Upload processor

processing_taivisionlm.py

@@ -1,27 +1,330 @@
-"""
-Processor class for TaiVisionLM.
-"""
-import transformers
+# """
+# Processor class for TaiVisionLM.
+# """
+# import transformers
+# import logging
+# from typing import List, Optional, Union
+
+# 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 import (
+#     AddedToken,
+#     PaddingStrategy,
+#     PreTokenizedInput,
+#     TextInput,
+#     TruncationStrategy,  
+# )
+# from transformers.utils import TensorType
+# from .configuration_taivisionlm import TaiVisionLMConfig
+
+# logger = logging.getLogger(__name__)
+
+# IMAGE_TOKEN = "<image>"
+
+# # 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)
+
+# # Copied from transformers.models.paligemma.processing_paligemma._is_str_or_image
+# def _is_str_or_image(elem):
+#     return isinstance(elem, (str)) or is_image_or_image_url(elem)
+
+
+# def build_string_from_input(image_seq_len, image_token):
+#     """
+#     Builds a string from the input prompt and image tokens.
+#     For example, for the call:
+#     build_string_from_input(
+#         image_seq_len=3,
+#         image_token="<im>",
+#     )
+#     The output will be:
+#     "<im><im><im>"
+#     Args:
+#         image_seq_len (`int`): The length of the image sequence.
+#         image_token (`str`): The image token.
+#     """
+#     return f"{image_token * image_seq_len}"
+
+
+# class TaiVisionProcessor(ProcessorMixin):
+#     r"""
+#     Constructs a TraVision processor which wraps a SigLIP image processor and a GPT2 tokenizer into a single processor.
+
+#     [`TaiVisionProcessor`] offers all the functionalities of [`SiglipImageProcessor`] and [`LlamaTokenizerFast`]. See the
+#     [`~TaiVisionProcessor.__call__`] and [`~TaiVisionProcessor.decode`] for more information.
+
+#     Args:
+#         image_processor ([`SiglipImageProcessor`], *optional*):
+#             The image processor is a required input.
+#         tokenizer ([`LlamaTokenizerFast`], *optional*):
+#             The tokenizer is a required input.
+#         chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+#             in a chat into a tokenizable string.
+#     """
+
+#     attributes = ["image_processor", "tokenizer"]
+#     valid_kwargs = ["chat_template"]
+#     image_processor_class = "SiglipImageProcessor"
+#     tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
+
+#     def __init__(
+#         self,
+#         image_processor=None,
+#         tokenizer=None,
+#         chat_template=None,
+#         **kwargs,
+#     ):
+#         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
+
+#         image_token = AddedToken(IMAGE_TOKEN, normalized=False, special=True)
+#         tokens_to_add = {"additional_special_tokens": [image_token]}
+#         tokenizer.add_special_tokens(tokens_to_add)
+#         self.image_token_id = tokenizer.convert_tokens_to_ids(IMAGE_TOKEN)
+#         tokenizer.add_bos_token = False
+#         tokenizer.add_eos_token = False
+
+#         super().__init__(image_processor, tokenizer, chat_template=chat_template)
+        
+#     def __call__(
+#         self,
+#         prompts: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+#         images: ImageInput = None,
+#         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,
+#         labels: Optional[Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]] = 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 GPT2TokenizerFast's [`~GPT2TokenizerFast.__call__`] if `text` is not `None` to encode
+#         the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+#         SiglipImageProcessor's [`~SiglipImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
+#         of the above two methods for more information.
+
+#         The usage for TraVisionLM fine-tuning preparation follows a standard 4D causal mask where only the prompt and label tokens
+#         are attended in an auto-regressive manner. The label in `text` are to be passed separately to the __call__ function and 
+#         will be placed after the prompt, which is the instruction to steer the model generation. 
+
+#         Args:
+#             prompts (`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 `False`):
+#                 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.
+#             labels (`str`, `List[str]`, `List[List[str]]`):
+#                 The label or batch of labels to be encoded. Only necessary for training. 
