Spaces:
Sleeping
Sleeping
File size: 4,779 Bytes
28c6826 |
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 |
import warnings
from dataclasses import dataclass, asdict
from typing import Any, Dict, Optional, Sequence, Tuple, Union
import torch
import torch.nn as nn
import torchvision.transforms.functional as F
from torchvision.transforms import Normalize, Compose, RandomResizedCrop, InterpolationMode, ToTensor, Resize, \
CenterCrop
from .constants import OPENAI_DATASET_MEAN, OPENAI_DATASET_STD
@dataclass
class AugmentationCfg:
scale: Tuple[float, float] = (0.9, 1.0)
ratio: Optional[Tuple[float, float]] = None
color_jitter: Optional[Union[float, Tuple[float, float, float]]] = None
interpolation: Optional[str] = None
re_prob: Optional[float] = None
re_count: Optional[int] = None
use_timm: bool = False
class ResizeMaxSize(nn.Module):
def __init__(self, max_size, interpolation=InterpolationMode.BICUBIC, fn='max', fill=0):
super().__init__()
if not isinstance(max_size, int):
raise TypeError(f"Size should be int. Got {type(max_size)}")
self.max_size = max_size
self.interpolation = interpolation
self.fn = min if fn == 'min' else min
self.fill = fill
def forward(self, img):
if isinstance(img, torch.Tensor):
height, width = img.shape[:2]
else:
width, height = img.size
scale = self.max_size / float(max(height, width))
if scale != 1.0:
new_size = tuple(round(dim * scale) for dim in (height, width))
img = F.resize(img, new_size, self.interpolation)
pad_h = self.max_size - new_size[0]
pad_w = self.max_size - new_size[1]
img = F.pad(img, padding=[pad_w//2, pad_h//2, pad_w - pad_w//2, pad_h - pad_h//2], fill=self.fill)
return img
def _convert_to_rgb(image):
return image.convert('RGB')
def image_transform(
image_size: int,
is_train: bool,
mean: Optional[Tuple[float, ...]] = None,
std: Optional[Tuple[float, ...]] = None,
resize_longest_max: bool = False,
fill_color: int = 0,
aug_cfg: Optional[Union[Dict[str, Any], AugmentationCfg]] = None,
):
mean = mean or OPENAI_DATASET_MEAN
if not isinstance(mean, (list, tuple)):
mean = (mean,) * 3
std = std or OPENAI_DATASET_STD
if not isinstance(std, (list, tuple)):
std = (std,) * 3
if isinstance(image_size, (list, tuple)) and image_size[0] == image_size[1]:
# for square size, pass size as int so that Resize() uses aspect preserving shortest edge
image_size = image_size[0]
if isinstance(aug_cfg, dict):
aug_cfg = AugmentationCfg(**aug_cfg)
else:
aug_cfg = aug_cfg or AugmentationCfg()
normalize = Normalize(mean=mean, std=std)
if is_train:
aug_cfg_dict = {k: v for k, v in asdict(aug_cfg).items() if v is not None}
use_timm = aug_cfg_dict.pop('use_timm', False)
if use_timm:
from timm.data import create_transform # timm can still be optional
if isinstance(image_size, (tuple, list)):
assert len(image_size) >= 2
input_size = (3,) + image_size[-2:]
else:
input_size = (3, image_size, image_size)
# by default, timm aug randomly alternates bicubic & bilinear for better robustness at inference time
aug_cfg_dict.setdefault('interpolation', 'random')
aug_cfg_dict.setdefault('color_jitter', None) # disable by default
train_transform = create_transform(
input_size=input_size,
is_training=True,
hflip=0.,
mean=mean,
std=std,
re_mode='pixel',
**aug_cfg_dict,
)
else:
train_transform = Compose([
RandomResizedCrop(
image_size,
scale=aug_cfg_dict.pop('scale'),
interpolation=InterpolationMode.BICUBIC,
),
_convert_to_rgb,
ToTensor(),
normalize,
])
if aug_cfg_dict:
warnings.warn(f'Unused augmentation cfg items, specify `use_timm` to use ({list(aug_cfg_dict.keys())}).')
return train_transform
else:
if resize_longest_max:
transforms = [
ResizeMaxSize(image_size, fill=fill_color)
]
else:
transforms = [
Resize(image_size, interpolation=InterpolationMode.BICUBIC),
CenterCrop(image_size),
]
transforms.extend([
_convert_to_rgb,
ToTensor(),
normalize,
])
return Compose(transforms)
|