Upload processor
Browse files- image_processor.py +257 -0
- preprocessor_config.json +22 -0
image_processor.py
ADDED
@@ -0,0 +1,257 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, List, Optional, Tuple, Union, Iterable
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import transformers
|
6 |
+
from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
|
7 |
+
from transformers.image_transforms import (
|
8 |
+
ChannelDimension,
|
9 |
+
get_resize_output_image_size,
|
10 |
+
rescale,
|
11 |
+
resize,
|
12 |
+
to_channel_dimension_format,
|
13 |
+
)
|
14 |
+
from transformers.image_utils import (
|
15 |
+
ImageInput,
|
16 |
+
PILImageResampling,
|
17 |
+
infer_channel_dimension_format,
|
18 |
+
get_channel_dimension_axis,
|
19 |
+
make_list_of_images,
|
20 |
+
to_numpy_array,
|
21 |
+
valid_images,
|
22 |
+
)
|
23 |
+
from transformers.utils import is_torch_tensor
|
24 |
+
|
25 |
+
|
26 |
+
class FaceSegformerImageProcessor(BaseImageProcessor):
|
27 |
+
def __init__(self, **kwargs):
|
28 |
+
super().__init__(**kwargs)
|
29 |
+
self.image_size = kwargs.get("image_size", (224, 224))
|
30 |
+
self.normalize_mean = kwargs.get("normalize_mean", [0.485, 0.456, 0.406])
|
31 |
+
self.normalize_std = kwargs.get("normalize_std", [0.229, 0.224, 0.225])
|
32 |
+
self.resample = kwargs.get("resample", PILImageResampling.BILINEAR)
|
33 |
+
self.data_format = kwargs.get("data_format", ChannelDimension.FIRST)
|
34 |
+
|
35 |
+
@staticmethod
|
36 |
+
def normalize(
|
37 |
+
image: np.ndarray,
|
38 |
+
mean: Union[float, Iterable[float]],
|
39 |
+
std: Union[float, Iterable[float]],
|
40 |
+
max_pixel_value: float = 255.0,
|
41 |
+
data_format: Optional[ChannelDimension] = None,
|
42 |
+
input_data_format: Optional[Union[str, ChannelDimension]] = None,
|
43 |
+
) -> np.ndarray:
|
44 |
+
"""
|
45 |
+
Copied from:
|
46 |
+
https://github.com/huggingface/transformers/blob/3eddda1111f70f3a59485e08540e8262b927e867/src/transformers/image_transforms.py#L209
|
47 |
+
|
48 |
+
BUT uses the formula from albumentations:
|
49 |
+
https://albumentations.ai/docs/api_reference/augmentations/transforms/#albumentations.augmentations.transforms.Normalize
|
50 |
+
|
51 |
+
img = (img - mean * max_pixel_value) / (std * max_pixel_value)
|
52 |
+
"""
|
53 |
+
if not isinstance(image, np.ndarray):
|
54 |
+
raise ValueError("image must be a numpy array")
|
55 |
+
|
56 |
+
if input_data_format is None:
|
57 |
+
input_data_format = infer_channel_dimension_format(image)
|
58 |
+
channel_axis = get_channel_dimension_axis(
|
59 |
+
image, input_data_format=input_data_format
|
60 |
+
)
|
61 |
+
num_channels = image.shape[channel_axis]