+#             text (`str`, `List[str]`, `List[List[str]]`):
+#                 The text or batch of text to be encoded. If provided, the prompt and label should be 
+
+#         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 `label`
+#               is provided, the `input_ids` will also contain the label 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 `label` is not None
+#         """
+
+#         # return_token_type_ids = True if labels is not None else False
+#         return_token_type_ids = True
+
+#         if images is None:
+#             raise ValueError("`images` are expected as arguments to a `TraVisionProcessor` instance.")
+        
+#         images = [images] if not isinstance(images, list) else images
+        
+#         if prompts is None:
+#             logger.warning_once(
+#                 "You are using TaiVisionLM without a text prefix. It will perform as a picture-captioning model."
+#             )
+#             prompts = "描述這張圖片"  # default prompt if it is not provided as an argument
+#             if len(images) != 1:
+#                 prompts = [prompts] * len(images)
+                
+#         if isinstance(prompts, List) and isinstance(images, List):
+#             if len(images) < len(text):
+#                 raise ValueError(
+#                     f"Received {len(images)} images for {len(prompts)} prompts. Each prompt should be associated with an image."
+#                 )
+#         if _is_str_or_image(prompts):
+#             prompts = [prompts]
+#         elif isinstance(prompts, list) and _is_str_or_image(prompts[0]):
+#             pass
+        
+#         # add \n after image tokens
+#         prompts = [f"\n<|user|>\n{prompt}{self.tokenizer.eos_token}\n" for prompt in prompts]
+#         # TODO: tokenize the prompt twice, and check if the prompt is too long
+#         prompt_length = [len(self.tokenizer.tokenize(prompt)) + self.image_seq_length for prompt in prompts]
+        
+        
+#         if labels is not None:
+#             if _is_str_or_image(labels):
+#                 labels = [labels] # convert it to list if it is a string
+#             labels = [f"<|assistant|>\n{label}{self.tokenizer.eos_token}" for label in labels]
+
+#             text = [f"{prompt}{label}" for prompt, label in zip(prompts, labels)]
+            
+#         else:
+#             text = prompts
+        
+#         assert len(images) == len(text), "The number of images and text should be the same."
+        
+#         input_strings = [
+#             build_string_from_input(
+#                 image_seq_len=self.image_seq_length,
+#                 image_token=IMAGE_TOKEN,
+#             )            
+#             for _ in text
+#         ]
+        
+#         # this will do some image processing, like resizing, normalizing, etc.
+#         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
+
+#         # modify the token_type_ids here by text_pair ?
+#         inputs = self.tokenizer(
+#             input_strings,
+#             text_pair=labels,
+#             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}
+
+#         # we are doing training, so we need to return the labels
+#         if labels is not None:
+#             # fill the labels with -100 where we don't have to compute the loss
+#             # mask the padding part
+#             labels = inputs["input_ids"].masked_fill(inputs["attention_mask"] == 0, -100)
+#             # mask the image + prompt part, so that we don't train the model to predict the image tokens
+#             import torch
+#             prompt_length_tensor = torch.tensor(prompt_length)
+#             labels = labels.masked_fill(torch.arange(labels.size(1)).unsqueeze(0) < prompt_length_tensor.unsqueeze(1), -100)
+#             return_data.update({"labels": labels})
+            
+#         return BatchFeature(data=return_data)
+
+#     # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->GPT2
+#     def batch_decode(self, *args, **kwargs):
+#         """
+#         This method forwards all its arguments to GPT2TokenizerFast'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->GPT2
+#     def decode(self, *args, **kwargs):
+#         """
+#         This method forwards all its arguments to GPT2TokenizerFast'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->TraVision
+#     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))
+    
+
+
+
+
+
 import logging
 from typing import List, Optional, Union
 
