File size: 11,820 Bytes
8c31d70 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 |
# 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.
"""The patcher and unpatcher implementation for 2D and 3D data."""
import torch
import torch.nn.functional as F
from einops import rearrange
_WAVELETS = {
"haar": torch.tensor([0.7071067811865476, 0.7071067811865476]),
"rearrange": torch.tensor([1.0, 1.0]),
}
_PERSISTENT = False
class Patcher(torch.nn.Module):
"""A module to convert image tensors into patches using torch operations.
The main difference from `class Patching` is that this module implements
all operations using torch, rather than python or numpy, for efficiency purpose.
It's bit-wise identical to the Patching module outputs, with the added
benefit of being torch.jit scriptable.
"""
def __init__(self, patch_size=1, patch_method="haar"):
super().__init__()
self.patch_size = patch_size
self.patch_method = patch_method
self.register_buffer("wavelets", _WAVELETS[patch_method], persistent=_PERSISTENT)
self.range = range(int(torch.log2(torch.tensor(self.patch_size)).item()))
self.register_buffer("_arange", torch.arange(_WAVELETS[patch_method].shape[0]), persistent=_PERSISTENT)
for param in self.parameters():
param.requires_grad = False
def forward(self, x):
if self.patch_method == "haar":
return self._haar(x)
elif self.patch_method == "rearrange":
return self._arrange(x)
else:
raise ValueError("Unknown patch method: " + self.patch_method)
def _dwt(self, x, mode="reflect", rescale=False):
dtype = x.dtype
h = self.wavelets
n = h.shape[0]
g = x.shape[1]
hl = h.flip(0).reshape(1, 1, -1).repeat(g, 1, 1)
hh = (h * ((-1) ** self._arange)).reshape(1, 1, -1).repeat(g, 1, 1)
hh = hh.to(dtype=dtype)
hl = hl.to(dtype=dtype)
x = F.pad(x, pad=(n - 2, n - 1, n - 2, n - 1), mode=mode).to(dtype)
xl = F.conv2d(x, hl.unsqueeze(2), groups=g, stride=(1, 2))
xh = F.conv2d(x, hh.unsqueeze(2), groups=g, stride=(1, 2))
xll = F.conv2d(xl, hl.unsqueeze(3), groups=g, stride=(2, 1))
xlh = F.conv2d(xl, hh.unsqueeze(3), groups=g, stride=(2, 1))
xhl = F.conv2d(xh, hl.unsqueeze(3), groups=g, stride=(2, 1))
xhh = F.conv2d(xh, hh.unsqueeze(3), groups=g, stride=(2, 1))
out = torch.cat([xll, xlh, xhl, xhh], dim=1)
if rescale:
out = out / 2
return out
def _haar(self, x):
for _ in self.range:
x = self._dwt(x, rescale=True)
return x
def _arrange(self, x):
x = rearrange(x, "b c (h p1) (w p2) -> b (c p1 p2) h w", p1=self.patch_size, p2=self.patch_size).contiguous()
return x
class Patcher3D(Patcher):
"""A 3D discrete wavelet transform for video data, expects 5D tensor, i.e. a batch of videos."""
def __init__(self, patch_size=1, patch_method="haar"):
super().__init__(patch_method=patch_method, patch_size=patch_size)
self.register_buffer(
"patch_size_buffer", patch_size * torch.ones([1], dtype=torch.int32), persistent=_PERSISTENT
)
def _dwt(self, x, mode="reflect", rescale=False):
dtype = x.dtype
h = self.wavelets
n = h.shape[0]
g = x.shape[1]
hl = h.flip(0).reshape(1, 1, -1).repeat(g, 1, 1)
hh = (h * ((-1) ** self._arange)).reshape(1, 1, -1).repeat(g, 1, 1)
hh = hh.to(dtype=dtype)
hl = hl.to(dtype=dtype)
# Handles temporal axis.
x = F.pad(x, pad=(max(0, n - 2), n - 1, n - 2, n - 1, n - 2, n - 1), mode=mode).to(dtype)
xl = F.conv3d(x, hl.unsqueeze(3).unsqueeze(4), groups=g, stride=(2, 1, 1))
xh = F.conv3d(x, hh.unsqueeze(3).unsqueeze(4), groups=g, stride=(2, 1, 1))
# Handles spatial axes.
xll = F.conv3d(xl, hl.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
xlh = F.conv3d(xl, hh.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
xhl = F.conv3d(xh, hl.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
xhh = F.conv3d(xh, hh.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
xlll = F.conv3d(xll, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xllh = F.conv3d(xll, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xlhl = F.conv3d(xlh, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xlhh = F.conv3d(xlh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xhll = F.conv3d(xhl, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xhlh = F.conv3d(xhl, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xhhl = F.conv3d(xhh, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xhhh = F.conv3d(xhh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
out = torch.cat([xlll, xllh, xlhl, xlhh, xhll, xhlh, xhhl, xhhh], dim=1)
if rescale:
out = out / (2 * torch.sqrt(torch.tensor(2.0)))
return out
def _haar(self, x):
xi, xv = torch.split(x, [1, x.shape[2] - 1], dim=2)
x = torch.cat([xi.repeat_interleave(self.patch_size, dim=2), xv], dim=2)
for _ in self.range:
x = self._dwt(x, rescale=True)
return x
def _arrange(self, x):
xi, xv = torch.split(x, [1, x.shape[2] - 1], dim=2)
x = torch.cat([xi.repeat_interleave(self.patch_size, dim=2), xv], dim=2)
x = rearrange(
x,
"b c (t p1) (h p2) (w p3) -> b (c p1 p2 p3) t h w",
p1=self.patch_size,
p2=self.patch_size,
p3=self.patch_size,
).contiguous()
return x
class UnPatcher(torch.nn.Module):
"""A module to convert patches into image tensorsusing torch operations.
