tts/modules/wavvae/encoder/common_modules/seanet.py

# MIT License

# Copyright (c) Meta Platforms, Inc. and affiliates.

# 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.

# Copyright (c) [2023] [Meta Platforms, Inc. and affiliates.] 
# Copyright (c) [2025] [Ziyue Jiang] 
# SPDX-License-Identifier: MIT
# This file has been modified by Ziyue Jiang on 2025/03/19
# Original file was released under MIT, with the full license text # available at https://github.com/facebookresearch/encodec/blob/gh-pages/LICENSE.
# This modified file is released under the same license.

"""Encodec SEANet-based encoder and decoder implementation."""

import typing as tp

import numpy as np
import torch.nn as nn

from .conv import SConv1d
from .lstm import SLSTM


class SEANetResnetBlock(nn.Module):
    def __init__(self, dim: int, kernel_sizes: tp.List[int] = [3, 1], dilations: tp.List[int] = [1, 1],
                 activation: str = 'ELU', activation_params: dict = {'alpha': 1.0},
                 norm: str = 'weight_norm', norm_params: tp.Dict[str, tp.Any] = {}, causal: bool = False,
                 pad_mode: str = 'reflect', compress: int = 2, true_skip: bool = True):
        super().__init__()
        assert len(kernel_sizes) == len(dilations), 'Number of kernel sizes should match number of dilations'
        act = getattr(nn, activation)
        hidden = dim // compress
        block = []
        for i, (kernel_size, dilation) in enumerate(zip(kernel_sizes, dilations)):
            in_chs = dim if i == 0 else hidden
            out_chs = dim if i == len(kernel_sizes) - 1 else hidden
            block += [
                act(**activation_params),
                SConv1d(in_chs, out_chs, kernel_size=kernel_size, dilation=dilation,
                        norm=norm, norm_kwargs=norm_params,
                        causal=causal, pad_mode=pad_mode),
            ]
        self.block = nn.Sequential(*block)
        self.shortcut: nn.Module
        if true_skip:
            self.shortcut = nn.Identity()
        else:
            self.shortcut = SConv1d(dim, dim, kernel_size=1, norm=norm, norm_kwargs=norm_params,
                                    causal=causal, pad_mode=pad_mode)

    def forward(self, x):
        return self.shortcut(x) + self.block(x)


class SEANetEncoder(nn.Module):
    def __init__(self, channels: int = 1, dimension: int = 128, n_filters: int = 32, n_residual_layers: int = 1,
                 ratios: tp.List[int] = [8, 5, 4, 2], activation: str = 'ELU', activation_params: dict = {'alpha': 1.0},
                 norm: str = 'weight_norm', norm_params: tp.Dict[str, tp.Any] = {}, kernel_size: int = 7,
                 last_kernel_size: int = 7, residual_kernel_size: int = 3, dilation_base: int = 2, causal: bool = False,
                 pad_mode: str = 'reflect', true_skip: bool = False, compress: int = 2, lstm: int = 2):
        super().__init__()
        self.channels = channels
        self.dimension = dimension
        self.n_filters = n_filters
        self.ratios = list(reversed(ratios))
        del ratios
        self.n_residual_layers = n_residual_layers
        self.hop_length = np.prod(self.ratios)

        act = getattr(nn, activation)
        mult = 1
        model: tp.List[nn.Module] = [
            SConv1d(channels, mult * n_filters, kernel_size, norm=norm, norm_kwargs=norm_params,
                    causal=causal, pad_mode=pad_mode)
        ]
        # Downsample to raw audio scale
        for i, ratio in enumerate(self.ratios):
            # Add residual layers
            for j in range(n_residual_layers):
                model += [
                    SEANetResnetBlock(mult * n_filters, kernel_sizes=[residual_kernel_size, 1],
                                      dilations=[dilation_base ** j, 1],
                                      norm=norm, norm_params=norm_params,
                                      activation=activation, activation_params=activation_params,
                                      causal=causal, pad_mode=pad_mode, compress=compress, true_skip=true_skip)]

            # Add downsampling layers
            model += [
                act(**activation_params),
                SConv1d(mult * n_filters, mult * n_filters * 2,
                        kernel_size=ratio * 2, stride=ratio,
                        norm=norm, norm_kwargs=norm_params,
                        causal=causal, pad_mode=pad_mode),
            ]
            mult *= 2

        if lstm:
            model += [SLSTM(mult * n_filters, num_layers=lstm)]

        model += [
            act(**activation_params),
            SConv1d(mult * n_filters, dimension, last_kernel_size, norm=norm, norm_kwargs=norm_params,
                    causal=causal, pad_mode=pad_mode)
        ]

        self.model = nn.Sequential(*model)

    def forward(self, x):
        return self.model(x)