|
62 |
+
|
63 |
+
# We cast to float32 to avoid errors that can occur when subtracting uint8 values.
|
64 |
+
# We preserve the original dtype if it is a float type to prevent upcasting float16.
|
65 |
+
if not np.issubdtype(image.dtype, np.floating):
|
66 |
+
image = image.astype(np.float32)
|
67 |
+
|
68 |
+
if isinstance(mean, Iterable):
|
69 |
+
if len(mean) != num_channels:
|
70 |
+
raise ValueError(
|
71 |
+
f"mean must have {num_channels} elements if it is an iterable, got {len(mean)}"
|
72 |
+
)
|
73 |
+
else:
|
74 |
+
mean = [mean] * num_channels
|
75 |
+
mean = np.array(mean, dtype=image.dtype)
|
76 |
+
|
77 |
+
if isinstance(std, Iterable):
|
78 |
+
if len(std) != num_channels:
|
79 |
+
raise ValueError(
|
80 |
+
f"std must have {num_channels} elements if it is an iterable, got {len(std)}"
|
81 |
+
)
|
82 |
+
else:
|
83 |
+
std = [std] * num_channels
|
84 |
+
std = np.array(std, dtype=image.dtype)
|
85 |
+
|
86 |
+
# Uses max_pixel_value for normalization
|
87 |
+
if input_data_format == ChannelDimension.LAST:
|
88 |
+
image = (image - mean * max_pixel_value) / (std * max_pixel_value)
|
89 |
+
else:
|
90 |
+
image = ((image.T - mean * max_pixel_value) / (std * max_pixel_value)).T
|
91 |
+
|
92 |
+
image = (
|
93 |
+
to_channel_dimension_format(image, data_format, input_data_format)
|
94 |
+
if data_format is not None
|
95 |
+
else image
|
96 |
+
)
|
97 |
+
return image
|
98 |
+
|
99 |
+
def resize(
|
100 |
+
self,
|
101 |
+
image: np.ndarray,
|
102 |
+
size: Dict[str, int],
|
103 |
+
resample: PILImageResampling = PILImageResampling.BICUBIC,
|
104 |
+
data_format: Optional[Union[str, ChannelDimension]] = None,
|
105 |
+
input_data_format: Optional[Union[str, ChannelDimension]] = None,
|
106 |
+
**kwargs,
|
107 |
+
) -> np.ndarray:
|
108 |
+
"""
|
109 |
+
Copied from:
|
110 |
+
https://github.com/huggingface/transformers/blob/3eddda1111f70f3a59485e08540e8262b927e867/src/transformers/models/mobilenet_v2/image_processing_mobilenet_v2.py
|
111 |
+
"""
|
112 |
+
default_to_square = True
|
113 |
+
if "shortest_edge" in size:
|
114 |
+
size = size["shortest_edge"]
|
115 |
+
default_to_square = False
|
116 |
+
elif "height" in size and "width" in size:
|
117 |
+
size = (size["height"], size["width"])
|
118 |
+
else:
|
119 |
+
raise ValueError(
|
120 |
+
"Size must contain either 'shortest_edge' or 'height' and 'width'."
|
121 |
+
)
|
122 |
+
|
123 |
+
output_size = get_resize_output_image_size(
|
124 |
+
image,
|
125 |
+
size=size,
|
126 |
+
default_to_square=default_to_square,
|
127 |
+
input_data_format=input_data_format,
|
128 |
+
)
|
129 |
+
return resize(
|
130 |
+
image,
|
131 |
+
size=output_size,
|
132 |
+
resample=resample,
|
133 |
+
data_format=data_format,
|
134 |
+
input_data_format=input_data_format,
|
135 |
+
**kwargs,
|
136 |
+
)
|
137 |
+
|
138 |
+
def __call__(self, images: ImageInput, masks: ImageInput = None, **kwargs):
|
139 |
+
"""
|
140 |
+
Adapted from:
|
141 |
+
https://github.com/huggingface/transformers/blob/3eddda1111f70f3a59485e08540e8262b927e867/src/transformers/models/mobilenet_v2/image_processing_mobilenet_v2.py
|
142 |
+
"""
|
143 |
+
# single to iterable if needed
|
144 |
+
images = make_list_of_images(images)
|
145 |
+
|
146 |
+
# validate
|
147 |
+
if not valid_images(images):
|
148 |
+
raise ValueError(
|
149 |
+
"Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
|
150 |
+
"torch.Tensor, tf.Tensor or jax.ndarray."
|
151 |
+
)
|
152 |
+
|
153 |
+
# make numpy arrays
|
154 |
+
images = [to_numpy_array(image) for image in images]
|
155 |
+
|
156 |
+
# get channel dimensions
|
157 |
+
input_data_format = kwargs.get("input_data_format")
|
158 |
+
if input_data_format is None:
|
159 |
+
# We assume that all images have the same channel dimension format.
|
160 |
+
input_data_format = infer_channel_dimension_format(images[0])
|
161 |
+
|
162 |
+
# check if training
|
163 |
+
# todo: can also assume if masks are passed that we are doing training?
|
164 |
+
if kwargs.get("do_training", False) is True:
|
165 |
+
if mask is None:
|
166 |
+
raise ValueError("must pass masks if doing training.")
|
167 |
+
# todo: implement this soon.
|
168 |
+
raise NotImplementedError("not yet implemented.")
