File size: 5,776 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 | # 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 __future__ import annotations
from typing import Sequence
import torch
from einops import rearrange
from einops.layers.torch import Rearrange
from torch import nn
def _parse_conv_layers_spec(
conv_layers_spec: str | Sequence[Sequence[int]] | Sequence[tuple[int, int, int]],
) -> list[tuple[int, int, int]]:
if isinstance(conv_layers_spec, str):
# Config-driven expression style used by wavjepa, e.g.
# "[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512, 2, 2)]"
parsed = eval(conv_layers_spec, {"__builtins__": {}}, {}) # noqa: S307
else:
parsed = conv_layers_spec
out: list[tuple[int, int, int]] = []
for layer in parsed:
if len(layer) != 3:
raise ValueError(f"Invalid conv layer spec {layer}, expected (dim, k, s)")
dim, kernel, stride = layer
out.append((int(dim), int(kernel), int(stride)))
if len(out) == 0:
raise ValueError("conv_layers_spec must contain at least one layer")
return out
class WaveformFeatureEncoder(nn.Module):
"""
Convolutional waveform feature encoder that outputs a token sequence.
Input shape: [B, C, T]
Output shape: [B, N, F]
"""
def __init__(
self,
conv_layers_spec: str
| Sequence[Sequence[int]]
| Sequence[
tuple[int, int, int]
] = "[(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512, 2, 2)]",
in_channels: int = 1,
dropout: float = 0.0,
mode: str = "default",
conv_bias: bool = False,
depthwise: bool = False,
) -> None:
super().__init__()
if mode not in {"default", "layer_norm"}:
raise ValueError(
f"Unknown mode='{mode}', expected 'default' or 'layer_norm'"
)
self.conv_layers_spec = _parse_conv_layers_spec(conv_layers_spec)
self.in_channels = in_channels
self.depthwise = depthwise
layers: list[nn.Module] = []
in_dim = in_channels
for idx, (out_dim, kernel, stride) in enumerate(self.conv_layers_spec):
layers.append(
self._make_block(
in_dim=in_dim,
out_dim=out_dim,
kernel=kernel,
stride=stride,
dropout=dropout,
mode=mode,
conv_bias=conv_bias,
depthwise=depthwise,
is_first=idx == 0,
)
)
in_dim = out_dim
self.cnn = nn.Sequential(*layers)
self.embedding_dim = self.conv_layers_spec[-1][0]
@staticmethod
def _make_block(
in_dim: int,
out_dim: int,
kernel: int,
stride: int,
dropout: float,
mode: str,
conv_bias: bool,
depthwise: bool,
is_first: bool,
) -> nn.Module:
if depthwise:
if out_dim % in_dim != 0:
raise ValueError(
"Depthwise mode requires out_dim to be a multiple of in_dim, "
f"got out_dim={out_dim}, in_dim={in_dim}"
)
conv = nn.Conv1d(
in_dim,
out_dim,
kernel_size=kernel,
stride=stride,
bias=conv_bias,
groups=in_dim,
)
else:
conv = nn.Conv1d(
in_dim,
out_dim,
kernel_size=kernel,
stride=stride,
bias=conv_bias,
)
nn.init.kaiming_normal_(conv.weight)
if mode == "layer_norm":
return nn.Sequential(
conv,
nn.Dropout(p=dropout),
Rearrange("... c t -> ... t c"),
nn.LayerNorm(out_dim, elementwise_affine=True),
Rearrange("... t c -> ... c t"),
nn.GELU(),
)
if mode == "default" and is_first:
return nn.Sequential(
conv,
nn.Dropout(p=dropout),
nn.GroupNorm(out_dim, out_dim, affine=True),
nn.GELU(),
)
return nn.Sequential(conv, nn.Dropout(p=dropout), nn.GELU())
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.cnn(x)
return rearrange(x, "b f n -> b n f")
def total_patches(self, time_samples: int) -> int:
n = int(time_samples)
for _, kernel, stride in self.conv_layers_spec:
if n < kernel:
return 0
n = (n - kernel) // stride + 1
return int(n)
|