File size: 6,323 Bytes
3b96cb1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
# Copyright (c) OpenMMLab. All rights reserved.
from pathlib import Path
from typing import Callable, List, Optional

import numpy as np
from mmcv.image import imread
from mmengine.config import Config
from mmengine.dataset import Compose, default_collate

from mmpretrain.registry import TRANSFORMS
from mmpretrain.structures import DataSample
from .base import BaseInferencer, InputType
from .model import list_models


class ImageCaptionInferencer(BaseInferencer):
    """The inferencer for image caption.

    Args:
        model (BaseModel | str | Config): A model name or a path to the config
            file, or a :obj:`BaseModel` object. The model name can be found
            by ``ImageCaptionInferencer.list_models()`` and you can also
            query it in :doc:`/modelzoo_statistics`.
        pretrained (str, optional): Path to the checkpoint. If None, it will
            try to find a pre-defined weight from the model you specified
            (only work if the ``model`` is a model name). Defaults to None.
        device (str, optional): Device to run inference. If None, the available
            device will be automatically used. Defaults to None.
        **kwargs: Other keyword arguments to initialize the model (only work if
            the ``model`` is a model name).

    Example:
        >>> from mmpretrain import ImageCaptionInferencer
        >>> inferencer = ImageCaptionInferencer('blip-base_3rdparty_caption')
        >>> inferencer('demo/cat-dog.png')[0]
        {'pred_caption': 'a puppy and a cat sitting on a blanket'}
    """  # noqa: E501

    visualize_kwargs: set = {'resize', 'show', 'show_dir', 'wait_time'}

    def __call__(self,
                 images: InputType,
                 return_datasamples: bool = False,
                 batch_size: int = 1,
                 **kwargs) -> dict:
        """Call the inferencer.

        Args:
            images (str | array | list): The image path or array, or a list of
                images.
            return_datasamples (bool): Whether to return results as
                :obj:`DataSample`. Defaults to False.
            batch_size (int): Batch size. Defaults to 1.
            resize (int, optional): Resize the short edge of the image to the
                specified length before visualization. Defaults to None.
            draw_score (bool): Whether to draw the prediction scores
                of prediction categories. Defaults to True.
            show (bool): Whether to display the visualization result in a
                window. Defaults to False.
            wait_time (float): The display time (s). Defaults to 0, which means
                "forever".
            show_dir (str, optional): If not None, save the visualization
                results in the specified directory. Defaults to None.

        Returns:
            list: The inference results.
        """
        return super().__call__(images, return_datasamples, batch_size,
                                **kwargs)

    def _init_pipeline(self, cfg: Config) -> Callable:
        test_pipeline_cfg = cfg.test_dataloader.dataset.pipeline
        from mmpretrain.datasets import remove_transform

        # Image loading is finished in `self.preprocess`.
        test_pipeline_cfg = remove_transform(test_pipeline_cfg,
                                             'LoadImageFromFile')
        test_pipeline = Compose(
            [TRANSFORMS.build(t) for t in test_pipeline_cfg])
        return test_pipeline

    def preprocess(self, inputs: List[InputType], batch_size: int = 1):

        def load_image(input_):
            img = imread(input_)
            if img is None:
                raise ValueError(f'Failed to read image {input_}.')
            return dict(
                img=img,
                img_shape=img.shape[:2],
                ori_shape=img.shape[:2],
            )

        pipeline = Compose([load_image, self.pipeline])

        chunked_data = self._get_chunk_data(map(pipeline, inputs), batch_size)
        yield from map(default_collate, chunked_data)

    def visualize(self,
                  ori_inputs: List[InputType],
                  preds: List[DataSample],
                  show: bool = False,
                  wait_time: int = 0,
                  resize: Optional[int] = None,
                  show_dir=None):
        if not show and show_dir is None:
            return None

        if self.visualizer is None:
            from mmpretrain.visualization import UniversalVisualizer
            self.visualizer = UniversalVisualizer()

        visualization = []
        for i, (input_, data_sample) in enumerate(zip(ori_inputs, preds)):
            image = imread(input_)
            if isinstance(input_, str):
                # The image loaded from path is BGR format.
                image = image[..., ::-1]
                name = Path(input_).stem
            else:
                name = str(i)

            if show_dir is not None:
                show_dir = Path(show_dir)
                show_dir.mkdir(exist_ok=True)
                out_file = str((show_dir / name).with_suffix('.png'))
            else:
                out_file = None

            self.visualizer.visualize_image_caption(
                image,
                data_sample,
                resize=resize,
                show=show,
                wait_time=wait_time,
                name=name,
                out_file=out_file)
            visualization.append(self.visualizer.get_image())
        if show:
            self.visualizer.close()
        return visualization

    def postprocess(self,
                    preds: List[DataSample],
                    visualization: List[np.ndarray],
                    return_datasamples=False) -> dict:
        if return_datasamples:
            return preds

        results = []
        for data_sample in preds:
            results.append({'pred_caption': data_sample.get('pred_caption')})

        return results

    @staticmethod
    def list_models(pattern: Optional[str] = None):
        """List all available model names.

        Args:
            pattern (str | None): A wildcard pattern to match model names.

        Returns:
            List[str]: a list of model names.
        """
        return list_models(pattern=pattern, task='Image Caption')