v1.0

src/third_party/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
+
+import activations
+from utils import init_weights, get_padding
+from alias_free_activation.torch.act import Activation1d as TorchActivation1d
+from env import AttrDict
+
+from huggingface_hub import PyTorchModelHubMixin, hf_hub_download
+import sys
+sys.path.append(os.path.dirname(os.path.abspath(__file__)))
+
+
+def load_hparams_from_json(path) -> AttrDict:
+    with open(path) as f:
+        data = f.read()
+    return AttrDict(json.loads(data))
+
+
+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,
+        h: AttrDict,
+        channels: int,
+        kernel_size: int = 3,
+        dilation: tuple = (1, 3, 5),
+        activation: str = None,
+    ):
+        super().__init__()
+        
+        self.h = h
+
+        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 self.h.get("use_cuda_kernel", False):
+            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=h.snake_logscale
+                        )
+                    )
+                    for _ in range(self.num_layers)
+                ]
+            )
+        elif activation == "snakebeta":
+            self.activations = nn.ModuleList(
+                [
+                    Activation1d(
+                        activation=activations.SnakeBeta(
+                            channels, alpha_logscale=h.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,
+        h: AttrDict,
+        channels: int,
+        kernel_size: int = 3,
+        dilation: tuple = (1, 3, 5),
+        activation: str = None,
+    ):
+        super().__init__()
+        
+        self.h = h
+
+        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 self.h.get("use_cuda_kernel", False):
+            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=h.snake_logscale
+                        )
+                    )
+                    for _ in range(self.num_layers)
+                ]
+            )
+        elif activation == "snakebeta":
+            self.activations = nn.ModuleList(
+                [
+                    Activation1d(
+                        activation=activations.SnakeBeta(
+                            channels, alpha_logscale=h.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,
+    PyTorchModelHubMixin,
+    library_name="bigvgan",
+    repo_url="https://github.com/NVIDIA/BigVGAN",
+    docs_url="https://github.com/NVIDIA/BigVGAN/blob/main/README.md",
+    pipeline_tag="audio-to-audio",
+    license="mit",
+    tags=["neural-vocoder", "audio-generation", "arxiv:2206.04658"],
+):
+    """
+    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:
+        h (AttrDict): Hyperparameters.
+        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, h: AttrDict, use_cuda_kernel: bool = False):
+        super().__init__()
+        self.h = h
+        self.h["use_cuda_kernel"] = use_cuda_kernel
+
+        # Select which Activation1d, lazy-load cuda version to ensure backward compatibility
+        if self.h.get("use_cuda_kernel", False):
+            from alias_free_activation.cuda.activation1d import (
+                Activation1d as CudaActivation1d,
+            )
+
+            Activation1d = CudaActivation1d
+        else:
+            Activation1d = TorchActivation1d
+
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+
+        # Pre-conv
+        self.conv_pre = weight_norm(
+            Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3)
+        )
+
+        # Define which AMPBlock to use. BigVGAN uses AMPBlock1 as default
+        if h.resblock == "1":
+            resblock_class = AMPBlock1
+        elif h.resblock == "2":
+            resblock_class = AMPBlock2
+        else:
+            raise ValueError(
+                f"Incorrect resblock class specified in hyperparameters. Got {h.resblock}"
+            )
+
+        # Transposed conv-based upsamplers. does not apply anti-aliasing
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                nn.ModuleList(
+                    [
+                        weight_norm(
+                            ConvTranspose1d(
+                                h.upsample_initial_channel // (2**i),
+                                h.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 = h.upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
+            ):
+                self.resblocks.append(
+                    resblock_class(h, ch, k, d, activation=h.activation)
+                )
+
+        # Post-conv
+        activation_post = (
+            activations.Snake(ch, alpha_logscale=h.snake_logscale)
+            if h.activation == "snake"
+            else (
+                activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale)
+                if h.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 = h.get("use_bias_at_final", True)
+        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 = h.get("use_tanh_at_final", True)
+
+    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
+
+    # Additional methods for huggingface_hub support
+    def _save_pretrained(self, save_directory: Path) -> None:
+        """Save weights and config.json from a Pytorch model to a local directory."""
+
+        model_path = save_directory / "bigvgan_generator.pt"
+        torch.save({"generator": self.state_dict()}, model_path)
+
+        config_path = save_directory / "config.json"
+        with open(config_path, "w") as config_file:
+            json.dump(self.h, config_file, indent=4)
+
+    @classmethod
+    def _from_pretrained(
+        cls,
+        *,
+        model_id: str,
+        revision: str,
+        cache_dir: str,
+        force_download: bool,
+        proxies: Optional[Dict],
+        resume_download: bool,
+        local_files_only: bool,
+        token: Union[str, bool, None],
+        map_location: str = "cpu",  # Additional argument
+        strict: bool = False,  # Additional argument
+        use_cuda_kernel: bool = False,
+        **model_kwargs,
+    ):
+        """Load Pytorch pretrained weights and return the loaded model."""
