_grounding_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                )
+                parsed_dict['phrase_grounding'] = instances
+            elif task == 'pure_text':
+                parsed_dict['pure_text'] = text 
+            elif task == 'description_with_bboxes':
+                instances = self.parse_description_with_bboxes_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                )
+                parsed_dict['description_with_bboxes'] = instances
+            elif task == 'description_with_polygons':
+                instances = self.parse_description_with_polygons_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                )
+                parsed_dict['description_with_polygons'] = instances
+            elif task == 'polygons':
+                instances = self.parse_description_with_polygons_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                    allow_empty_phrase=True,
+                )
+                parsed_dict['polygons'] = instances
+            elif task == 'bboxes':
+                instances = self.parse_description_with_bboxes_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                    allow_empty_phrase=True,
+                )
+                parsed_dict['bboxes'] = instances
+            elif task == 'description_with_bboxes_or_polygons':
+                if '<poly>' in text:
+                    # only support either polygons or bboxes, not both at the same time
+                    instances = self.parse_description_with_polygons_from_text_and_spans( 
+                        text,
+                        pattern=pattern,
+                        image_size=image_size,
+                    )
+                else:
+                    instances = self.parse_description_with_bboxes_from_text_and_spans( 
+                        text,
+                        pattern=pattern,
+                        image_size=image_size,
+                    )
+                parsed_dict['description_with_bboxes_or_polygons'] = instances
+            else:
+                raise ValueError("task {} is not supported".format(task))
+
+        return parsed_dict

processor_config.json

@@ -0,0 +1,6 @@
+{
+  "auto_map": {
+    "AutoProcessor": "processing_florence2.Florence2Processor"
+  },
+  "processor_class": "Florence2Processor"
+}

tokenizer.json

@@ -1,14 +1,7 @@
 {
   "version": "1.0",
   "truncation": null,
-  "padding": {
-    "strategy": "BatchLongest",
-    "direction": "Right",
-    "pad_to_multiple_of": null,
-    "pad_id": 1,
-    "pad_type_id": 0,
-    "pad_token": "<pad>"
-  },
+  "padding": null,
   "added_tokens": [
     {
       "id": 0,

tokenizer_config.json

