image_processor.py 17.7 KB
Newer Older
YiYi Xu's avatar
YiYi Xu committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# 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.

import warnings
16
from typing import List, Optional, Union
YiYi Xu's avatar
YiYi Xu committed
17
18
19
20
21
22
23

import numpy as np
import PIL
import torch
from PIL import Image

from .configuration_utils import ConfigMixin, register_to_config
24
from .utils import CONFIG_NAME, PIL_INTERPOLATION, deprecate
YiYi Xu's avatar
YiYi Xu committed
25
26


27
28
29
30
31
32
33
34
35
36
PipelineImageInput = Union[
    PIL.Image.Image,
    np.ndarray,
    torch.FloatTensor,
    List[PIL.Image.Image],
    List[np.ndarray],
    List[torch.FloatTensor],
]


YiYi Xu's avatar
YiYi Xu committed
37
38
class VaeImageProcessor(ConfigMixin):
    """
Steven Liu's avatar
Steven Liu committed
39
    Image processor for VAE.
YiYi Xu's avatar
YiYi Xu committed
40
41
42

    Args:
        do_resize (`bool`, *optional*, defaults to `True`):
43
            Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept
Steven Liu's avatar
Steven Liu committed
44
            `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method.
YiYi Xu's avatar
YiYi Xu committed
45
        vae_scale_factor (`int`, *optional*, defaults to `8`):
Steven Liu's avatar
Steven Liu committed
46
            VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
YiYi Xu's avatar
YiYi Xu committed
47
48
49
        resample (`str`, *optional*, defaults to `lanczos`):
            Resampling filter to use when resizing the image.
        do_normalize (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
50
            Whether to normalize the image to [-1,1].
51
52
        do_binarize (`bool`, *optional*, defaults to `True`):
            Whether to binarize the image to 0/1.
53
54
        do_convert_rgb (`bool`, *optional*, defaults to be `False`):
            Whether to convert the images to RGB format.
55
56
        do_convert_grayscale (`bool`, *optional*, defaults to be `False`):
            Whether to convert the images to grayscale format.
YiYi Xu's avatar
YiYi Xu committed
57
58
59
60
61
62
63
64
65
66
67
    """

    config_name = CONFIG_NAME

    @register_to_config
    def __init__(
        self,
        do_resize: bool = True,
        vae_scale_factor: int = 8,
        resample: str = "lanczos",
        do_normalize: bool = True,
68
        do_binarize: bool = False,
69
        do_convert_rgb: bool = False,
70
        do_convert_grayscale: bool = False,
YiYi Xu's avatar
YiYi Xu committed
71
72
    ):
        super().__init__()
73
74
75
76
77
78
79
        if do_convert_rgb and do_convert_grayscale:
            raise ValueError(
                "`do_convert_rgb` and `do_convert_grayscale` can not both be set to `True`,"
                " if you intended to convert the image into RGB format, please set `do_convert_grayscale = False`.",
                " if you intended to convert the image into grayscale format, please set `do_convert_rgb = False`",
            )
            self.config.do_convert_rgb = False
YiYi Xu's avatar
YiYi Xu committed
80
81

    @staticmethod
82
    def numpy_to_pil(images: np.ndarray) -> PIL.Image.Image:
YiYi Xu's avatar
YiYi Xu committed
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
        """
        Convert a numpy image or a batch of images to a PIL image.
        """
        if images.ndim == 3:
            images = images[None, ...]
        images = (images * 255).round().astype("uint8")
        if images.shape[-1] == 1:
            # special case for grayscale (single channel) images
            pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
        else:
            pil_images = [Image.fromarray(image) for image in images]

        return pil_images

    @staticmethod
98
99
    def pil_to_numpy(images: Union[List[PIL.Image.Image], PIL.Image.Image]) -> np.ndarray:
        """
Steven Liu's avatar
Steven Liu committed
100
        Convert a PIL image or a list of PIL images to NumPy arrays.
101
102
103
104
105
106
107
108
109
110
        """
        if not isinstance(images, list):
            images = [images]
        images = [np.array(image).astype(np.float32) / 255.0 for image in images]
        images = np.stack(images, axis=0)

        return images

    @staticmethod
    def numpy_to_pt(images: np.ndarray) -> torch.FloatTensor:
YiYi Xu's avatar
YiYi Xu committed
111
        """
Steven Liu's avatar
Steven Liu committed
112
        Convert a NumPy image to a PyTorch tensor.
YiYi Xu's avatar
YiYi Xu committed
113
114
115
116
117
118
119
120
        """
        if images.ndim == 3:
            images = images[..., None]

        images = torch.from_numpy(images.transpose(0, 3, 1, 2))
        return images

    @staticmethod
121
    def pt_to_numpy(images: torch.FloatTensor) -> np.ndarray:
YiYi Xu's avatar
YiYi Xu committed
122
        """
Steven Liu's avatar
Steven Liu committed
123
        Convert a PyTorch tensor to a NumPy image.
YiYi Xu's avatar
YiYi Xu committed
124
125
126
127
128
129
130
        """
        images = images.cpu().permute(0, 2, 3, 1).float().numpy()
        return images

