stepvocoder/cosyvoice2/bigvgan/bigvgan.py

# Copyright (c) 2024 NVIDIA CORPORATION.
#   Licensed under the MIT license.

# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
#   LICENSE is in incl_licenses directory.

import os
import json
from pathlib import Path
from typing import Optional, Union, Dict

import torch
import torch.nn as nn
from torch.nn import Conv1d, ConvTranspose1d
from torch.nn.utils import weight_norm, remove_weight_norm

from stepvocoder.cosyvoice2.bigvgan import activations
from stepvocoder.cosyvoice2.bigvgan.bigvgan_utils import init_weights, get_padding
from stepvocoder.cosyvoice2.bigvgan.alias_free_activation.torch.act import Activation1d as TorchActivation1d


class AMPBlock1(torch.nn.Module):
    """
    AMPBlock applies Snake / SnakeBeta activation functions with trainable parameters that control periodicity, defined for each layer.
    AMPBlock1 has additional self.convs2 that contains additional Conv1d layers with a fixed dilation=1 followed by each layer in self.convs1

    Args:
        h (AttrDict): Hyperparameters.
        channels (int): Number of convolution channels.
        kernel_size (int): Size of the convolution kernel. Default is 3.
        dilation (tuple): Dilation rates for the convolutions. Each dilation layer has two convolutions. Default is (1, 3, 5).
        activation (str): Activation function type. Should be either 'snake' or 'snakebeta'. Default is None.
    """

    def __init__(
        self,
        channels: int,
        kernel_size: int = 3,
        dilation: tuple = (1, 3, 5),
        activation: str = None,
        use_cuda_kernel: bool = False,
        snake_logscale: bool = True
    ):
        super().__init__()
        
        self.convs1 = nn.ModuleList(
            [
                weight_norm(
                    Conv1d(
                        channels,
                        channels,
                        kernel_size,
                        stride=1,
                        dilation=d,
                        padding=get_padding(kernel_size, d),
                    )
                )
                for d in dilation
            ]
        )
        self.convs1.apply(init_weights)

        self.convs2 = nn.ModuleList(
            [
                weight_norm(
                    Conv1d(
                        channels,
                        channels,
                        kernel_size,
                        stride=1,
                        dilation=1,
                        padding=get_padding(kernel_size, 1),
                    )
                )
                for _ in range(len(dilation))
            ]
        )
        self.convs2.apply(init_weights)

        self.num_layers = len(self.convs1) + len(
            self.convs2
        )  # Total number of conv layers

        # Select which Activation1d, lazy-load cuda version to ensure backward compatibility
        if use_cuda_kernel:
            from alias_free_activation.cuda.activation1d import (
                Activation1d as CudaActivation1d,
            )

            Activation1d = CudaActivation1d
        else:
            Activation1d = TorchActivation1d

        # Activation functions
        if activation == "snake":
            self.activations = nn.ModuleList(
                [
                    Activation1d(
                        activation=activations.Snake(
                            channels, alpha_logscale=snake_logscale
                        )
                    )
                    for _ in range(self.num_layers)
                ]
            )
        elif activation == "snakebeta":
            self.activations = nn.ModuleList(
                [
                    Activation1d(
                        activation=activations.SnakeBeta(
                            channels, alpha_logscale=snake_logscale
                        )
                    )
                    for _ in range(self.num_layers)
                ]
            )
        else:
            raise NotImplementedError(
                "activation incorrectly specified. check the config file and look for 'activation'."
            )

    def forward(self, x):
        acts1, acts2 = self.activations[::2], self.activations[1::2]
        for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2):
            xt = a1(x)
            xt = c1(xt)
            xt = a2(xt)
            xt = c2(xt)
            x = xt + x

        return x

    def remove_weight_norm(self):
        for l in self.convs1:
            remove_weight_norm(l)
        for l in self.convs2:
            remove_weight_norm(l)


class AMPBlock2(torch.nn.Module):
    """
    AMPBlock applies Snake / SnakeBeta activation functions with trainable parameters that control periodicity, defined for each layer.
    Unlike AMPBlock1, AMPBlock2 does not contain extra Conv1d layers with fixed dilation=1