@@ -9258,16 +9258,1047 @@
     "</proposal>",
     "<poly>",
     "</poly>",
+    "<and>",
+    "<od>",
+    "</od>",
+    "<ocr>",
+    "</ocr>",
+    "<loc_0>",
+    "<loc_1>",
+    "<loc_2>",
+    "<loc_3>",
+    "<loc_4>",
+    "<loc_5>",
+    "<loc_6>",
+    "<loc_7>",
+    "<loc_8>",
+    "<loc_9>",
+    "<loc_10>",
+    "<loc_11>",
+    "<loc_12>",
+    "<loc_13>",
+    "<loc_14>",
+    "<loc_15>",
+    "<loc_16>",
+    "<loc_17>",
+    "<loc_18>",
+    "<loc_19>",
+    "<loc_20>",
+    "<loc_21>",
+    "<loc_22>",
+    "<loc_23>",
+    "<loc_24>",
+    "<loc_25>",
+    "<loc_26>",
+    "<loc_27>",
+    "<loc_28>",
+    "<loc_29>",
+    "<loc_30>",
+    "<loc_31>",
+    "<loc_32>",
+    "<loc_33>",
+    "<loc_34>",
+    "<loc_35>",
+    "<loc_36>",
+    "<loc_37>",
+    "<loc_38>",
+    "<loc_39>",
+    "<loc_40>",
+    "<loc_41>",
+    "<loc_42>",
+    "<loc_43>",
+    "<loc_44>",
+    "<loc_45>",
+    "<loc_46>",
+    "<loc_47>",
+    "<loc_48>",
+    "<loc_49>",
+    "<loc_50>",
+    "<loc_51>",
+    "<loc_52>",
+    "<loc_53>",
+    "<loc_54>",
+    "<loc_55>",
+    "<loc_56>",
+    "<loc_57>",
+    "<loc_58>",
+    "<loc_59>",
+    "<loc_60>",
+    "<loc_61>",
+    "<loc_62>",
+    "<loc_63>",
+    "<loc_64>",
+    "<loc_65>",
+    "<loc_66>",
+    "<loc_67>",
+    "<loc_68>",
+    "<loc_69>",
+    "<loc_70>",
+    "<loc_71>",
+    "<loc_72>",
+    "<loc_73>",
+    "<loc_74>",
+    "<loc_75>",
+    "<loc_76>",
+    "<loc_77>",
+    "<loc_78>",
+    "<loc_79>",
+    "<loc_80>",
+    "<loc_81>",
+    "<loc_82>",
+    "<loc_83>",
+    "<loc_84>",
+    "<loc_85>",
+    "<loc_86>",
+    "<loc_87>",
+    "<loc_88>",
+    "<loc_89>",
+    "<loc_90>",
+    "<loc_91>",
+    "<loc_92>",
+    "<loc_93>",
+    "<loc_94>",
+    "<loc_95>",
+    "<loc_96>",
+    "<loc_97>",
+    "<loc_98>",
+    "<loc_99>",
+    "<loc_100>",
+    "<loc_101>",
+    "<loc_102>",
+    "<loc_103>",
+    "<loc_104>",
+    "<loc_105>",
+    "<loc_106>",
+    "<loc_107>",
+    "<loc_108>",
+    "<loc_109>",
+    "<loc_110>",
+    "<loc_111>",
+    "<loc_112>",
+    "<loc_113>",
+    "<loc_114>",
+    "<loc_115>",
+    "<loc_116>",
+    "<loc_117>",
+    "<loc_118>",
+    "<loc_119>",
+    "<loc_120>",
+    "<loc_121>",
+    "<loc_122>",
+    "<loc_123>",
+    "<loc_124>",
+    "<loc_125>",
+    "<loc_126>",
+    "<loc_127>",
+    "<loc_128>",
+    "<loc_129>",
+    "<loc_130>",
+    "<loc_131>",
+    "<loc_132>",
+    "<loc_133>",
+    "<loc_134>",
+    "<loc_135>",
+    "<loc_136>",
+    "<loc_137>",
+    "<loc_138>",
+    "<loc_139>",
+    "<loc_140>",
+    "<loc_141>",
+    "<loc_142>",
+    "<loc_143>",
+    "<loc_144>",
+    "<loc_145>",
+    "<loc_146>",
+    "<loc_147>",
