File size: 6,057 Bytes
f9561b9 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 |
import random
import math
import numpy as np
import torch
from torch.nn import functional as F
class SceneMaskGenerator:
def __init__(
self,
input_size,
min_num_patches=16,
max_num_patches_ratio=0.5,
min_aspect=0.3,
):
if not isinstance(input_size, tuple):
input_size = (input_size,) * 2
self.input_size = input_size
self.num_patches = input_size[0] * input_size[1]
self.min_num_patches = min_num_patches
self.max_num_patches = max_num_patches_ratio * self.num_patches
self.log_aspect_ratio = (math.log(min_aspect), -math.log(min_aspect))
def _mask(self, mask, max_mask_patches):
delta = 0
for _ in range(4):
target_area = random.uniform(self.min_num_patches, max_mask_patches)
aspect_ratio = math.exp(random.uniform(*self.log_aspect_ratio))
h = int(round(math.sqrt(target_area * aspect_ratio)))
w = int(round(math.sqrt(target_area / aspect_ratio)))
height, width = self.input_size
if w < width and h < height:
top = random.randint(0, height - h)
left = random.randint(0, width - w)
num_masked = mask[top : top + h, left : left + w].sum()
# Overlap
if 0 < h * w - num_masked <= max_mask_patches:
mask[top : top + h, left : left + w] = 1
delta = h * w - num_masked
break
return delta
def __call__(self, head_mask):
mask = np.zeros(shape=self.input_size, dtype=bool)
mask_count = 0
num_masking_patches = random.uniform(self.min_num_patches, self.max_num_patches)
while mask_count < num_masking_patches:
max_mask_patches = num_masking_patches - mask_count
delta = self._mask(mask, max_mask_patches)
if delta == 0:
break
else:
mask_count += delta
mask = torch.from_numpy(mask).unsqueeze(0)
head_mask = (
F.interpolate(head_mask.unsqueeze(0), mask.shape[-2:]).squeeze(0) < 0.5
)
return torch.logical_and(mask, head_mask).squeeze(0)
class HeadMaskGenerator:
def __init__(
self,
input_size,
min_num_patches=4,
max_num_patches_ratio=0.5,
min_aspect=0.3,
):
if not isinstance(input_size, tuple):
input_size = (input_size,) * 2
self.input_size = input_size
self.num_patches = input_size[0] * input_size[1]
self.min_num_patches = min_num_patches
self.max_num_patches_ratio = max_num_patches_ratio
self.log_aspect_ratio = (math.log(min_aspect), -math.log(min_aspect))
def __call__(
self,
x_min,
y_min,
x_max,
y_max, # coords in [0,1]
):
height = math.floor((y_max - y_min) * self.input_size[0])
width = math.floor((x_max - x_min) * self.input_size[1])
origin_area = width * height
if origin_area < self.min_num_patches:
return torch.zeros(size=self.input_size, dtype=bool)
target_area = random.uniform(
self.min_num_patches, self.max_num_patches_ratio * origin_area
)
aspect_ratio = math.exp(random.uniform(*self.log_aspect_ratio))
h = min(int(round(math.sqrt(target_area * aspect_ratio))), height)
w = min(int(round(math.sqrt(target_area / aspect_ratio))), width)
top = random.randint(0, height - h) + int(y_min * self.input_size[0])
left = random.randint(0, width - w) + int(x_min * self.input_size[1])
mask = torch.zeros(size=self.input_size, dtype=bool)
mask[top : top + h, left : left + w] = True
return mask
class MaskGenerator:
def __init__(
self,
input_size,
mask_scene: bool = False,
mask_head: bool = False,
min_scene_patches=16,
max_scene_patches_ratio=0.5,
min_head_patches=4,
max_head_patches_ratio=0.5,
min_aspect=0.3,
mask_prob=0.2,
head_prob=0.2,
):
if not isinstance(input_size, tuple):
input_size = (input_size,) * 2
self.input_size = input_size
if mask_scene:
self.scene_mask_generator = SceneMaskGenerator(
input_size, min_scene_patches, max_scene_patches_ratio, min_aspect
)
else:
self.scene_mask_generator = None
if mask_head:
self.head_mask_generator = HeadMaskGenerator(
input_size, min_head_patches, max_head_patches_ratio, min_aspect
)
else:
self.head_mask_generator = None
self.no_mask = not (mask_scene or mask_head)
self.mask_head = mask_head and not mask_scene
self.mask_scene = mask_scene and not mask_head
self.scene_prob = mask_prob
self.head_prob = head_prob
def __call__(
self,
x_min,
y_min,
x_max,
y_max,
head_mask,
):
mask_scene = random.random() < self.scene_prob
mask_head = random.random() < self.head_prob
no_mask = (
self.no_mask
or (self.mask_head and not mask_head)
or (self.mask_scene and not mask_scene)
or not (mask_scene or mask_head)
)
if no_mask:
return torch.zeros(size=self.input_size, dtype=bool)
if self.mask_scene:
return self.scene_mask_generator(head_mask)
if self.mask_head:
return self.head_mask_generator(x_min, y_min, x_max, y_max)
if mask_head and mask_scene:
return torch.logical_or(
self.scene_mask_generator(head_mask),
self.head_mask_generator(x_min, y_min, x_max, y_max),
)
elif mask_head:
return self.head_mask_generator(x_min, y_min, x_max, y_max)
return self.scene_mask_generator(head_mask)
|