 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 import (
+from transformers.tokenization_utils_base import (
     AddedToken,
     PaddingStrategy,
     PreTokenizedInput,
     TextInput,
-    TruncationStrategy,  
+    TruncationStrategy,
 )
 from transformers.utils import TensorType
-from .configuration_taivisionlm import TaiVisionLMConfig
+
 
 logger = logging.getLogger(__name__)
 
 IMAGE_TOKEN = "<image>"
 
+
 # Copied from transformers.models.idefics2.processing_idefics2.is_url
 def is_url(val) -> bool:
     return isinstance(val, str) and val.startswith("http")
@@ -31,34 +334,38 @@ def is_url(val) -> bool:
 def is_image_or_image_url(elem):
     return is_url(elem) or is_valid_image(elem)
 
-# Copied from transformers.models.paligemma.processing_paligemma._is_str_or_image
+
 def _is_str_or_image(elem):
     return isinstance(elem, (str)) or is_image_or_image_url(elem)
 
 
-def build_string_from_input(image_seq_len, image_token):
+def build_string_from_input(prompt, bos_token, image_seq_len, image_token):
     """
     Builds a string from the input prompt and image tokens.
     For example, for the call:
     build_string_from_input(
+        prompt="Prefix str"
+        bos_token="<s>",
         image_seq_len=3,
         image_token="<im>",
     )
     The output will be:
-    "<im><im><im>"
+    "<im><im><im><s>Initial str"
     Args:
+        prompt (`List[Union[str, ImageInput]]`): The input prompt.
+        bos_token (`str`): The beginning of sentence token.
         image_seq_len (`int`): The length of the image sequence.
         image_token (`str`): The image token.
     """
-    return f"{image_token * image_seq_len}"
+    return f"{image_token * image_seq_len}{bos_token}{prompt}\n"
 
 
 class TaiVisionProcessor(ProcessorMixin):
     r"""
-    Constructs a TraVision processor which wraps a SigLIP image processor and a GPT2 tokenizer into a single processor.
+    Constructs a PaliGemma processor which wraps a PaliGemma image processor and a PaliGemma tokenizer into a single processor.
 
-    [`TaiVisionProcessor`] offers all the functionalities of [`SiglipImageProcessor`] and [`LlamaTokenizerFast`]. See the
-    [`~TaiVisionProcessor.__call__`] and [`~TaiVisionProcessor.decode`] for more information.
+    [`PaliGemmaProcessor`] offers all the functionalities of [`SiglipImageProcessor`] and [`LlamaTokenizerFast`]. See the
+    [`~PaliGemmaProcessor.__call__`] and [`~PaliGemmaProcessor.decode`] for more information.
 
     Args:
         image_processor ([`SiglipImageProcessor`], *optional*):
@@ -98,11 +405,12 @@ class TaiVisionProcessor(ProcessorMixin):
         tokenizer.add_eos_token = False
 
         super().__init__(image_processor, tokenizer, chat_template=chat_template)
-        
+
     def __call__(
         self,
-        prompts: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        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,
@@ -120,30 +428,45 @@ class TaiVisionProcessor(ProcessorMixin):
         do_thumbnail: bool = None,
         do_align_long_axis: bool = None,
         do_rescale: bool = None,
-        labels: Optional[Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]] = None,
+        suffix: Optional[Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]]] = 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 GPT2TokenizerFast's [`~GPT2TokenizerFast.__call__`] if `text` is not `None` to encode
+        and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
         the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
         SiglipImageProcessor's [`~SiglipImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
         of the above two methods for more information.
 
-        The usage for TraVisionLM fine-tuning preparation follows a standard 4D causal mask where only the prompt and label tokens
-        are attended in an auto-regressive manner. The label in `text` are to be passed separately to the __call__ function and 
-        will be placed after the prompt, which is the instruction to steer the model generation. 
+        The usage for PaliGemma fine-tuning preparation is slightly different than usual. suffix passed are suffixes to
+        the prompt in `text`, and will be placed after the prompt. This is because attention is handled differently for
+        the prefix and the suffix. For instance,
+        ```python
+        image = PIL_cow_image
+        prompt = "answer en Where is the cow standing?"
+        suffix = "on the beach"
+        inputs = processor(text=prompt, images=image, suffix=suffix)
+        ```
+        Here `inputs` will contain the `input_ids` and `token_type_ids` that follow
+        ```python
+        inputs["input_ids"][:, 256:]
+        # tensor([[     2,   6006,    603,    573,  13910,   9980, 235336,    108,    477,   573,   8318]])
+        inputs["token_type_ids"][:, 256:]
+        tensor([[0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1]])
+        ```
+        Meaning the last three tokens are of "label" ("suffix") type while the other ones are of "prefix" type.
+
 