+    "<loc_148>",
+    "<loc_149>",
+    "<loc_150>",
+    "<loc_151>",
+    "<loc_152>",
+    "<loc_153>",
+    "<loc_154>",
+    "<loc_155>",
+    "<loc_156>",
+    "<loc_157>",
+    "<loc_158>",
+    "<loc_159>",
+    "<loc_160>",
+    "<loc_161>",
+    "<loc_162>",
+    "<loc_163>",
+    "<loc_164>",
+    "<loc_165>",
+    "<loc_166>",
+    "<loc_167>",
+    "<loc_168>",
+    "<loc_169>",
+    "<loc_170>",
+    "<loc_171>",
+    "<loc_172>",
+    "<loc_173>",
+    "<loc_174>",
+    "<loc_175>",
+    "<loc_176>",
+    "<loc_177>",
+    "<loc_178>",
+    "<loc_179>",
+    "<loc_180>",
+    "<loc_181>",
+    "<loc_182>",
+    "<loc_183>",
+    "<loc_184>",
+    "<loc_185>",
+    "<loc_186>",
+    "<loc_187>",
+    "<loc_188>",
+    "<loc_189>",
+    "<loc_190>",
+    "<loc_191>",
+    "<loc_192>",
+    "<loc_193>",
+    "<loc_194>",
+    "<loc_195>",
+    "<loc_196>",
+    "<loc_197>",
+    "<loc_198>",
+    "<loc_199>",
+    "<loc_200>",
+    "<loc_201>",
+    "<loc_202>",
+    "<loc_203>",
+    "<loc_204>",
+    "<loc_205>",
+    "<loc_206>",
+    "<loc_207>",
+    "<loc_208>",
+    "<loc_209>",
+    "<loc_210>",
+    "<loc_211>",
+    "<loc_212>",
+    "<loc_213>",
+    "<loc_214>",
+    "<loc_215>",
+    "<loc_216>",
+    "<loc_217>",
+    "<loc_218>",
+    "<loc_219>",
+    "<loc_220>",
+    "<loc_221>",
+    "<loc_222>",
+    "<loc_223>",
+    "<loc_224>",
+    "<loc_225>",
+    "<loc_226>",
+    "<loc_227>",
+    "<loc_228>",
+    "<loc_229>",
+    "<loc_230>",
+    "<loc_231>",
+    "<loc_232>",
+    "<loc_233>",
+    "<loc_234>",
+    "<loc_235>",
+    "<loc_236>",
+    "<loc_237>",
+    "<loc_238>",
+    "<loc_239>",
+    "<loc_240>",
+    "<loc_241>",
+    "<loc_242>",
+    "<loc_243>",
+    "<loc_244>",
+    "<loc_245>",
+    "<loc_246>",
+    "<loc_247>",
+    "<loc_248>",
+    "<loc_249>",
+    "<loc_250>",
+    "<loc_251>",
+    "<loc_252>",
+    "<loc_253>",
+    "<loc_254>",
+    "<loc_255>",
+    "<loc_256>",
+    "<loc_257>",
+    "<loc_258>",
+    "<loc_259>",
+    "<loc_260>",
+    "<loc_261>",
+    "<loc_262>",
+    "<loc_263>",
+    "<loc_264>",
+    "<loc_265>",
+    "<loc_266>",
+    "<loc_267>",
+    "<loc_268>",
+    "<loc_269>",
+    "<loc_270>",
+    "<loc_271>",
+    "<loc_272>",
+    "<loc_273>",
+    "<loc_274>",
+    "<loc_275>",
+    "<loc_276>",
+    "<loc_277>",
+    "<loc_278>",
+    "<loc_279>",
+    "<loc_280>",
+    "<loc_281>",
+    "<loc_282>",
+    "<loc_283>",
+    "<loc_284>",
+    "<loc_285>",
+    "<loc_286>",
+    "<loc_287>",
+    "<loc_288>",
+    "<loc_289>",
+    "<loc_290>",
+    "<loc_291>",
+    "<loc_292>",
+    "<loc_293>",
+    "<loc_294>",
+    "<loc_295>",
+    "<loc_296>",
+    "<loc_297>",