+
+        # Download and load hyperparameters (h) used by BigVGAN
+        if os.path.isdir(model_id):
+            print("Loading config.json from local directory")
+            config_file = os.path.join(model_id, "config.json")
+        else:
+            config_file = hf_hub_download(
+                repo_id=model_id,
+                filename="config.json",
+                revision=revision,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                token=token,
+                local_files_only=local_files_only,
+            )
+        h = load_hparams_from_json(config_file)
+
+        # instantiate BigVGAN using h
+        if use_cuda_kernel:
+            print(
+                f"[WARNING] You have specified use_cuda_kernel=True during BigVGAN.from_pretrained(). Only inference is supported (training is not implemented)!"
+            )
+            print(
+                f"[WARNING] You need nvcc and ninja installed in your system that matches your PyTorch build is using to build the kernel. If not, the model will fail to initialize or generate incorrect waveform!"
+            )
+            print(
+                f"[WARNING] For detail, see the official GitHub repository: https://github.com/NVIDIA/BigVGAN?tab=readme-ov-file#using-custom-cuda-kernel-for-synthesis"
+            )
+        model = cls(h, use_cuda_kernel=use_cuda_kernel)
+
+        # Download and load pretrained generator weight
+        if os.path.isdir(model_id):
+            print("Loading weights from local directory")
+            model_file = os.path.join(model_id, "bigvgan_generator.pt")
+        else:
+            print(f"Loading weights from {model_id}")
+            model_file = hf_hub_download(
+                repo_id=model_id,
+                filename="bigvgan_generator.pt",
+                revision=revision,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                proxies=proxies,
+                resume_download=resume_download,
+                token=token,
+                local_files_only=local_files_only,
+            )
+
+        checkpoint_dict = torch.load(model_file, map_location=map_location)
+
+        try:
+            model.load_state_dict(checkpoint_dict["generator"])
+        except RuntimeError:
+            print(
+                f"[INFO] the pretrained checkpoint does not contain weight norm. Loading the checkpoint after removing weight norm!"
+            )
+            model.remove_weight_norm()
+            model.load_state_dict(checkpoint_dict["generator"])
+
+        return model

src/third_party/BigVGAN/configs/bigvgan_22khz_80band.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 32,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [4,4,2,2,2,2],
+    "upsample_kernel_sizes": [8,8,4,4,4,4],
+    "upsample_initial_channel": 1536,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "resolutions": [[1024, 120, 600], [2048, 240, 1200], [512, 50, 240]],
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+
+    "segment_size": 8192,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 22050,
+
+    "fmin": 0,
+    "fmax": 8000,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/configs/bigvgan_24khz_100band.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 32,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [4,4,2,2,2,2],
+    "upsample_kernel_sizes": [8,8,4,4,4,4],
+    "upsample_initial_channel": 1536,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "resolutions": [[1024, 120, 600], [2048, 240, 1200], [512, 50, 240]],
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+
+    "segment_size": 8192,
+    "num_mels": 100,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 24000,
+
+    "fmin": 0,
+    "fmax": 12000,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/configs/bigvgan_base_22khz_80band.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 32,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [8,8,2,2],
+    "upsample_kernel_sizes": [16,16,4,4],
+    "upsample_initial_channel": 512,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "resolutions": [[1024, 120, 600], [2048, 240, 1200], [512, 50, 240]],
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+
+    "segment_size": 8192,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 22050,
+
+    "fmin": 0,
+    "fmax": 8000,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/configs/bigvgan_base_24khz_100band.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 32,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [8,8,2,2],
+    "upsample_kernel_sizes": [16,16,4,4],
+    "upsample_initial_channel": 512,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "resolutions": [[1024, 120, 600], [2048, 240, 1200], [512, 50, 240]],
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+
+    "segment_size": 8192,
+    "num_mels": 100,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 24000,