    Args:
        h (AttrDict): Hyperparameters.
        channels (int): Number of convolution channels.
        kernel_size (int): Size of the convolution kernel. Default is 3.
        dilation (tuple): Dilation rates for the convolutions. Each dilation layer has two convolutions. Default is (1, 3, 5).
        activation (str): Activation function type. Should be either 'snake' or 'snakebeta'. Default is None.
    """

    def __init__(
        self,
        channels: int,
        kernel_size: int = 3,
        dilation: tuple = (1, 3, 5),
        activation: str = None,
        use_cuda_kernel: bool = False,
        snake_logscale: bool = True
    ):
        super().__init__()
        
        self.convs = nn.ModuleList(
            [
                weight_norm(
                    Conv1d(
                        channels,
                        channels,
                        kernel_size,
                        stride=1,
                        dilation=d,
                        padding=get_padding(kernel_size, d),
                    )
                )
                for d in dilation
            ]
        )
        self.convs.apply(init_weights)

        self.num_layers = len(self.convs)  # Total number of conv layers

        # Select which Activation1d, lazy-load cuda version to ensure backward compatibility
        if use_cuda_kernel:
            from alias_free_activation.cuda.activation1d import (
                Activation1d as CudaActivation1d,
            )

            Activation1d = CudaActivation1d
        else:
            Activation1d = TorchActivation1d

        # Activation functions
        if activation == "snake":
            self.activations = nn.ModuleList(
                [
                    Activation1d(
                        activation=activations.Snake(
                            channels, alpha_logscale=snake_logscale
                        )
                    )
                    for _ in range(self.num_layers)
                ]
            )
        elif activation == "snakebeta":
            self.activations = nn.ModuleList(
                [
                    Activation1d(
                        activation=activations.SnakeBeta(
                            channels, alpha_logscale=snake_logscale
                        )
                    )
                    for _ in range(self.num_layers)
                ]
            )
        else:
            raise NotImplementedError(
                "activation incorrectly specified. check the config file and look for 'activation'."
            )

    def forward(self, x):
        for c, a in zip(self.convs, self.activations):
            xt = a(x)
            xt = c(xt)
            x = xt + x
        return x

    def remove_weight_norm(self):
        for l in self.convs:
            remove_weight_norm(l)


class BigVGAN(torch.nn.Module):
    """
    BigVGAN is a neural vocoder model that applies anti-aliased periodic activation for residual blocks (resblocks).
    New in BigVGAN-v2: it can optionally use optimized CUDA kernels for AMP (anti-aliased multi-periodicity) blocks.

    Args:
        use_cuda_kernel (bool): If set to True, loads optimized CUDA kernels for AMP. This should be used for inference only, as training is not supported with CUDA kernels.

    Note:
        - The `use_cuda_kernel` parameter should be used for inference only, as training with CUDA kernels is not supported.
        - Ensure that the activation function is correctly specified in the hyperparameters (h.activation).
    """

    def __init__(
            self,
            use_cuda_kernel: bool = False,
            num_mels: int = 80,
            upsample_initial_channel: int = 512,
            upsample_rates: list[int] = [5, 4, 3, 2, 2, 2],
            upsample_kernel_sizes: list[int] = [11, 8, 7, 4, 4, 4],
            resblock: str = "1",
            resblock_kernel_sizes: list[int] = [3, 7, 11],
            resblock_dilation_sizes: list[tuple] = [(1, 3, 5), (1, 3, 5), (1, 3, 5)],
            activation: str = "snakebeta",
            snake_logscale: bool = True,
            use_bias_at_final: bool = False,
            use_tanh_at_final: bool = False,
    ):
        super().__init__()
        self.use_cuda_kernel = use_cuda_kernel

        # Select which Activation1d, lazy-load cuda version to ensure backward compatibility
        if self.use_cuda_kernel:
            from alias_free_activation.cuda.activation1d import (
                Activation1d as CudaActivation1d,
            )