+    "<loc_298>",
+    "<loc_299>",
+    "<loc_300>",
+    "<loc_301>",
+    "<loc_302>",
+    "<loc_303>",
+    "<loc_304>",
+    "<loc_305>",
+    "<loc_306>",
+    "<loc_307>",
+    "<loc_308>",
+    "<loc_309>",
+    "<loc_310>",
+    "<loc_311>",
+    "<loc_312>",
+    "<loc_313>",
+    "<loc_314>",
+    "<loc_315>",
+    "<loc_316>",
+    "<loc_317>",
+    "<loc_318>",
+    "<loc_319>",
+    "<loc_320>",
+    "<loc_321>",
+    "<loc_322>",
+    "<loc_323>",
+    "<loc_324>",
+    "<loc_325>",
+    "<loc_326>",
+    "<loc_327>",
+    "<loc_328>",
+    "<loc_329>",
+    "<loc_330>",
+    "<loc_331>",
+    "<loc_332>",
+    "<loc_333>",
+    "<loc_334>",
+    "<loc_335>",
+    "<loc_336>",
+    "<loc_337>",
+    "<loc_338>",
+    "<loc_339>",
+    "<loc_340>",
+    "<loc_341>",
+    "<loc_342>",
+    "<loc_343>",
+    "<loc_344>",
+    "<loc_345>",
+    "<loc_346>",
+    "<loc_347>",
+    "<loc_348>",
+    "<loc_349>",
+    "<loc_350>",
+    "<loc_351>",
+    "<loc_352>",
+    "<loc_353>",
+    "<loc_354>",
+    "<loc_355>",
+    "<loc_356>",
+    "<loc_357>",
+    "<loc_358>",
+    "<loc_359>",
+    "<loc_360>",
+    "<loc_361>",
+    "<loc_362>",
+    "<loc_363>",
+    "<loc_364>",
+    "<loc_365>",
+    "<loc_366>",
+    "<loc_367>",
+    "<loc_368>",
+    "<loc_369>",
+    "<loc_370>",
+    "<loc_371>",
+    "<loc_372>",
+    "<loc_373>",
+    "<loc_374>",
+    "<loc_375>",
+    "<loc_376>",
+    "<loc_377>",
+    "<loc_378>",
+    "<loc_379>",
+    "<loc_380>",
+    "<loc_381>",
+    "<loc_382>",
+    "<loc_383>",
+    "<loc_384>",
+    "<loc_385>",
+    "<loc_386>",
+    "<loc_387>",
+    "<loc_388>",
+    "<loc_389>",
+    "<loc_390>",
+    "<loc_391>",
+    "<loc_392>",
+    "<loc_393>",
+    "<loc_394>",
+    "<loc_395>",
+    "<loc_396>",
+    "<loc_397>",
+    "<loc_398>",
+    "<loc_399>",
+    "<loc_400>",
+    "<loc_401>",
+    "<loc_402>",
+    "<loc_403>",
+    "<loc_404>",
+    "<loc_405>",
+    "<loc_406>",
+    "<loc_407>",
+    "<loc_408>",
+    "<loc_409>",
+    "<loc_410>",
+    "<loc_411>",
+    "<loc_412>",
+    "<loc_413>",
+    "<loc_414>",
+    "<loc_415>",
+    "<loc_416>",
+    "<loc_417>",
+    "<loc_418>",
+    "<loc_419>",
+    "<loc_420>",
+    "<loc_421>",
+    "<loc_422>",
+    "<loc_423>",
+    "<loc_424>",
+    "<loc_425>",
+    "<loc_426>",
+    "<loc_427>",
+    "<loc_428>",
+    "<loc_429>",
+    "<loc_430>",
+    "<loc_431>",
+    "<loc_432>",
+    "<loc_433>",
+    "<loc_434>",
+    "<loc_435>",
+    "<loc_436>",
+    "<loc_437>",
+    "<loc_438>",
+    "<loc_439>",
+    "<loc_440>",
+    "<loc_441>",
+    "<loc_442>",
+    "<loc_443>",
+    "<loc_444>",
+    "<loc_445>",
+    "<loc_446>",
+    "<loc_447>",
