waveform_feature_encoder.py

# 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)