    @staticmethod
    def normalize(images):
        """
Steven Liu's avatar
Steven Liu committed
131
        Normalize an image array to [-1,1].
YiYi Xu's avatar
YiYi Xu committed
132
133
134
        """
        return 2.0 * images - 1.0

135
136
137
    @staticmethod
    def denormalize(images):
        """
Steven Liu's avatar
Steven Liu committed
138
        Denormalize an image array to [0,1].
139
140
141
        """
        return (images / 2 + 0.5).clamp(0, 1)

142
143
144
    @staticmethod
    def convert_to_rgb(image: PIL.Image.Image) -> PIL.Image.Image:
        """
145
        Converts a PIL image to RGB format.
146
147
        """
        image = image.convert("RGB")
148

149
150
        return image

151
152
153
154
155
156
157
158
159
160
    @staticmethod
    def convert_to_grayscale(image: PIL.Image.Image) -> PIL.Image.Image:
        """
        Converts a PIL image to grayscale format.
        """
        image = image.convert("L")

        return image

    def get_default_height_width(
161
        self,
162
        image: [PIL.Image.Image, np.ndarray, torch.Tensor],
163
164
        height: Optional[int] = None,
        width: Optional[int] = None,
165
    ):
YiYi Xu's avatar
YiYi Xu committed
166
        """
167
168
169
170
171
172
173
174
175
176
177
178
        This function return the height and width that are downscaled to the next integer multiple of
        `vae_scale_factor`.

        Args:
            image(`PIL.Image.Image`, `np.ndarray` or `torch.Tensor`):
                The image input, can be a PIL image, numpy array or pytorch tensor. if it is a numpy array, should have
                shape `[batch, height, width]` or `[batch, height, width, channel]` if it is a pytorch tensor, should
                have shape `[batch, channel, height, width]`.
            height (`int`, *optional*, defaults to `None`):
                The height in preprocessed image. If `None`, will use the height of `image` input.
            width (`int`, *optional*`, defaults to `None`):
                The width in preprocessed. If `None`, will use the width of the `image` input.
YiYi Xu's avatar
YiYi Xu committed
179
        """
180

181
        if height is None:
182
183
184
185
186
187
188
            if isinstance(image, PIL.Image.Image):
                height = image.height
            elif isinstance(image, torch.Tensor):
                height = image.shape[2]
            else:
                height = image.shape[1]

189
        if width is None:
190
191
192
193
194
195
            if isinstance(image, PIL.Image.Image):
                width = image.width
            elif isinstance(image, torch.Tensor):
                width = image.shape[3]
            else:
                height = image.shape[2]
196
197
198
199

        width, height = (
            x - x % self.config.vae_scale_factor for x in (width, height)
        )  # resize to integer multiple of vae_scale_factor
200
201
202
203
204
205
206
207
208
209
210
211

        return height, width

    def resize(
        self,
        image: PIL.Image.Image,
        height: Optional[int] = None,
        width: Optional[int] = None,
    ) -> PIL.Image.Image:
        """
        Resize a PIL image.
        """
212
213
        image = image.resize((width, height), resample=PIL_INTERPOLATION[self.config.resample])
        return image
YiYi Xu's avatar
YiYi Xu committed
214

215
216
217
218
219
220
221
222
    def binarize(self, image: PIL.Image.Image) -> PIL.Image.Image:
        """
        create a mask
        """
        image[image < 0.5] = 0
        image[image >= 0.5] = 1
        return image

YiYi Xu's avatar
YiYi Xu committed
223
224
225
    def preprocess(
        self,
        image: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
226
227
        height: Optional[int] = None,
        width: Optional[int] = None,
YiYi Xu's avatar
YiYi Xu committed
228
229
    ) -> torch.Tensor:
        """
Steven Liu's avatar
Steven Liu committed
230
        Preprocess the image input. Accepted formats are PIL images, NumPy arrays or PyTorch tensors.
YiYi Xu's avatar
YiYi Xu committed
231
232
        """
        supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor)
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251

        # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image
        if self.config.do_convert_grayscale and isinstance(image, (torch.Tensor, np.ndarray)) and image.ndim == 3:
            if isinstance(image, torch.Tensor):
                # if image is a pytorch tensor could have 2 possible shapes:
                #    1. batch x height x width: we should insert the channel dimension at position 1
                #    2. channnel x height x width: we should insert batch dimension at position 0,
                #       however, since both channel and batch dimension has same size 1, it is same to insert at position 1
                #    for simplicity, we insert a dimension of size 1 at position 1 for both cases
                image = image.unsqueeze(1)
            else:
                # if it is a numpy array, it could have 2 possible shapes:
                #   1. batch x height x width: insert channel dimension on last position
                #   2. height x width x channel: insert batch dimension on first position
                if image.shape[-1] == 1:
                    image = np.expand_dims(image, axis=0)
                else:
                    image = np.expand_dims(image, axis=-1)

YiYi Xu's avatar
YiYi Xu committed
252
253
254
255
256
257
258
259
        if isinstance(image, supported_formats):
            image = [image]
        elif not (isinstance(image, list) and all(isinstance(i, supported_formats) for i in image)):
            raise ValueError(
                f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support {', '.join(supported_formats)}"
            )

        if isinstance(image[0], PIL.Image.Image):
260
261
            if self.config.do_convert_rgb:
                image = [self.convert_to_rgb(i) for i in image]
262
263
            elif self.config.do_convert_grayscale:
                image = [self.convert_to_grayscale(i) for i in image]
264
            if self.config.do_resize:
265
                height, width = self.get_default_height_width(image[0], height, width)
266
267
                image = [self.resize(i, height, width) for i in image]
            image = self.pil_to_numpy(image)  # to np
YiYi Xu's avatar
YiYi Xu committed
268
269
270
271
            image = self.numpy_to_pt(image)  # to pt

        elif isinstance(image[0], np.ndarray):
            image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
272

YiYi Xu's avatar
YiYi Xu committed
273
            image = self.numpy_to_pt(image)
274
275
276

            height, width = self.get_default_height_width(image, height, width)
            if self.config.do_resize and (image.shape[2] != height or image.shape[3] != width):
YiYi Xu's avatar
YiYi Xu committed
277
                raise ValueError(
278
279
                    f"Currently we only support resizing for PIL image - please resize your numpy array to be {height} and {width}"
                    f"currently the sizes are {image.shape[2]} and {image.shape[3]}. You can also pass a PIL image instead to use resize option in VAEImageProcessor"
YiYi Xu's avatar
YiYi Xu committed
280
281
282
283
                )

        elif isinstance(image[0], torch.Tensor):
            image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
284

285
286
287
288
            if self.config.do_convert_grayscale and image.ndim == 3:
                image = image.unsqueeze(1)

            channel = image.shape[1]
289
290
291
292
            # don't need any preprocess if the image is latents
            if channel == 4:
                return image

293
294
            height, width = self.get_default_height_width(image, height, width)
            if self.config.do_resize and (image.shape[2] != height or image.shape[3] != width):
YiYi Xu's avatar
YiYi Xu committed
295
                raise ValueError(
296
297
                    f"Currently we only support resizing for PIL image - please resize your torch tensor to be {height} and {width}"
                    f"currently the sizes are {image.shape[2]} and {image.shape[3]}. You can also pass a PIL image instead to use resize option in VAEImageProcessor"
YiYi Xu's avatar
YiYi Xu committed
298
299
300
                )

        # expected range [0,1], normalize to [-1,1]
301
        do_normalize = self.config.do_normalize
302
        if image.min() < 0 and do_normalize:
YiYi Xu's avatar
YiYi Xu committed
303
304
305
306
307
308
309
310
311
312
            warnings.warn(
                "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] "
                f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{image.min()},{image.max()}]",
                FutureWarning,
            )
            do_normalize = False

        if do_normalize:
            image = self.normalize(image)

313
314
315
        if self.config.do_binarize:
            image = self.binarize(image)

YiYi Xu's avatar
YiYi Xu committed
316
317
318
319
        return image

    def postprocess(
        self,
320
        image: torch.FloatTensor,
YiYi Xu's avatar
YiYi Xu committed
321
        output_type: str = "pil",
322
        do_denormalize: Optional[List[bool]] = None,
YiYi Xu's avatar
YiYi Xu committed
323
    ):
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
        if not isinstance(image, torch.Tensor):
            raise ValueError(
                f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor"
            )
        if output_type not in ["latent", "pt", "np", "pil"]:
            deprecation_message = (
                f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: "
                "`pil`, `np`, `pt`, `latent`"
            )
            deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False)
            output_type = "np"