+    "<loc_448>",
+    "<loc_449>",
+    "<loc_450>",
+    "<loc_451>",
+    "<loc_452>",
+    "<loc_453>",
+    "<loc_454>",
+    "<loc_455>",
+    "<loc_456>",
+    "<loc_457>",
+    "<loc_458>",
+    "<loc_459>",
+    "<loc_460>",
+    "<loc_461>",
+    "<loc_462>",
+    "<loc_463>",
+    "<loc_464>",
+    "<loc_465>",
+    "<loc_466>",
+    "<loc_467>",
+    "<loc_468>",
+    "<loc_469>",
+    "<loc_470>",
+    "<loc_471>",
+    "<loc_472>",
+    "<loc_473>",
+    "<loc_474>",
+    "<loc_475>",
+    "<loc_476>",
+    "<loc_477>",
+    "<loc_478>",
+    "<loc_479>",
+    "<loc_480>",
+    "<loc_481>",
+    "<loc_482>",
+    "<loc_483>",
+    "<loc_484>",
+    "<loc_485>",
+    "<loc_486>",
+    "<loc_487>",
+    "<loc_488>",
+    "<loc_489>",
+    "<loc_490>",
+    "<loc_491>",
+    "<loc_492>",
+    "<loc_493>",
+    "<loc_494>",
+    "<loc_495>",
+    "<loc_496>",
+    "<loc_497>",
+    "<loc_498>",
+    "<loc_499>",
+    "<loc_500>",
+    "<loc_501>",
+    "<loc_502>",
+    "<loc_503>",
+    "<loc_504>",
+    "<loc_505>",
+    "<loc_506>",
+    "<loc_507>",
+    "<loc_508>",
+    "<loc_509>",
+    "<loc_510>",
+    "<loc_511>",
+    "<loc_512>",
+    "<loc_513>",
+    "<loc_514>",
+    "<loc_515>",
+    "<loc_516>",
+    "<loc_517>",
+    "<loc_518>",
+    "<loc_519>",
+    "<loc_520>",
+    "<loc_521>",
+    "<loc_522>",
+    "<loc_523>",
+    "<loc_524>",
+    "<loc_525>",
+    "<loc_526>",
+    "<loc_527>",
+    "<loc_528>",
+    "<loc_529>",
+    "<loc_530>",
+    "<loc_531>",
+    "<loc_532>",
+    "<loc_533>",
+    "<loc_534>",
+    "<loc_535>",
+    "<loc_536>",
+    "<loc_537>",
+    "<loc_538>",
+    "<loc_539>",
+    "<loc_540>",
+    "<loc_541>",
+    "<loc_542>",
+    "<loc_543>",
+    "<loc_544>",
+    "<loc_545>",
+    "<loc_546>",
+    "<loc_547>",
+    "<loc_548>",
+    "<loc_549>",
+    "<loc_550>",
+    "<loc_551>",
+    "<loc_552>",
+    "<loc_553>",
+    "<loc_554>",
+    "<loc_555>",
+    "<loc_556>",
+    "<loc_557>",
+    "<loc_558>",
+    "<loc_559>",
+    "<loc_560>",
+    "<loc_561>",
+    "<loc_562>",
+    "<loc_563>",
+    "<loc_564>",
+    "<loc_565>",
+    "<loc_566>",
+    "<loc_567>",
+    "<loc_568>",
+    "<loc_569>",
+    "<loc_570>",
+    "<loc_571>",
+    "<loc_572>",
+    "<loc_573>",
+    "<loc_574>",
+    "<loc_575>",
+    "<loc_576>",
+    "<loc_577>",
+    "<loc_578>",
+    "<loc_579>",
+    "<loc_580>",
+    "<loc_581>",
+    "<loc_582>",
+    "<loc_583>",
+    "<loc_584>",
+    "<loc_585>",
+    "<loc_586>",
+    "<loc_587>",
+    "<loc_588>",
+    "<loc_589>",
+    "<loc_590>",
+    "<loc_591>",
+    "<loc_592>",
+    "<loc_593>",
+    "<loc_594>",
+    "<loc_595>",
+    "<loc_596>",
+    "<loc_597>",
+    "<loc_598>",