+
+    "fmin": 0,
+    "fmax": 12000,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/configs/bigvgan_v2_22khz_80band_256x.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 4,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [4,4,2,2,2,2],
+    "upsample_kernel_sizes": [8,8,4,4,4,4],
+    "upsample_initial_channel": 1536,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "use_tanh_at_final": false,
+    "use_bias_at_final": false,
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "use_cqtd_instead_of_mrd": true,
+    "cqtd_filters": 128,
+    "cqtd_max_filters": 1024,
+    "cqtd_filters_scale": 1,
+    "cqtd_dilations": [1, 2, 4],
+    "cqtd_hop_lengths": [512, 256, 256],
+    "cqtd_n_octaves": [9, 9, 9],
+    "cqtd_bins_per_octaves": [24, 36, 48],
+
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+    
+    "use_multiscale_melloss": true,
+    "lambda_melloss": 15,
+
+    "clip_grad_norm": 500,
+
+    "segment_size": 65536,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 22050,
+
+    "fmin": 0,
+    "fmax": null,
+    "fmax_for_loss": null,
+    
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/configs/bigvgan_v2_22khz_80band_fmax8k_256x.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 4,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [4,4,2,2,2,2],
+    "upsample_kernel_sizes": [8,8,4,4,4,4],
+    "upsample_initial_channel": 1536,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "use_tanh_at_final": false,
+    "use_bias_at_final": false,
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "use_cqtd_instead_of_mrd": true,
+    "cqtd_filters": 128,
+    "cqtd_max_filters": 1024,
+    "cqtd_filters_scale": 1,
+    "cqtd_dilations": [1, 2, 4],
+    "cqtd_hop_lengths": [512, 256, 256],
+    "cqtd_n_octaves": [9, 9, 9],
+    "cqtd_bins_per_octaves": [24, 36, 48],
+    
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+
+    "use_multiscale_melloss": true,
+    "lambda_melloss": 15,
+
+    "clip_grad_norm": 500,
+
+    "segment_size": 65536,
+    "num_mels": 80,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 22050,
+
+    "fmin": 0,
+    "fmax": 8000,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/configs/bigvgan_v2_24khz_100band_256x.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 4,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [4,4,2,2,2,2],
+    "upsample_kernel_sizes": [8,8,4,4,4,4],
+    "upsample_initial_channel": 1536,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "use_tanh_at_final": false,
+    "use_bias_at_final": false,
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "use_cqtd_instead_of_mrd": true,
+    "cqtd_filters": 128,
+    "cqtd_max_filters": 1024,
+    "cqtd_filters_scale": 1,
+    "cqtd_dilations": [1, 2, 4],
+    "cqtd_hop_lengths": [512, 256, 256],
+    "cqtd_n_octaves": [9, 9, 9],
+    "cqtd_bins_per_octaves": [24, 36, 48],
+
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+
+    "use_multiscale_melloss": true,
+    "lambda_melloss": 15,
+
+    "clip_grad_norm": 500,
+
+    "segment_size": 65536,
+    "num_mels": 100,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 24000,
+
+    "fmin": 0,
+    "fmax": null,
+    "fmax_for_loss": null,
+    
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/configs/bigvgan_v2_44khz_128band_256x.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 4,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [4,4,2,2,2,2],
+    "upsample_kernel_sizes": [8,8,4,4,4,4],
+    "upsample_initial_channel": 1536,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "use_tanh_at_final": false,
+    "use_bias_at_final": false,
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "use_cqtd_instead_of_mrd": true,
+    "cqtd_filters": 128,
+    "cqtd_max_filters": 1024,
+    "cqtd_filters_scale": 1,
+    "cqtd_dilations": [1, 2, 4],
+    "cqtd_hop_lengths": [512, 256, 256],
+    "cqtd_n_octaves": [9, 9, 9],
+    "cqtd_bins_per_octaves": [24, 36, 48],
+
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+
+    "use_multiscale_melloss": true,
+    "lambda_melloss": 15,
+
+    "clip_grad_norm": 500,
+
+    "segment_size": 65536,
+    "num_mels": 128,
+    "num_freq": 1025,
+    "n_fft": 1024,
+    "hop_size": 256,
+    "win_size": 1024,
+
+    "sampling_rate": 44100,
+
+    "fmin": 0,
+    "fmax": null,
+    "fmax_for_loss": null,
+    
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/configs/bigvgan_v2_44khz_128band_512x.json

+{
+    "resblock": "1",
+    "num_gpus": 0,
+    "batch_size": 4,
+    "learning_rate": 0.0001,
+    "adam_b1": 0.8,
+    "adam_b2": 0.99,
+    "lr_decay": 0.9999996,
+    "seed": 1234,
+
+    "upsample_rates": [8,4,2,2,2,2],
+    "upsample_kernel_sizes": [16,8,4,4,4,4],
+    "upsample_initial_channel": 1536,
+    "resblock_kernel_sizes": [3,7,11],