            Activation1d = CudaActivation1d
        else:
            Activation1d = TorchActivation1d

        self.num_kernels = len(resblock_kernel_sizes)
        self.num_upsamples = len(upsample_rates)

        # Pre-conv
        # for context smoothing, the padding=3 in the first layer conv_pre is removed
        self.conv_pre = weight_norm(
            Conv1d(num_mels, upsample_initial_channel, 7, 1, padding=0) 
        )

        # Define which AMPBlock to use. BigVGAN uses AMPBlock1 as default
        if resblock == "1":
            resblock_class = AMPBlock1
        elif resblock == "2":
            resblock_class = AMPBlock2
        else:
            raise ValueError(
                f"Incorrect resblock class specified in hyperparameters. Got {resblock}"
            )

        # Transposed conv-based upsamplers. does not apply anti-aliasing
        self.ups = nn.ModuleList()
        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
            self.ups.append(
                nn.ModuleList(
                    [
                        weight_norm(
                            ConvTranspose1d(
                                upsample_initial_channel // (2**i),
                                upsample_initial_channel // (2 ** (i + 1)),
                                k,
                                u,
                                padding=(k - u) // 2,
                            )
                        )
                    ]
                )
            )

        # Residual blocks using anti-aliased multi-periodicity composition modules (AMP)
        self.resblocks = nn.ModuleList()
        for i in range(len(self.ups)):
            ch = upsample_initial_channel // (2 ** (i + 1))
            for j, (k, d) in enumerate(
                zip(resblock_kernel_sizes, resblock_dilation_sizes)
            ):
                self.resblocks.append(
                    resblock_class(ch, k, d, activation=activation, use_cuda_kernel=self.use_cuda_kernel, snake_logscale=snake_logscale)
                )

        # Post-conv
        activation_post = (
            activations.Snake(ch, alpha_logscale=snake_logscale)
            if activation == "snake"
            else (
                activations.SnakeBeta(ch, alpha_logscale=snake_logscale)
                if activation == "snakebeta"
                else None
            )
        )
        if activation_post is None:
            raise NotImplementedError(
                "activation incorrectly specified. check the config file and look for 'activation'."
            )

        self.activation_post = Activation1d(activation=activation_post)

        # Whether to use bias for the final conv_post. Default to True for backward compatibility
        self.use_bias_at_final = use_bias_at_final
        self.conv_post = weight_norm(
            Conv1d(ch, 1, 7, 1, padding=3, bias=self.use_bias_at_final)
        )

        # Weight initialization
        for i in range(len(self.ups)):
            self.ups[i].apply(init_weights)
        self.conv_post.apply(init_weights)

        # Final tanh activation. Defaults to True for backward compatibility
        self.use_tanh_at_final = use_tanh_at_final

    def forward(self, x):
        # Pre-conv
        x = self.conv_pre(x)

        for i in range(self.num_upsamples):
            # Upsampling
            for i_up in range(len(self.ups[i])):
                x = self.ups[i][i_up](x)
            # AMP blocks
            xs = None
            for j in range(self.num_kernels):
                if xs is None:
                    xs = self.resblocks[i * self.num_kernels + j](x)
                else:
                    xs += self.resblocks[i * self.num_kernels + j](x)
            x = xs / self.num_kernels

        # Post-conv
        x = self.activation_post(x)
        x = self.conv_post(x)
        # Final tanh activation
        if self.use_tanh_at_final:
            x = torch.tanh(x)
        else:
            x = torch.clamp(x, min=-1.0, max=1.0)  # Bound the output to [-1, 1]

        return x

    def remove_weight_norm(self):
        try:
            print("Removing weight norm...")
            for l in self.ups:
                for l_i in l:
                    remove_weight_norm(l_i)
            for l in self.resblocks:
                l.remove_weight_norm()
            remove_weight_norm(self.conv_pre)
            remove_weight_norm(self.conv_post)
        except ValueError:
            print("[INFO] Model already removed weight norm. Skipping!")
            pass

    def _init_cuda_graph(self):
        pass

    @torch.inference_mode()
    def inference(self, x):
        x = self.forward(x)
        return x