forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathdownsampling.py
More file actions
399 lines (342 loc) 路 15.5 KB
/
Copy pathdownsampling.py
File metadata and controls
399 lines (342 loc) 路 15.5 KB
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
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
# Copyright 2026 The HuggingFace Team. All rights reserved.
#
# 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.
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..utils import deprecate
from .normalization import RMSNorm
from .upsampling import upfirdn2d_native
class Downsample1D(nn.Module):
"""A 1D downsampling layer with an optional convolution.
Parameters:
channels (`int`):
number of channels in the inputs and outputs.
use_conv (`bool`, default `False`):
option to use a convolution.
out_channels (`int`, optional):
number of output channels. Defaults to `channels`.
padding (`int`, default `1`):
padding for the convolution.
name (`str`, default `conv`):
name of the downsampling 1D layer.
"""
def __init__(
self,
channels: int,
use_conv: bool = False,
out_channels: int | None = None,
padding: int = 1,
name: str = "conv",
):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.padding = padding
stride = 2
self.name = name
if use_conv:
self.conv = nn.Conv1d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
else:
assert self.channels == self.out_channels
self.conv = nn.AvgPool1d(kernel_size=stride, stride=stride)
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
assert inputs.shape[1] == self.channels
return self.conv(inputs)
class Downsample2D(nn.Module):
"""A 2D downsampling layer with an optional convolution.
Parameters:
channels (`int`):
number of channels in the inputs and outputs.
use_conv (`bool`, default `False`):
option to use a convolution.
out_channels (`int`, optional):
number of output channels. Defaults to `channels`.
padding (`int`, default `1`):
padding for the convolution.
name (`str`, default `conv`):
name of the downsampling 2D layer.
"""
def __init__(
self,
channels: int,
use_conv: bool = False,
out_channels: int | None = None,
padding: int = 1,
name: str = "conv",
kernel_size=3,
norm_type=None,
eps=None,
elementwise_affine=None,
bias=True,
):
super().__init__()
self.channels = channels
self.out_channels = out_channels or channels
self.use_conv = use_conv
self.padding = padding
stride = 2
self.name = name
if norm_type == "ln_norm":
self.norm = nn.LayerNorm(channels, eps, elementwise_affine)
elif norm_type == "rms_norm":
self.norm = RMSNorm(channels, eps, elementwise_affine)
elif norm_type is None:
self.norm = None
else:
raise ValueError(f"unknown norm_type: {norm_type}")
if use_conv:
conv = nn.Conv2d(
self.channels, self.out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=bias
)
else:
assert self.channels == self.out_channels
conv = nn.AvgPool2d(kernel_size=stride, stride=stride)
# TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
if name == "conv":
self.Conv2d_0 = conv
self.conv = conv
elif name == "Conv2d_0":
self.conv = conv
else:
self.conv = conv
def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
if len(args) > 0 or kwargs.get("scale", None) is not None:
deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
deprecate("scale", "1.0.0", deprecation_message)
assert hidden_states.shape[1] == self.channels
if self.norm is not None:
hidden_states = self.norm(hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
if self.use_conv and self.padding == 0:
pad = (0, 1, 0, 1)
hidden_states = F.pad(hidden_states, pad, mode="constant", value=0)
assert hidden_states.shape[1] == self.channels
hidden_states = self.conv(hidden_states)
return hidden_states
class FirDownsample2D(nn.Module):
"""A 2D FIR downsampling layer with an optional convolution.
Parameters:
channels (`int`):
number of channels in the inputs and outputs.
use_conv (`bool`, default `False`):
option to use a convolution.
out_channels (`int`, optional):
number of output channels. Defaults to `channels`.
fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
kernel for the FIR filter.
"""
def __init__(
self,
channels: int | None = None,
out_channels: int | None = None,
use_conv: bool = False,
fir_kernel: tuple[int, int, int, int] = (1, 3, 3, 1),
):
super().__init__()
out_channels = out_channels if out_channels else channels
if use_conv:
self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
self.fir_kernel = fir_kernel
self.use_conv = use_conv
self.out_channels = out_channels
def _downsample_2d(
self,
hidden_states: torch.Tensor,
weight: torch.Tensor | None = None,
kernel: torch.Tensor | None = None,
factor: int = 2,
gain: float = 1,
) -> torch.Tensor:
"""Fused `Conv2d()` followed by `downsample_2d()`.
Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
arbitrary order.
Args:
hidden_states (`torch.Tensor`):
Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
weight (`torch.Tensor`, *optional*):
Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
performed by `inChannels = x.shape[0] // numGroups`.
kernel (`torch.Tensor`, *optional*):
FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
corresponds to average pooling.
factor (`int`, *optional*, default to `2`):
Integer downsampling factor.
gain (`float`, *optional*, default to `1.0`):
Scaling factor for signal magnitude.
Returns:
output (`torch.Tensor`):
Tensor of the shape `[N, C, H // factor, W // factor]` or `[N, H // factor, W // factor, C]`, and same
datatype as `x`.
