File size: 8,038 Bytes
4336727 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 | import torch
import torch.nn.functional as F
import torch.nn as nn
def check_image_size(x, padder_size, mode='reflect'):
_, _, h, w = x.size()
if isinstance(padder_size, int):
padder_size_h = padder_size
padder_size_w = padder_size
else:
padder_size_h, padder_size_w = padder_size
mod_pad_h = (padder_size_h - h % padder_size_h) % padder_size_h
mod_pad_w = (padder_size_w - w % padder_size_w) % padder_size_w
x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), mode=mode)
return x
def window_partitions(x, window_size):
"""
Args:
x: (B, C, H, W)
window_size (int): window size
Returns:
windows: (num_windows*B, C, window_size, window_size)
"""
if isinstance(window_size, int):
window_size = [window_size, window_size]
B, C, H, W = x.shape
x = x.view(B, C, H // window_size[0], window_size[0], W // window_size[1], window_size[1])
windows = x.permute(0, 2, 4, 1, 3, 5).contiguous().view(-1, C, window_size[0], window_size[1])
return windows
def window_reverses(windows, window_size, H, W):
"""
Args:
windows: (num_windows*B, C, window_size, window_size)
window_size (int): Window size
H (int): Height of image
W (int): Width of image
Returns:
x: (B, C, H, W)
"""
# B = int(windows.shape[0] / (H * W / window_size / window_size))
# print('B: ', B)
# print(H // window_size)
# print(W // window_size)
if isinstance(window_size, int):
window_size = [window_size, window_size]
C = windows.shape[1]
# print('C: ', C)
x = windows.view(-1, H // window_size[0], W // window_size[1], C, window_size[0], window_size[1])
x = x.permute(0, 3, 1, 4, 2, 5).contiguous().view(-1, C, H, W)
return x
def window_partitionx(x, window_size):
_, _, H, W = x.shape
h, w = window_size * (H // window_size), window_size * (W // window_size)
x_main = window_partitions(x[:, :, :h, :w], window_size)
b_main = x_main.shape[0]
if h == H and w == W:
return x_main, [b_main]
if h != H and w != W:
x_r = window_partitions(x[:, :, :h, -window_size:], window_size)
b_r = x_r.shape[0] + b_main
x_d = window_partitions(x[:, :, -window_size:, :w], window_size)
b_d = x_d.shape[0] + b_r
x_dd = x[:, :, -window_size:, -window_size:]
b_dd = x_dd.shape[0] + b_d
# batch_list = [b_main, b_r, b_d, b_dd]
return torch.cat([x_main, x_r, x_d, x_dd], dim=0), [b_main, b_r, b_d, b_dd]
if h == H and w != W:
x_r = window_partitions(x[:, :, :h, -window_size:], window_size)
b_r = x_r.shape[0] + b_main
return torch.cat([x_main, x_r], dim=0), [b_main, b_r]
if h != H and w == W:
x_d = window_partitions(x[:, :, -window_size:, :w], window_size)
b_d = x_d.shape[0] + b_main
return torch.cat([x_main, x_d], dim=0), [b_main, b_d]
def window_reversex(windows, window_size, H, W, batch_list):
h, w = window_size * (H // window_size), window_size * (W // window_size)
# print(windows[:batch_list[0], ...].shape)
x_main = window_reverses(windows[:batch_list[0], ...], window_size, h, w)
B, C, _, _ = x_main.shape
# print('windows: ', windows.shape)
# print('batch_list: ', batch_list)
if torch.is_complex(windows):
res = torch.complex(torch.zeros([B, C, H, W]), torch.zeros([B, C, H, W]))
res = res.to(windows.device)
else:
res = torch.zeros([B, C, H, W], device=windows.device)
res[:, :, :h, :w] = x_main
if h == H and w == W:
return res
if h != H and w != W and len(batch_list) == 4:
x_dd = window_reverses(windows[batch_list[2]:, ...], window_size, window_size, window_size)
res[:, :, h:, w:] = x_dd[:, :, h - H:, w - W:]
x_r = window_reverses(windows[batch_list[0]:batch_list[1], ...], window_size, h, window_size)
