File size: 8,993 Bytes
86dc2b6 | 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 | # MIT License
#
# Copyright (c) 2026 audio-embeddings contributors
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from collections.abc import Sequence
from math import prod
import torch
import torch.nn as nn
from timm.layers import PatchEmbed as TimmPatchEmbed
def _is_power_of_two(value: int) -> bool:
return value > 0 and value & (value - 1) == 0
def _build_hmlp_kernel_schedule(
patch_size: tuple[int, int],
) -> tuple[tuple[int, int], ...]:
"""Build non-overlapping aggregation steps for one target patch.
The paper starts from 4x4 subpatches and doubles both axes until reaching
16x16. For rectangular patches, an axis stops growing once it reaches its
target while the other axis keeps doubling.
"""
patch_height, patch_width = patch_size
if not _is_power_of_two(patch_height) or not _is_power_of_two(patch_width):
raise ValueError(
"hMLP patch dimensions must be powers of two so each axis can be "
f"doubled exactly, got {patch_size}"
)
first_kernel = (min(4, patch_height), min(4, patch_width))
schedule = [first_kernel]
current_height, current_width = first_kernel
while (current_height, current_width) != patch_size:
kernel_height = 2 if current_height < patch_height else 1
kernel_width = 2 if current_width < patch_width else 1
schedule.append((kernel_height, kernel_width))
current_height *= kernel_height
current_width *= kernel_width
return tuple(schedule)
def _resolve_hmlp_kernel_schedule(
patch_size: tuple[int, int],
kernel_schedule: Sequence[Sequence[int]] | None,
) -> tuple[tuple[int, int], ...]:
if kernel_schedule is None:
return _build_hmlp_kernel_schedule(patch_size)
schedule: list[tuple[int, int]] = []
for stage_index, stage in enumerate(kernel_schedule):
try:
stage_values = tuple(stage)
except TypeError as error:
raise ValueError(
"Each hMLP stage must contain [kernel_height, kernel_width], "
f"stage {stage_index} has {stage}"
) from error
if len(stage_values) != 2:
raise ValueError(
"Each hMLP stage must contain [kernel_height, kernel_width], "
f"stage {stage_index} has {stage}"
)
try:
kernel_size = tuple(int(value) for value in stage_values)
except (TypeError, ValueError) as error:
raise ValueError(
f"hMLP stage kernels must be integers, got {stage_values}"
) from error
if any(
isinstance(original, bool) or normalized != original
for normalized, original in zip(kernel_size, stage_values)
):
raise ValueError(f"hMLP stage kernels must be integers, got {stage_values}")
if any(value <= 0 for value in kernel_size):
raise ValueError(f"hMLP stage kernels must be positive, got {kernel_size}")
schedule.append(kernel_size)
if not schedule:
raise ValueError("hMLP kernel_schedule must contain at least one stage")
aggregated_patch_size = tuple(
prod(kernel_size[axis] for kernel_size in schedule) for axis in range(2)
)
if aggregated_patch_size != patch_size:
raise ValueError(
"hMLP kernel_schedule stages must multiply to patch_size; "
f"got {aggregated_patch_size} from {tuple(schedule)}, expected {patch_size}"
)
return tuple(schedule)
class HierarchicalMLPPatchEmbed(nn.Module):
"""hMLP patch stem with independent, hierarchical patch aggregation."""
def __init__(
self,
img_size: tuple[int, int] = (128, 256),
patch_size: tuple[int, int] = (16, 16),
in_chans: int = 1,
embed_dim: int = 768,
bias: bool = True,
kernel_schedule: Sequence[Sequence[int]] | None = None,
) -> None:
super().__init__()
self.img_size = tuple(img_size)
self.patch_size = tuple(patch_size)
self.in_chans = in_chans
self.embed_dim = embed_dim
self.bias = bias
self.kernel_schedule = _resolve_hmlp_kernel_schedule(
self.patch_size,
kernel_schedule,
)
self.num_patches = (self.img_size[0] // self.patch_size[0]) * (
self.img_size[1] // self.patch_size[1]
)
hidden_dim = max(1, embed_dim // 4)
layers: list[nn.Module] = []
input_dim = in_chans
for stage_index, kernel_size in enumerate(self.kernel_schedule):
is_last = stage_index == len(self.kernel_schedule) - 1
output_dim = embed_dim if is_last else hidden_dim
layers.extend(
[
nn.Conv2d(
input_dim,
output_dim,
kernel_size=kernel_size,
stride=kernel_size,
bias=bias,
),
nn.SyncBatchNorm(output_dim),
]
)
if not is_last:
layers.append(nn.GELU())
input_dim = output_dim
self.proj = nn.Sequential(*layers)
def forward(self, x: torch.Tensor) -> torch.Tensor:
if x.ndim != 4:
raise ValueError(f"Expected input with shape [B, C, H, W], got {x.shape}")
if x.shape[1] != self.in_chans:
raise ValueError(
f"Expected {self.in_chans} input channels, got {x.shape[1]}"
)
if x.shape[2] % self.patch_size[0] != 0 or x.shape[3] % self.patch_size[1] != 0:
raise ValueError(
"Input spatial dimensions must be divisible by the hMLP patch "
f"size {self.patch_size}, got {tuple(x.shape[2:])}"
)
return self.proj(x).flatten(2).transpose(1, 2)
class PatchEmbed(nn.Module):
"""
2D Image to Patch Embedding.
Args:
img_size (tuple[int, int]): Input image size (H, W).
patch_size (tuple[int, int]): Patch size (H, W).
in_chans (int): Number of input channels.
embed_dim (int): Embedding dimension.
"""
def __init__(
self,
img_size: tuple[int, int] = (128, 256),
patch_size: tuple[int, int] = (16, 16),
in_chans: int = 1,
embed_dim: int = 768,
bias: bool = True,
stem_type: str = "linear",
hmlp_kernel_schedule: Sequence[Sequence[int]] | None = None,
):
super().__init__()
self.img_size = tuple(img_size)
self.patch_size = tuple(patch_size)
self.in_chans = in_chans
self.embed_dim = embed_dim
self.bias = bias
self.stem_type = stem_type.strip().lower().replace("-", "_")
if self.stem_type == "linear":
self.patch_embed = TimmPatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=in_chans,
embed_dim=embed_dim,
flatten=True,
bias=bias,
strict_img_size=False,
)
elif self.stem_type == "hmlp":
self.patch_embed = HierarchicalMLPPatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=in_chans,
embed_dim=embed_dim,
bias=bias,
kernel_schedule=hmlp_kernel_schedule,
)
else:
raise ValueError(
f"Unknown stem_type={stem_type!r}; expected 'linear' or 'hmlp'"
)
self.num_patches = self.patch_embed.num_patches
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Forward pass.
Args:
x (torch.Tensor): Input tensor [B, C, H, W].
Returns:
torch.Tensor: Patch embeddings [B, N, D].
"""
return self.patch_embed(x)
|