"""
assert isinstance(factor, int) and factor >= 1
if kernel is None:
kernel = [1] * factor
# setup kernel
kernel = torch.tensor(kernel, dtype=torch.float32)
if kernel.ndim == 1:
kernel = torch.outer(kernel, kernel)
kernel /= torch.sum(kernel)
kernel = kernel * gain
if self.use_conv:
_, _, convH, convW = weight.shape
pad_value = (kernel.shape[0] - factor) + (convW - 1)
stride_value = [factor, factor]
upfirdn_input = upfirdn2d_native(
hidden_states,
kernel.to(device=hidden_states.device, dtype=hidden_states.dtype),
pad=((pad_value + 1) // 2, pad_value // 2),
)
output = F.conv2d(upfirdn_input, weight, stride=stride_value, padding=0)
else:
pad_value = kernel.shape[0] - factor
output = upfirdn2d_native(
hidden_states,
kernel.to(device=hidden_states.device, dtype=hidden_states.dtype),
down=factor,
pad=((pad_value + 1) // 2, pad_value // 2),
)
return output
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
if self.use_conv:
downsample_input = self._downsample_2d(hidden_states, weight=self.Conv2d_0.weight, kernel=self.fir_kernel)
hidden_states = downsample_input + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
else:
hidden_states = self._downsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
return hidden_states
# downsample/upsample layer used in k-upscaler, might be able to use FirDownsample2D/DirUpsample2D instead
class KDownsample2D(nn.Module):
r"""A 2D K-downsampling layer.
Parameters:
pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
"""
def __init__(self, pad_mode: str = "reflect"):
super().__init__()
self.pad_mode = pad_mode
kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]])
self.pad = kernel_1d.shape[1] // 2 - 1
self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
inputs = F.pad(inputs, (self.pad,) * 4, self.pad_mode)
weight = inputs.new_zeros(
[
inputs.shape[1],
inputs.shape[1],
self.kernel.shape[0],
self.kernel.shape[1],
]
)
indices = torch.arange(inputs.shape[1], device=inputs.device)
kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
weight[indices, indices] = kernel
return F.conv2d(inputs, weight, stride=2)
class CogVideoXDownsample3D(nn.Module):
# Todo: Wait for paper release.
r"""
A 3D Downsampling layer using in [CogVideoX]() by Tsinghua University & ZhipuAI
Args:
in_channels (`int`):
Number of channels in the input image.
out_channels (`int`):
Number of channels produced by the convolution.
kernel_size (`int`, defaults to `3`):
Size of the convolving kernel.
stride (`int`, defaults to `2`):
Stride of the convolution.
padding (`int`, defaults to `0`):
Padding added to all four sides of the input.
compress_time (`bool`, defaults to `False`):
Whether or not to compress the time dimension.
"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: int = 3,
stride: int = 2,
padding: int = 0,
compress_time: bool = False,
):
super().__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
self.compress_time = compress_time
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.compress_time:
batch_size, channels, frames, height, width = x.shape
# (batch_size, channels, frames, height, width) -> (batch_size, height, width, channels, frames) -> (batch_size * height * width, channels, frames)
x = x.permute(0, 3, 4, 1, 2).reshape(batch_size * height * width, channels, frames)
if x.shape[-1] % 2 == 1:
x_first, x_rest = x[..., 0], x[..., 1:]
if x_rest.shape[-1] > 0:
# (batch_size * height * width, channels, frames - 1) -> (batch_size * height * width, channels, (frames - 1) // 2)
x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
x = torch.cat([x_first[..., None], x_rest], dim=-1)
# (batch_size * height * width, channels, (frames // 2) + 1) -> (batch_size, height, width, channels, (frames // 2) + 1) -> (batch_size, channels, (frames // 2) + 1, height, width)
x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
else:
# (batch_size * height * width, channels, frames) -> (batch_size * height * width, channels, frames // 2)
x = F.avg_pool1d(x, kernel_size=2, stride=2)
# (batch_size * height * width, channels, frames // 2) -> (batch_size, height, width, channels, frames // 2) -> (batch_size, channels, frames // 2, height, width)
x = x.reshape(batch_size, height, width, channels, x.shape[-1]).permute(0, 3, 4, 1, 2)
# Pad the tensor
pad = (0, 1, 0, 1)
x = F.pad(x, pad, mode="constant", value=0)
batch_size, channels, frames, height, width = x.shape
# (batch_size, channels, frames, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size * frames, channels, height, width)
x = x.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channels, height, width)
x = self.conv(x)
# (batch_size * frames, channels, height, width) -> (batch_size, frames, channels, height, width) -> (batch_size, channels, frames, height, width)
x = x.reshape(batch_size, frames, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
return x
def downsample_2d(
hidden_states: torch.Tensor,
kernel: torch.Tensor | None = None,
factor: int = 2,
gain: float = 1,
) -> torch.Tensor:
r"""Downsample2D a batch of 2D images with the given filter.
Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and downsamples each image with the
given filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the
specified `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its
shape is a multiple of the downsampling factor.
Args:
hidden_states (`torch.Tensor`)
Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
kernel (`torch.Tensor`, *optional*):
FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
corresponds to average pooling.
factor (`int`, *optional*, default to `2`):
Integer downsampling factor.
gain (`float`, *optional*, default to `1.0`):
Scaling factor for signal magnitude.
Returns:
output (`torch.Tensor`):
Tensor of the shape `[N, C, H // factor, W // factor]`
"""
assert isinstance(factor, int) and factor >= 1
if kernel is None:
kernel = [1] * factor
kernel = torch.tensor(kernel, dtype=torch.float32)
if kernel.ndim == 1:
kernel = torch.outer(kernel, kernel)
kernel /= torch.sum(kernel)
kernel = kernel * gain
pad_value = kernel.shape[0] - factor
output = upfirdn2d_native(
hidden_states,
kernel.to(device=hidden_states.device, dtype=hidden_states.dtype),
down=factor,
pad=((pad_value + 1) // 2, pad_value // 2),
)
return output