+    "<loc_599>",
+    "<loc_600>",
+    "<loc_601>",
+    "<loc_602>",
+    "<loc_603>",
+    "<loc_604>",
+    "<loc_605>",
+    "<loc_606>",
+    "<loc_607>",
+    "<loc_608>",
+    "<loc_609>",
+    "<loc_610>",
+    "<loc_611>",
+    "<loc_612>",
+    "<loc_613>",
+    "<loc_614>",
+    "<loc_615>",
+    "<loc_616>",
+    "<loc_617>",
+    "<loc_618>",
+    "<loc_619>",
+    "<loc_620>",
+    "<loc_621>",
+    "<loc_622>",
+    "<loc_623>",
+    "<loc_624>",
+    "<loc_625>",
+    "<loc_626>",
+    "<loc_627>",
+    "<loc_628>",
+    "<loc_629>",
+    "<loc_630>",
+    "<loc_631>",
+    "<loc_632>",
+    "<loc_633>",
+    "<loc_634>",
+    "<loc_635>",
+    "<loc_636>",
+    "<loc_637>",
+    "<loc_638>",
+    "<loc_639>",
+    "<loc_640>",
+    "<loc_641>",
+    "<loc_642>",
+    "<loc_643>",
+    "<loc_644>",
+    "<loc_645>",
+    "<loc_646>",
+    "<loc_647>",
+    "<loc_648>",
+    "<loc_649>",
+    "<loc_650>",
+    "<loc_651>",
+    "<loc_652>",
+    "<loc_653>",
+    "<loc_654>",
+    "<loc_655>",
+    "<loc_656>",
+    "<loc_657>",
+    "<loc_658>",
+    "<loc_659>",
+    "<loc_660>",
+    "<loc_661>",
+    "<loc_662>",
+    "<loc_663>",
+    "<loc_664>",
+    "<loc_665>",
+    "<loc_666>",
+    "<loc_667>",
+    "<loc_668>",
+    "<loc_669>",
+    "<loc_670>",
+    "<loc_671>",
+    "<loc_672>",
+    "<loc_673>",
+    "<loc_674>",
+    "<loc_675>",
+    "<loc_676>",
+    "<loc_677>",
+    "<loc_678>",
+    "<loc_679>",
+    "<loc_680>",
+    "<loc_681>",
+    "<loc_682>",
+    "<loc_683>",
+    "<loc_684>",
+    "<loc_685>",
+    "<loc_686>",
+    "<loc_687>",
+    "<loc_688>",
+    "<loc_689>",
+    "<loc_690>",
+    "<loc_691>",
+    "<loc_692>",
+    "<loc_693>",
+    "<loc_694>",
+    "<loc_695>",
+    "<loc_696>",
+    "<loc_697>",
+    "<loc_698>",
+    "<loc_699>",
+    "<loc_700>",
+    "<loc_701>",
+    "<loc_702>",
+    "<loc_703>",
+    "<loc_704>",
+    "<loc_705>",
+    "<loc_706>",
+    "<loc_707>",
+    "<loc_708>",
+    "<loc_709>",
+    "<loc_710>",
+    "<loc_711>",
+    "<loc_712>",
+    "<loc_713>",
+    "<loc_714>",
+    "<loc_715>",
+    "<loc_716>",
+    "<loc_717>",
+    "<loc_718>",
+    "<loc_719>",
+    "<loc_720>",
+    "<loc_721>",
+    "<loc_722>",
+    "<loc_723>",
+    "<loc_724>",
+    "<loc_725>",
+    "<loc_726>",
+    "<loc_727>",
+    "<loc_728>",
+    "<loc_729>",
+    "<loc_730>",
+    "<loc_731>",
+    "<loc_732>",
+    "<loc_733>",
+    "<loc_734>",
+    "<loc_735>",
+    "<loc_736>",
+    "<loc_737>",
+    "<loc_738>",
+    "<loc_739>",
+    "<loc_740>",
+    "<loc_741>",
+    "<loc_742>",
+    "<loc_743>",
+    "<loc_744>",
+    "<loc_745>",
+    "<loc_746>",
+    "<loc_747>",
+    "<loc_748>",