res[:, :, :h, w:] = x_r[:, :, :, w - W:]
x_d = window_reverses(windows[batch_list[1]:batch_list[2], ...], window_size, window_size, w)
res[:, :, h:, :w] = x_d[:, :, h - H:, :]
return res
if w != W and len(batch_list) == 2:
x_r = window_reverses(windows[batch_list[0]:batch_list[1], ...], window_size, h, window_size)
res[:, :, :h, w:] = x_r[:, :, :, w - W:]
if h != H and len(batch_list) == 2:
x_d = window_reverses(windows[batch_list[0]:batch_list[1], ...], window_size, window_size, w)
res[:, :, h:, :w] = x_d[:, :, h - H:, :]
return res
def window_partitionxy(x, window_size, start=[0, 0]):
s_h, s_w = start
assert 0 <= s_h < window_size and 0 <= s_w < window_size
_, _, H, W = x.shape
h, w = window_size * (H // window_size), window_size * (W // window_size)
x_main, b_main = window_partitionx(x[:, :, s_h:, s_w:], window_size)
# print(x_main.shape, b_main, x[:, :, s_h:, s_w:].shape)
if s_h == 0 and s_w == 0:
return x_main, b_main
if s_h != 0 and s_w != 0:
x_l = window_partitions(x[:, :, -h:, :window_size], window_size)
b_l = x_l.shape[0] + b_main[-1]
b_main.append(b_l)
x_u = window_partitions(x[:, :, :window_size, -w:], window_size)
b_u = x_u.shape[0] + b_l
b_main.append(b_u)
x_uu = x[:, :, :window_size, :window_size]
b_uu = x_uu.shape[0] + b_u
b_main.append(b_uu)
# batch_list = [b_main, b_r, b_d, b_dd]
return torch.cat([x_main, x_l, x_u, x_uu], dim=0), b_main
def window_reversexy(windows, window_size, H, W, batch_list, start=[0, 0]):
s_h, s_w = start
assert 0 <= s_h < window_size and 0 <= s_w < window_size
if s_h == 0 and s_w == 0:
x_main = window_reversex(windows, window_size, H, W, batch_list)
return x_main
else:
h, w = window_size * (H // window_size), window_size * (W // window_size)
# print(windows[:batch_list[-4], ...].shape, batch_list[:-3], H-s_h, W-s_w)
x_main = window_reversex(windows[:batch_list[-4], ...], window_size, H-s_h, W-s_w, batch_list[:-3])
B, C, _, _ = x_main.shape
res = torch.zeros([B, C, H, W], device=windows.device)
x_uu = window_reverses(windows[batch_list[-2]:, ...], window_size, window_size, window_size)
res[:, :, :window_size, :window_size] = x_uu[:, :, :, :]
x_l = window_reverses(windows[batch_list[-4]:batch_list[-3], ...], window_size, h, window_size)
res[:, :, -h:, :window_size] = x_l
x_u = window_reverses(windows[batch_list[-3]:batch_list[-2], ...], window_size, window_size, w)
res[:, :, :window_size, -w:] = x_u[:, :, :, :]
res[:, :, s_h:, s_w:] = x_main
return res
class WindowPartition(nn.Module):
def __init__(self, window_size=8, shift_size=0):
super().__init__()
self.window_size = window_size
self.shift_size = shift_size
def forward(self, x):
H, W = x.shape[-2:]
if self.window_size is not None and (H > self.window_size and W > self.window_size):
if not self.shift_size:
x, batch_list = window_partitionx(x, self.window_size)
return x, batch_list
else:
x, batch_list = window_partitionxy(x, self.window_size, [self.shift_size, self.shift_size])
return x, batch_list
else:
return x, []
class WindowReverse(nn.Module):
def __init__(self, window_size=8, shift_size=0):
super().__init__()
self.window_size = window_size
self.shift_size = shift_size
def forward(self, x, H, W, batch_list):
# print(x.shape, batch_list)
if len(batch_list) > 0 and self.window_size is not None and (H > self.window_size and W > self.window_size):
if not self.shift_size:
x = window_reversex(x, self.window_size, H, W, batch_list)
else:
x = window_reversexy(x, self.window_size, H, W, batch_list, [self.shift_size, self.shift_size])
return x |