        if output_type == "latent":
            return image

        if do_denormalize is None:
            do_denormalize = [self.config.do_normalize] * image.shape[0]

        image = torch.stack(
            [self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
        )

        if output_type == "pt":
YiYi Xu's avatar
YiYi Xu committed
347
348
349
350
351
352
            return image

        image = self.pt_to_numpy(image)

        if output_type == "np":
            return image
353
354

        if output_type == "pil":
YiYi Xu's avatar
YiYi Xu committed
355
            return self.numpy_to_pil(image)
estelleafl's avatar
estelleafl committed
356
357
358
359


class VaeImageProcessorLDM3D(VaeImageProcessor):
    """
Steven Liu's avatar
Steven Liu committed
360
    Image processor for VAE LDM3D.
estelleafl's avatar
estelleafl committed
361
362
363
364
365

    Args:
        do_resize (`bool`, *optional*, defaults to `True`):
            Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`.
        vae_scale_factor (`int`, *optional*, defaults to `8`):
Steven Liu's avatar
Steven Liu committed
366
            VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
estelleafl's avatar
estelleafl committed
367
368
369
        resample (`str`, *optional*, defaults to `lanczos`):
            Resampling filter to use when resizing the image.
        do_normalize (`bool`, *optional*, defaults to `True`):
Steven Liu's avatar
Steven Liu committed
370
            Whether to normalize the image to [-1,1].
estelleafl's avatar
estelleafl committed
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
    """

    config_name = CONFIG_NAME

    @register_to_config
    def __init__(
        self,
        do_resize: bool = True,
        vae_scale_factor: int = 8,
        resample: str = "lanczos",
        do_normalize: bool = True,
    ):
        super().__init__()

    @staticmethod
    def numpy_to_pil(images):
        """
Steven Liu's avatar
Steven Liu committed
388
        Convert a NumPy image or a batch of images to a PIL image.
estelleafl's avatar
estelleafl committed
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
        """
        if images.ndim == 3:
            images = images[None, ...]
        images = (images * 255).round().astype("uint8")
        if images.shape[-1] == 1:
            # special case for grayscale (single channel) images
            pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
        else:
            pil_images = [Image.fromarray(image[:, :, :3]) for image in images]

        return pil_images

    @staticmethod
    def rgblike_to_depthmap(image):
        """
        Args:
            image: RGB-like depth image

        Returns: depth map

        """
        return image[:, :, 1] * 2**8 + image[:, :, 2]

    def numpy_to_depth(self, images):
        """
Steven Liu's avatar
Steven Liu committed
414
        Convert a NumPy depth image or a batch of images to a PIL image.
estelleafl's avatar
estelleafl committed
415
416
417
        """
        if images.ndim == 3:
            images = images[None, ...]
418
419
420
421
422
423
424
425
426
        images_depth = images[:, :, :, 3:]
        if images.shape[-1] == 6:
            images_depth = (images_depth * 255).round().astype("uint8")
            pil_images = [
                Image.fromarray(self.rgblike_to_depthmap(image_depth), mode="I;16") for image_depth in images_depth
            ]
        elif images.shape[-1] == 4:
            images_depth = (images_depth * 65535.0).astype(np.uint16)
            pil_images = [Image.fromarray(image_depth, mode="I;16") for image_depth in images_depth]
estelleafl's avatar
estelleafl committed
427
        else:
428
            raise Exception("Not supported")
estelleafl's avatar
estelleafl committed
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459

        return pil_images

    def postprocess(
        self,
        image: torch.FloatTensor,
        output_type: str = "pil",
        do_denormalize: Optional[List[bool]] = None,
    ):
        if not isinstance(image, torch.Tensor):
            raise ValueError(
                f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor"
            )
        if output_type not in ["latent", "pt", "np", "pil"]:
            deprecation_message = (
                f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: "
                "`pil`, `np`, `pt`, `latent`"
            )
            deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False)
            output_type = "np"

        if do_denormalize is None:
            do_denormalize = [self.config.do_normalize] * image.shape[0]

        image = torch.stack(
            [self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
        )

        image = self.pt_to_numpy(image)

        if output_type == "np":
460
461
462
463
464
            if image.shape[-1] == 6:
                image_depth = np.stack([self.rgblike_to_depthmap(im[:, :, 3:]) for im in image], axis=0)
            else:
                image_depth = image[:, :, :, 3:]
            return image[:, :, :, :3], image_depth
estelleafl's avatar
estelleafl committed
465
466
467
468
469

        if output_type == "pil":
            return self.numpy_to_pil(image), self.numpy_to_depth(image)
        else:
            raise Exception(f"This type {output_type} is not supported")