Spaces:
Build error
Build error
File size: 6,569 Bytes
b6af722 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 |
# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0
#
# 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.
from typing import Any, Optional
import torch
import torchvision.transforms.functional as transforms_F
from PIL import Image
def obtain_image_size(data_dict: dict, input_keys: list) -> tuple[int, int]:
r"""Function for obtaining the image size from the data dict.
Args:
data_dict (dict): Input data dict
input_keys (list): List of input keys
Returns:
width (int): Width of the input image
height (int): Height of the input image
"""
data1 = data_dict[input_keys[0]]
if isinstance(data1, Image.Image):
width, height = data1.size
elif isinstance(data1, torch.Tensor):
height, width = data1.size()[-2:]
else:
raise ValueError("data to random crop should be PIL Image or tensor")
return width, height
class Augmentor:
def __init__(self, input_keys: list, output_keys: Optional[list] = None, args: Optional[dict] = None) -> None:
r"""Base augmentor class
Args:
input_keys (list): List of input keys
output_keys (list): List of output keys
args (dict): Arguments associated with the augmentation
"""
self.input_keys = input_keys
self.output_keys = output_keys
self.args = args
def __call__(self, *args: Any, **kwds: Any) -> Any:
raise ValueError("Augmentor not implemented")
class ResizeSmallestSideAspectPreserving(Augmentor):
def __init__(self, input_keys: list, output_keys: Optional[list] = None, args: Optional[dict] = None) -> None:
super().__init__(input_keys, output_keys, args)
def __call__(self, data_dict: dict) -> dict:
r"""Performs aspect-ratio preserving resizing.
Image is resized to the dimension which has the smaller ratio of (size / target_size).
First we compute (w_img / w_target) and (h_img / h_target) and resize the image
to the dimension that has the smaller of these ratios.
Args:
data_dict (dict): Input data dict
Returns:
data_dict (dict): Output dict where images are resized
"""
if self.output_keys is None:
self.output_keys = self.input_keys
assert self.args is not None, "Please specify args in augmentations"
img_w, img_h = self.args["img_w"], self.args["img_h"]
orig_w, orig_h = obtain_image_size(data_dict, self.input_keys)
scaling_ratio = max((img_w / orig_w), (img_h / orig_h))
target_size = (int(scaling_ratio * orig_h + 0.5), int(scaling_ratio * orig_w + 0.5))
assert (
target_size[0] >= img_h and target_size[1] >= img_w
), f"Resize error. orig {(orig_w, orig_h)} desire {(img_w, img_h)} compute {target_size}"
for inp_key, out_key in zip(self.input_keys, self.output_keys):
data_dict[out_key] = transforms_F.resize(
data_dict[inp_key],
size=target_size, # type: ignore
interpolation=getattr(self.args, "interpolation", transforms_F.InterpolationMode.BICUBIC),
antialias=True,
)
if out_key != inp_key:
del data_dict[inp_key]
return data_dict
class CenterCrop(Augmentor):
def __init__(self, input_keys: list, output_keys: Optional[list] = None, args: Optional[dict] = None) -> None:
super().__init__(input_keys, output_keys, args)
def __call__(self, data_dict: dict) -> dict:
r"""Performs center crop.
Args:
data_dict (dict): Input data dict
Returns:
data_dict (dict): Output dict where images are center cropped.
We also save the cropping parameters in the aug_params dict
so that it will be used by other transforms.
"""
assert (
(self.args is not None) and ("img_w" in self.args) and ("img_h" in self.args)
), "Please specify size in args"
img_w, img_h = self.args["img_w"], self.args["img_h"]
orig_w, orig_h = obtain_image_size(data_dict, self.input_keys)
for key in self.input_keys:
data_dict[key] = transforms_F.center_crop(data_dict[key], [img_h, img_w])
# We also add the aug params we use. This will be useful for other transforms
crop_x0 = (orig_w - img_w) // 2
crop_y0 = (orig_h - img_h) // 2
cropping_params = {
"resize_w": orig_w,
"resize_h": orig_h,
"crop_x0": crop_x0,
"crop_y0": crop_y0,
"crop_w": img_w,
"crop_h": img_h,
}
if "aug_params" not in data_dict:
data_dict["aug_params"] = dict()
data_dict["aug_params"]["cropping"] = cropping_params
data_dict["padding_mask"] = torch.zeros((1, cropping_params["crop_h"], cropping_params["crop_w"]))
return data_dict
class Normalize(Augmentor):
def __init__(self, input_keys: list, output_keys: Optional[list] = None, args: Optional[dict] = None) -> None:
super().__init__(input_keys, output_keys, args)
def __call__(self, data_dict: dict) -> dict:
r"""Performs data normalization.
Args:
data_dict (dict): Input data dict
Returns:
data_dict (dict): Output dict where images are center cropped.
"""
assert self.args is not None, "Please specify args"
mean = self.args["mean"]
std = self.args["std"]
for key in self.input_keys:
if isinstance(data_dict[key], torch.Tensor):
data_dict[key] = data_dict[key].to(dtype=torch.get_default_dtype()).div(255)
else:
data_dict[key] = transforms_F.to_tensor(data_dict[key]) # division by 255 is applied in to_tensor()
data_dict[key] = transforms_F.normalize(tensor=data_dict[key], mean=mean, std=std)
return data_dict
|