+    "<loc_749>",
+    "<loc_750>",
+    "<loc_751>",
+    "<loc_752>",
+    "<loc_753>",
+    "<loc_754>",
+    "<loc_755>",
+    "<loc_756>",
+    "<loc_757>",
+    "<loc_758>",
+    "<loc_759>",
+    "<loc_760>",
+    "<loc_761>",
+    "<loc_762>",
+    "<loc_763>",
+    "<loc_764>",
+    "<loc_765>",
+    "<loc_766>",
+    "<loc_767>",
+    "<loc_768>",
+    "<loc_769>",
+    "<loc_770>",
+    "<loc_771>",
+    "<loc_772>",
+    "<loc_773>",
+    "<loc_774>",
+    "<loc_775>",
+    "<loc_776>",
+    "<loc_777>",
+    "<loc_778>",
+    "<loc_779>",
+    "<loc_780>",
+    "<loc_781>",
+    "<loc_782>",
+    "<loc_783>",
+    "<loc_784>",
+    "<loc_785>",
+    "<loc_786>",
+    "<loc_787>",
+    "<loc_788>",
+    "<loc_789>",
+    "<loc_790>",
+    "<loc_791>",
+    "<loc_792>",
+    "<loc_793>",
+    "<loc_794>",
+    "<loc_795>",
+    "<loc_796>",
+    "<loc_797>",
+    "<loc_798>",
+    "<loc_799>",
+    "<loc_800>",
+    "<loc_801>",
+    "<loc_802>",
+    "<loc_803>",
+    "<loc_804>",
+    "<loc_805>",
+    "<loc_806>",
+    "<loc_807>",
+    "<loc_808>",
+    "<loc_809>",
+    "<loc_810>",
+    "<loc_811>",
+    "<loc_812>",
+    "<loc_813>",
+    "<loc_814>",
+    "<loc_815>",
+    "<loc_816>",
+    "<loc_817>",
+    "<loc_818>",
+    "<loc_819>",
+    "<loc_820>",
+    "<loc_821>",
+    "<loc_822>",
+    "<loc_823>",
+    "<loc_824>",
+    "<loc_825>",
+    "<loc_826>",
+    "<loc_827>",
+    "<loc_828>",
+    "<loc_829>",
+    "<loc_830>",
+    "<loc_831>",
+    "<loc_832>",
+    "<loc_833>",
+    "<loc_834>",
+    "<loc_835>",
+    "<loc_836>",
+    "<loc_837>",
+    "<loc_838>",
+    "<loc_839>",
+    "<loc_840>",
+    "<loc_841>",
+    "<loc_842>",
+    "<loc_843>",
+    "<loc_844>",
+    "<loc_845>",
+    "<loc_846>",
+    "<loc_847>",
+    "<loc_848>",
+    "<loc_849>",
+    "<loc_850>",
+    "<loc_851>",
+    "<loc_852>",
+    "<loc_853>",
+    "<loc_854>",
+    "<loc_855>",
+    "<loc_856>",
+    "<loc_857>",
+    "<loc_858>",
+    "<loc_859>",
+    "<loc_860>",
+    "<loc_861>",
+    "<loc_862>",
+    "<loc_863>",
+    "<loc_864>",
+    "<loc_865>",
+    "<loc_866>",
+    "<loc_867>",
+    "<loc_868>",
+    "<loc_869>",
+    "<loc_870>",
+    "<loc_871>",
+    "<loc_872>",
+    "<loc_873>",
+    "<loc_874>",
+    "<loc_875>",
+    "<loc_876>",
+    "<loc_877>",
+    "<loc_878>",
+    "<loc_879>",
+    "<loc_880>",
+    "<loc_881>",
+    "<loc_882>",
+    "<loc_883>",
+    "<loc_884>",
+    "<loc_885>",
+    "<loc_886>",
+    "<loc_887>",
+    "<loc_888>",
+    "<loc_889>",
+    "<loc_890>",
+    "<loc_891>",
+    "<loc_892>",
+    "<loc_893>",
+    "<loc_894>",
+    "<loc_895>",
+    "<loc_896>",
+    "<loc_897>",
+    "<loc_898>",