|
169 |
+
# Assume we want to do all transformations for training
|
170 |
+
else:
|
171 |
+
# do transformations for inference...
|
172 |
+
images = [
|
173 |
+
self.resize(
|
174 |
+
image=image,
|
175 |
+
size={
|
176 |
+
"shortest_edge": min(
|
177 |
+
kwargs.get("image_size") or self.image_size
|
178 |
+
)
|
179 |
+
},
|
180 |
+
resample=kwargs.get("resample") or self.resample,
|
181 |
+
input_data_format=input_data_format,
|
182 |
+
)
|
183 |
+
for image in images
|
184 |
+
]
|
185 |
+
images = [
|
186 |
+
self.normalize(
|
187 |
+
image=image,
|
188 |
+
mean=kwargs.get("normalize_mean") or self.normalize_mean,
|
189 |
+
std=kwargs.get("normalize_std") or self.normalize_std,
|
190 |
+
input_data_format=input_data_format,
|
191 |
+
)
|
192 |
+
for image in images
|
193 |
+
]
|
194 |
+
# fix dimensions
|
195 |
+
images = [
|
196 |
+
to_channel_dimension_format(
|
197 |
+
image,
|
198 |
+
kwargs.get("data_format") or self.data_format,
|
199 |
+
input_channel_dim=input_data_format,
|
200 |
+
)
|
201 |
+
for image in images
|
202 |
+
]
|
203 |
+
|
204 |
+
data = {"pixel_values": images}
|
205 |
+
return BatchFeature(data=data, tensor_type="pt")
|
206 |
+
|
207 |
+
# Copied from transformers.models.segformer.image_processing_segformer.SegformerImageProcessor.post_process_semantic_segmentation
|
208 |
+
def post_process_semantic_segmentation(
|
209 |
+
self, outputs, target_sizes: List[Tuple] = None
|
210 |
+
):
|
211 |
+
"""
|
212 |
+
Converts the output of [`SegformerForSemanticSegmentation`] into semantic segmentation maps. Only supports PyTorch.
|
213 |
+
|
214 |
+
Args:
|
215 |
+
outputs ([`SegformerForSemanticSegmentation`]):
|
216 |
+
Raw outputs of the model.
|
217 |
+
target_sizes (`List[Tuple]` of length `batch_size`, *optional*):
|
218 |
+
List of tuples corresponding to the requested final size (height, width) of each prediction. If unset,
|
219 |
+
predictions will not be resized.
|
220 |
+
|
221 |
+
Returns:
|
222 |
+
semantic_segmentation: `List[torch.Tensor]` of length `batch_size`, where each item is a semantic
|
223 |
+
segmentation map of shape (height, width) corresponding to the target_sizes entry (if `target_sizes` is
|
224 |
+
specified). Each entry of each `torch.Tensor` correspond to a semantic class id.
|
225 |
+
"""
|
226 |
+
# TODO: add support for other frameworks
|
227 |
+
logits = outputs.logits
|
228 |
+
|
229 |
+
# Resize logits and compute semantic segmentation maps
|
230 |
+
if target_sizes is not None:
|
231 |
+
if len(logits) != len(target_sizes):
|
232 |
+
raise ValueError(
|
233 |
+
"Make sure that you pass in as many target sizes as the batch dimension of the logits"
|
234 |
+
)
|
235 |
+
|
236 |
+
if is_torch_tensor(target_sizes):
|
237 |
+
target_sizes = target_sizes.numpy()
|
238 |
+
|
239 |
+
semantic_segmentation = []
|
240 |
+
|
241 |
+
for idx in range(len(logits)):
|
242 |
+
resized_logits = torch.nn.functional.interpolate(
|
243 |
+
logits[idx].unsqueeze(dim=0),
|
244 |
+
size=target_sizes[idx],
|
245 |
+
mode="bilinear",
|
246 |
+
align_corners=False,
|
247 |
+
)
|
248 |
+
semantic_map = resized_logits[0].argmax(dim=0)
|
249 |
+
semantic_segmentation.append(semantic_map)
|
250 |
+
else:
|
251 |
+
semantic_segmentation = logits.argmax(dim=1)
|
252 |
+
semantic_segmentation = [
|
253 |
+
semantic_segmentation[i] for i in range(semantic_segmentation.shape[0])
|
254 |
+
]
|
255 |
+
|
256 |
+
return semantic_segmentation
|
257 |
+
|
preprocessor_config.json
ADDED
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"auto_map": {
|
3 |
+
"AutoImageProcessor": "image_processor.FaceSegformerImageProcessor"
|
4 |
+
},
|
5 |
+
"data_format": "channels_first",
|
6 |
+
"image_processor_type": "FaceSegformerImageProcessor",
|
7 |
+
"image_size": [
|
8 |
+
224,
|
9 |
+
224
|
10 |
+
],
|
11 |
+
"normalize_mean": [
|
12 |
+
0.485,
|
13 |
+
0.456,
|
14 |
+
0.406
|
15 |
+
],
|
16 |
+
"normalize_std": [
|
17 |
+
0.229,
|
18 |
+
0.224,
|
19 |
+
0.225
|
20 |
+
],
|
21 |
+
"resample": 2
|
22 |
+
}
|