         Args:
-            prompts (`str`, `List[str]`, `List[List[str]]`):
+            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 `False`):
-                Adds a separately tokenized '\n' at the end of the prompt. 
+                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:
@@ -164,76 +487,56 @@ class TaiVisionProcessor(ProcessorMixin):
                 - `'pt'`: Return PyTorch `torch.Tensor` objects.
                 - `'np'`: Return NumPy `np.ndarray` objects.
                 - `'jax'`: Return JAX `jnp.ndarray` objects.
-            labels (`str`, `List[str]`, `List[List[str]]`):
-                The label or batch of labels to be encoded. Only necessary for training. 
-            text (`str`, `List[str]`, `List[List[str]]`):
-                The text or batch of text to be encoded. If provided, the prompt and label should be 
+            suffix (`str`, `List[str]`, `List[List[str]]`):
+                The suffixes or batch of suffixes to be encoded. Only necessary for finetuning. See https://github.com/google-research/big_vision/blob/main/big_vision/configs/proj/paligemma/README.md
+                for more information. If your prompt is "<image> What is on the image", the suffix corresponds to the expected prediction "a cow sitting on a bench".
 
         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 `label`
-              is provided, the `input_ids` will also contain the label input ids.
+            - **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 `label` is not None
+            - **labels** -- Labels compatible with training if `suffix` is not None
         """
 
-        # return_token_type_ids = True if labels is not None else False
-        return_token_type_ids = True
+        return_token_type_ids = True if suffix is not None else False
 
         if images is None:
-            raise ValueError("`images` are expected as arguments to a `TraVisionProcessor` instance.")
-        
-        images = [images] if not isinstance(images, list) else images
-        
-        if prompts is None:
+            raise ValueError("`images` are expected as arguments to a `PaliGemmaProcessor` instance.")
+        if text is None:
             logger.warning_once(
-                "You are using TaiVisionLM without a text prefix. It will perform as a picture-captioning model."
+                "You are using PaliGemma without a text prefix. It will perform as a picture-captioning model."
             )
-            prompts = "描述這張圖片"  # default prompt if it is not provided as an argument
-            if len(images) != 1:
-                prompts = [prompts] * len(images)
-                
-        if isinstance(prompts, List) and isinstance(images, List):
+            text = ""
+
+        if isinstance(text, List) and isinstance(images, List):
             if len(images) < len(text):
                 raise ValueError(
-                    f"Received {len(images)} images for {len(prompts)} prompts. Each prompt should be associated with an image."
+                    f"Received {len(images)} images for {len(text)} prompts. Each prompt should be associated with an image."
                 )
-        if _is_str_or_image(prompts):
-            prompts = [prompts]
-        elif isinstance(prompts, list) and _is_str_or_image(prompts[0]):
+        if _is_str_or_image(text):
+            text = [text]
+        elif isinstance(text, list) and _is_str_or_image(text[0]):
             pass
-        
-        # add \n after image tokens
-        prompts = [f"\n<|user|>\n{prompt}{self.tokenizer.eos_token}\n" for prompt in prompts]
-        # TODO: tokenize the prompt twice, and check if the prompt is too long
-        prompt_length = [len(self.tokenizer.tokenize(prompt)) + self.image_seq_length for prompt in prompts]
-        
-        
-        if labels is not None:
-            if _is_str_or_image(labels):
-                labels = [labels] # convert it to list if it is a string
-            labels = [f"<|assistant|>\n{label}{self.tokenizer.eos_token}" for label in labels]
+        if suffix is not None and _is_str_or_image(suffix):
+            suffix = [suffix]
+        if suffix is not None:
+            suffix = [sfx + self.tokenizer.eos_token for sfx in suffix]
 
-            text = [f"{prompt}{label}" for prompt, label in zip(prompts, labels)]
-            
-        else:
-            text = prompts
-        
-        assert len(images) == len(text), "The number of images and text should be the same."
-        
         input_strings = [
             build_string_from_input(
+                prompt=prompt,
+                bos_token=self.tokenizer.bos_token,
                 image_seq_len=self.image_seq_length,
                 image_token=IMAGE_TOKEN,
-            )            
-            for _ in text
+            )
+            for prompt in text
         ]
-        
-        # this will do some image processing, like resizing, normalizing, etc.
+
         pixel_values = self.image_processor(
             images,
             do_resize=do_resize,
@@ -250,10 +553,9 @@ class TaiVisionProcessor(ProcessorMixin):
         if max_length is not None:
             max_length += self.image_seq_length  # max_length has to account for the image tokens
 