+    "<loc_899>",
+    "<loc_900>",
+    "<loc_901>",
+    "<loc_902>",
+    "<loc_903>",
+    "<loc_904>",
+    "<loc_905>",
+    "<loc_906>",
+    "<loc_907>",
+    "<loc_908>",
+    "<loc_909>",
+    "<loc_910>",
+    "<loc_911>",
+    "<loc_912>",
+    "<loc_913>",
+    "<loc_914>",
+    "<loc_915>",
+    "<loc_916>",
+    "<loc_917>",
+    "<loc_918>",
+    "<loc_919>",
+    "<loc_920>",
+    "<loc_921>",
+    "<loc_922>",
+    "<loc_923>",
+    "<loc_924>",
+    "<loc_925>",
+    "<loc_926>",
+    "<loc_927>",
+    "<loc_928>",
+    "<loc_929>",
+    "<loc_930>",
+    "<loc_931>",
+    "<loc_932>",
+    "<loc_933>",
+    "<loc_934>",
+    "<loc_935>",
+    "<loc_936>",
+    "<loc_937>",
+    "<loc_938>",
+    "<loc_939>",
+    "<loc_940>",
+    "<loc_941>",
+    "<loc_942>",
+    "<loc_943>",
+    "<loc_944>",
+    "<loc_945>",
+    "<loc_946>",
+    "<loc_947>",
+    "<loc_948>",
+    "<loc_949>",
+    "<loc_950>",
+    "<loc_951>",
+    "<loc_952>",
+    "<loc_953>",
+    "<loc_954>",
+    "<loc_955>",
+    "<loc_956>",
+    "<loc_957>",
+    "<loc_958>",
+    "<loc_959>",
+    "<loc_960>",
+    "<loc_961>",
+    "<loc_962>",
+    "<loc_963>",
+    "<loc_964>",
+    "<loc_965>",
+    "<loc_966>",
+    "<loc_967>",
+    "<loc_968>",
+    "<loc_969>",
+    "<loc_970>",
+    "<loc_971>",
+    "<loc_972>",
+    "<loc_973>",
+    "<loc_974>",
+    "<loc_975>",
+    "<loc_976>",
+    "<loc_977>",
+    "<loc_978>",
+    "<loc_979>",
+    "<loc_980>",
+    "<loc_981>",
+    "<loc_982>",
+    "<loc_983>",
+    "<loc_984>",
+    "<loc_985>",
+    "<loc_986>",
+    "<loc_987>",
+    "<loc_988>",
+    "<loc_989>",
+    "<loc_990>",
+    "<loc_991>",
+    "<loc_992>",
+    "<loc_993>",
+    "<loc_994>",
+    "<loc_995>",
+    "<loc_996>",
+    "<loc_997>",
+    "<loc_998>",
+    "<loc_999>",
+    "<cap>",
+    "</cap>",
+    "<ncap>",
+    "</ncap>",
+    "<dcap>",
+    "</dcap>",
+    "<grounding>",
+    "</grounding>",
+    "<seg>",
+    "</seg>",
+    "<sep>",
+    "<region_cap>",
+    "</region_cap>",
+    "<region_to_desciption>",
+    "</region_to_desciption>",
+    "<proposal>",
+    "</proposal>",
+    "<poly>",
+    "</poly>",
     "<and>"
   ],
+  "auto_map": {
+    "AutoProcessor": "processing_florence2.Florence2Processor"
+  },
   "bos_token": "<s>",
   "clean_up_tokenization_spaces": false,
   "cls_token": "<s>",
   "eos_token": "</s>",
   "errors": "replace",
   "mask_token": "<mask>",
+  "max_length": null,
   "model_max_length": 1024,
+  "pad_to_multiple_of": null,
   "pad_token": "<pad>",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
   "processor_class": "Florence2Processor",
   "sep_token": "</s>",
   "tokenizer_class": "BartTokenizer",