File size: 6,542 Bytes
3ce0a28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f19d4b
3ce0a28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f19d4b
3ce0a28
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
# coding=utf-8
# Copyright 2024 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 MiniCPMV.
"""

from typing import List, Optional, Union
import torch
import re

from transformers.image_utils import ImageInput
from transformers.processing_utils import ProcessorMixin
from transformers.tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
from transformers.utils import TensorType

from .image_processing_minicpmv import MiniCPMVBatchFeature


class MiniCPMVProcessor(ProcessorMixin):
    r"""
    Constructs a MiniCPMV processor which wraps a MiniCPMV image processor and a MiniCPMV tokenizer into a single processor.

    [`MiniCPMVProcessor`] offers all the functionalities of [`MiniCPMVImageProcessor`] and [`LlamaTokenizerWrapper`]. See the
    [`~MiniCPMVProcessor.__call__`] and [`~MiniCPMVProcessor.decode`] for more information.

    Args:
        image_processor ([`MiniCPMVImageProcessor`], *optional*):
            The image processor is a required input.
        tokenizer ([`LlamaTokenizerWrapper`], *optional*):
            The tokenizer is a required input.
    """
    attributes = ["image_processor", "tokenizer"]
    image_processor_class = "AutoImageProcessor"
    tokenizer_class = "AutoTokenizer"

    def __init__(self, image_processor=None, tokenizer=None):
        super().__init__(image_processor, tokenizer)
    
    def __call__(
        self,
        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]],
        images: ImageInput = None,
        padding: Union[bool, str, PaddingStrategy] = False,
        max_length: Optional[int] = None,
        do_pad: Optional[bool] = True,
        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
    ) -> MiniCPMVBatchFeature:
        if images is not None:
            image_inputs = self.image_processor(images, do_pad=do_pad, return_tensors=return_tensors)
        return self._convert_images_texts_to_inputs(image_inputs, text, max_length=max_length, return_tensors=return_tensors)
    
    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
    def batch_decode(self, *args, **kwargs):
        """
        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
        refer to the docstring of this method for more information.
        """
        output_ids = args[0]
        result_text = []
        for result in output_ids:
            result = result[result != 0]
            if result[0] == self.tokenizer.bos_id:
                result = result[1:]
            if result[-1] == self.tokenizer.eos_id:
                result = result[:-1]
            result_text.append(self.tokenizer.decode(result, *args[1:], **kwargs).strip())
        return result_text
        # return self.tokenizer.batch_decode(*args, **kwargs)
    
    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
    def decode(self, *args, **kwargs):
        """
        This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
        the docstring of this method for more information.
        """
        result = args[0]
        result = result[result != 0]
        if result[0] == self.tokenizer.bos_id:
            result = result[1:]
        if result[-1] == self.tokenizer.eos_id:
            result = result[:-1]
        return self.tokenizer.decode(result, *args[1:], **kwargs).strip()

    def _convert(
        self, input_str, max_inp_length: Optional[int] = None
    ):
        if self.tokenizer.add_bos_token:
            input_ids = self.tokenizer.encode(input_str)
        else:
            input_ids = [self.tokenizer.bos_id] + self.tokenizer.encode(input_str)
        if max_inp_length is not None:
            input_ids = input_ids[:max_inp_length]
        input_ids = torch.tensor(input_ids, dtype=torch.int32)

        image_start_tokens = torch.where(input_ids == self.tokenizer.im_start_id)[0]
        image_start_tokens += 1
        image_end_tokens = torch.where(input_ids == self.tokenizer.im_end_id)[0]
        valid_image_nums = max(len(image_start_tokens), len(image_end_tokens))
        image_bounds = torch.hstack(
            [
                image_start_tokens[:valid_image_nums].unsqueeze(-1),
                image_end_tokens[:valid_image_nums].unsqueeze(-1),
            ]
        )
        return input_ids.unsqueeze(0), image_bounds

    def _convert_images_texts_to_inputs(self, images, texts, do_pad=False, truncation=None, max_length=None, return_tensors=None):
        if not len(images):
            model_inputs = self.tokenizer(texts, return_tensors=return_tensors, padding=do_pad, truncation=truncation, max_length=max_length)
            return MiniCPMVBatchFeature(data={**model_inputs})
        
        pattern = "(<image>./</image>)"
        images, image_sizes = images["pixel_values"], images["image_sizes"]

        image_tags = re.findall(pattern, texts)
        assert len(image_tags) <= 1 and len(image_sizes) == 1
        text_chunks = texts.split(pattern)
        final_texts = text_chunks[0] + self.image_processor.get_slice_image_placeholder(image_sizes[0]) \
            + text_chunks[1] + "<AI>"
        input_ids, image_bounds = self._convert(final_texts, max_length)
        
        return MiniCPMVBatchFeature(data={
            "input_ids": input_ids,
            "pixel_values": images,
            "image_sizes": [image_sizes],
            "image_bounds": [image_bounds]
        }, tensor_type=return_tensors)

    @property
    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
    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))