-        # modify the token_type_ids here by text_pair ?
         inputs = self.tokenizer(
             input_strings,
-            text_pair=labels,
+            text_pair=suffix,
             return_tensors=return_tensors,
             padding=padding,
             max_length=max_length,
@@ -263,59 +565,58 @@ class TaiVisionProcessor(ProcessorMixin):
 
         return_data = {**inputs, "pixel_values": pixel_values}
 
-        # we are doing training, so we need to return the labels
-        if labels is not None:
-            # fill the labels with -100 where we don't have to compute the loss
-            # mask the padding part
-            labels = inputs["input_ids"].masked_fill(inputs["attention_mask"] == 0, -100)
-            # mask the image + prompt part, so that we don't train the model to predict the image tokens
-            import torch
-            prompt_length_tensor = torch.tensor(prompt_length)
-            labels = labels.masked_fill(torch.arange(labels.size(1)).unsqueeze(0) < prompt_length_tensor.unsqueeze(1), -100)
+        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->GPT2
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Gemma
     def batch_decode(self, *args, **kwargs):
         """
-        This method forwards all its arguments to GPT2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        This method forwards all its arguments to GemmaTokenizerFast'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->GPT2
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Gemma
     def decode(self, *args, **kwargs):
         """
-        This method forwards all its arguments to GPT2TokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        This method forwards all its arguments to GemmaTokenizerFast'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->TraVision
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names with CLIP->PaliGemma
     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))
     
-
-# if __name__ == '__main__':
-#     config = TaiVisionLMConfig.from_pretrained("./")
-#     preprocessor = transformers.SiglipImageProcessor.from_pretrained("google/siglip-base-patch16-224")
-#     preprocessor.image_seq_length = config.num_image_tokens
-#     tokenizer = transformers.AutoTokenizer.from_pretrained("benchang1110/Taiwan-tinyllama-v1.0-chat")
-#     processor = TaiVisionProcessor(tokenizer=tokenizer, image_processor=preprocessor)
-#     processor.save_pretrained("./")
     
-    # from PIL import Image
-    # import requests
-    # processor = TaiVisionProcessor.from_pretrained("./")
-    # url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg"
-    # image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
-    # prompt = "Hello< what is your name?"
-    # label = "I am fine, thank you."
-    # inputs = processor(prompts=prompt, labels=label,images=image, return_tensors="pt",padding="max_length",max_length=512)
-    # for key, value in inputs.items():
-    #     print(f"{key}: {value}")
-    # print(processor.decode(inputs.input_ids.tolist()[0]))
+    
+if __name__ == '__main__':
+    from configuration_taivisionlm import TaiVisionLMConfig
+    import transformers
+    import torch
+    config = TaiVisionLMConfig.from_pretrained("./")
+    preprocessor = transformers.SiglipImageProcessor.from_pretrained("google/siglip-base-patch16-224")
+    preprocessor.image_seq_length = config.num_image_tokens
+    tokenizer = transformers.AutoTokenizer.from_pretrained("benchang1110/Taiwan-tinyllama-v1.0-chat")
+    processor = TaiVisionProcessor(tokenizer=tokenizer, image_processor=preprocessor)
+    processor.save_pretrained("./")
+    
+    from PIL import Image
+    import requests
+    processor = TaiVisionProcessor.from_pretrained("./")
+    url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/car.jpg"
+    image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
+    text = "Hello< what is your name?"
+    suffix = "I am fine, thank you."
+    inputs = processor(text=text,suffix=suffix,images=image, return_tensors="pt",padding="max_length",max_length=512)
+    print(inputs['attention_mask'].shape)
+    print(inputs['input_ids'].shape)
+    print(inputs['token_type_ids'].shape)
+    # print number of 0 in token_type_ids
+    print(torch.sum(inputs['token_type_ids']==0))
+    print(inputs)