The main difference from `class Unpatching` is that this module implements
all operations using torch, rather than python or numpy, for efficiency purpose.
It's bit-wise identical to the Unpatching module outputs, with the added
benefit of being torch.jit scriptable.
"""
def __init__(self, patch_size=1, patch_method="haar"):
super().__init__()
self.patch_size = patch_size
self.patch_method = patch_method
self.register_buffer("wavelets", _WAVELETS[patch_method], persistent=_PERSISTENT)
self.range = range(int(torch.log2(torch.tensor(self.patch_size)).item()))
self.register_buffer("_arange", torch.arange(_WAVELETS[patch_method].shape[0]), persistent=_PERSISTENT)
for param in self.parameters():
param.requires_grad = False
def forward(self, x):
if self.patch_method == "haar":
return self._ihaar(x)
elif self.patch_method == "rearrange":
return self._iarrange(x)
else:
raise ValueError("Unknown patch method: " + self.patch_method)
def _idwt(self, x, rescale=False):
dtype = x.dtype
h = self.wavelets
n = h.shape[0]
g = x.shape[1] // 4
hl = h.flip([0]).reshape(1, 1, -1).repeat([g, 1, 1])
hh = (h * ((-1) ** self._arange)).reshape(1, 1, -1).repeat(g, 1, 1)
hh = hh.to(dtype=dtype)
hl = hl.to(dtype=dtype)
xll, xlh, xhl, xhh = torch.chunk(x.to(dtype), 4, dim=1)
# Inverse transform.
yl = torch.nn.functional.conv_transpose2d(xll, hl.unsqueeze(3), groups=g, stride=(2, 1), padding=(n - 2, 0))
yl += torch.nn.functional.conv_transpose2d(xlh, hh.unsqueeze(3), groups=g, stride=(2, 1), padding=(n - 2, 0))
yh = torch.nn.functional.conv_transpose2d(xhl, hl.unsqueeze(3), groups=g, stride=(2, 1), padding=(n - 2, 0))
yh += torch.nn.functional.conv_transpose2d(xhh, hh.unsqueeze(3), groups=g, stride=(2, 1), padding=(n - 2, 0))
y = torch.nn.functional.conv_transpose2d(yl, hl.unsqueeze(2), groups=g, stride=(1, 2), padding=(0, n - 2))
y += torch.nn.functional.conv_transpose2d(yh, hh.unsqueeze(2), groups=g, stride=(1, 2), padding=(0, n - 2))
if rescale:
y = y * 2
return y
def _ihaar(self, x):
for _ in self.range:
x = self._idwt(x, rescale=True)
return x
def _iarrange(self, x):
x = rearrange(x, "b (c p1 p2) h w -> b c (h p1) (w p2)", p1=self.patch_size, p2=self.patch_size)
return x
class UnPatcher3D(UnPatcher):
"""A 3D inverse discrete wavelet transform for video wavelet decompositions."""
def __init__(self, patch_size=1, patch_method="haar"):
super().__init__(patch_method=patch_method, patch_size=patch_size)
def _idwt(self, x, rescale=False):
dtype = x.dtype
h = self.wavelets
g = x.shape[1] // 8 # split into 8 spatio-temporal filtered tesnors.
hl = h.flip([0]).reshape(1, 1, -1).repeat([g, 1, 1])
hh = (h * ((-1) ** self._arange)).reshape(1, 1, -1).repeat(g, 1, 1)
hl = hl.to(dtype=dtype)
hh = hh.to(dtype=dtype)
xlll, xllh, xlhl, xlhh, xhll, xhlh, xhhl, xhhh = torch.chunk(x, 8, dim=1)
# Height height transposed convolutions.
xll = F.conv_transpose3d(xlll, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xll += F.conv_transpose3d(xllh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xlh = F.conv_transpose3d(xlhl, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xlh += F.conv_transpose3d(xlhh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xhl = F.conv_transpose3d(xhll, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xhl += F.conv_transpose3d(xhlh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xhh = F.conv_transpose3d(xhhl, hl.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
xhh += F.conv_transpose3d(xhhh, hh.unsqueeze(2).unsqueeze(3), groups=g, stride=(1, 1, 2))
# Handles width transposed convolutions.
xl = F.conv_transpose3d(xll, hl.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
xl += F.conv_transpose3d(xlh, hh.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
xh = F.conv_transpose3d(xhl, hl.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
xh += F.conv_transpose3d(xhh, hh.unsqueeze(2).unsqueeze(4), groups=g, stride=(1, 2, 1))
# Handles time axis transposed convolutions.
x = F.conv_transpose3d(xl, hl.unsqueeze(3).unsqueeze(4), groups=g, stride=(2, 1, 1))
x += F.conv_transpose3d(xh, hh.unsqueeze(3).unsqueeze(4), groups=g, stride=(2, 1, 1))
if rescale:
x = x * (2 * torch.sqrt(torch.tensor(2.0)))
return x
def _ihaar(self, x):
for _ in self.range:
x = self._idwt(x, rescale=True)
x = x[:, :, self.patch_size - 1 :, ...]
return x
def _iarrange(self, x):
x = rearrange(
x,
"b (c p1 p2 p3) t h w -> b c (t p1) (h p2) (w p3)",
p1=self.patch_size,
p2=self.patch_size,
p3=self.patch_size,
)
x = x[:, :, self.patch_size - 1 :, ...]
return x
|