+    "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+
+    "use_tanh_at_final": false,
+    "use_bias_at_final": false,
+
+    "activation": "snakebeta",
+    "snake_logscale": true,
+
+    "use_cqtd_instead_of_mrd": true,
+    "cqtd_filters": 128,
+    "cqtd_max_filters": 1024,
+    "cqtd_filters_scale": 1,
+    "cqtd_dilations": [1, 2, 4],
+    "cqtd_hop_lengths": [512, 256, 256],
+    "cqtd_n_octaves": [9, 9, 9],
+    "cqtd_bins_per_octaves": [24, 36, 48],
+
+    "mpd_reshapes": [2, 3, 5, 7, 11],
+    "use_spectral_norm": false,
+    "discriminator_channel_mult": 1,
+
+    "use_multiscale_melloss": true,
+    "lambda_melloss": 15,
+
+    "clip_grad_norm": 500,
+
+    "segment_size": 65536,
+    "num_mels": 128,
+    "num_freq": 2049,
+    "n_fft": 2048,
+    "hop_size": 512,
+    "win_size": 2048,
+
+    "sampling_rate": 44100,
+
+    "fmin": 0,
+    "fmax": null,
+    "fmax_for_loss": null,
+
+    "num_workers": 4,
+
+    "dist_config": {
+        "dist_backend": "nccl",
+        "dist_url": "tcp://localhost:54321",
+        "world_size": 1
+    }
+}

src/third_party/BigVGAN/demo/app.py

+# Copyright (c) 2024 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+import spaces
+import gradio as gr
+import pandas as pd
+import torch
+import os
+import sys
+
+# to import modules from parent_dir
+parent_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
+sys.path.append(parent_dir)
+
+from meldataset import get_mel_spectrogram, MAX_WAV_VALUE
+from bigvgan import BigVGAN
+import librosa
+import numpy as np
+from utils import plot_spectrogram
+import PIL
+
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+    torch.backends.cudnn.benchmark = False
+    print(f"using GPU")
+else:
+    device = torch.device("cpu")
+    print(f"using CPU")
+
+
+def inference_gradio(input, model_choice):  # Input is audio waveform in [T, channel]
+    sr, audio = input  # Unpack input to sampling rate and audio itself
+    audio = np.transpose(audio)  # Transpose to [channel, T] for librosa
+    audio = audio / MAX_WAV_VALUE  # Convert int16 to float range used by BigVGAN
+
+    model = dict_model[model_choice]
+
+    if sr != model.h.sampling_rate:  # Convert audio to model's sampling rate
+        audio = librosa.resample(audio, orig_sr=sr, target_sr=model.h.sampling_rate)
+    if len(audio.shape) == 2:  # Stereo
+        audio = librosa.to_mono(audio)  # Convert to mono if stereo
+    audio = librosa.util.normalize(audio) * 0.95
+
+    output, spec_gen = inference_model(
+        audio, model
+    )  # Output is generated audio in ndarray, int16
+
+    spec_plot_gen = plot_spectrogram(spec_gen)
+
+    output_audio = (model.h.sampling_rate, output)  # Tuple for gr.Audio output
+
+    buffer = spec_plot_gen.canvas.buffer_rgba()
+    output_image = PIL.Image.frombuffer(
+        "RGBA", spec_plot_gen.canvas.get_width_height(), buffer, "raw", "RGBA", 0, 1
+    )
+
+    return output_audio, output_image
+
+
+@spaces.GPU(duration=120)
+def inference_model(audio_input, model):
+    # Load model to device
+    model.to(device)
+
+    with torch.inference_mode():
+        wav = torch.FloatTensor(audio_input)
+        # Compute mel spectrogram from the ground truth audio
+        spec_gt = get_mel_spectrogram(wav.unsqueeze(0), model.h).to(device)
+
+        y_g_hat = model(spec_gt)
+
+        audio_gen = y_g_hat.squeeze().cpu()
+        spec_gen = get_mel_spectrogram(audio_gen.unsqueeze(0), model.h)
+        audio_gen = audio_gen.numpy()  # [T], float [-1, 1]
+        audio_gen = (audio_gen * MAX_WAV_VALUE).astype("int16")  # [T], int16
+        spec_gen = spec_gen.squeeze().numpy()  # [C, T_frame]
+
+    # Unload to CPU
+    model.to("cpu")
+    # Delete GPU tensor
+    del spec_gt, y_g_hat
+
+    return audio_gen, spec_gen
+
+
+css = """
+        a {
+            color: inherit;
+            text-decoration: underline;
+        }
+        .gradio-container {
+            font-family: 'IBM Plex Sans', sans-serif;
+        }
+        .gr-button {
+            color: white;
+            border-color: #000000;
+            background: #000000;
+        }
+        input[type='range'] {
+            accent-color: #000000;
+        }
+        .dark input[type='range'] {
+            accent-color: #dfdfdf;
+        }
+        .container {
+            max-width: 730px;
+            margin: auto;
+            padding-top: 1.5rem;
+        }
+        #gallery {
+            min-height: 22rem;
+            margin-bottom: 15px;
+            margin-left: auto;
+            margin-right: auto;
+            border-bottom-right-radius: .5rem !important;
+            border-bottom-left-radius: .5rem !important;
+        }
+        #gallery>div>.h-full {
+            min-height: 20rem;
+        }
+        .details:hover {
+            text-decoration: underline;
+        }
+        .gr-button {
+            white-space: nowrap;
+        }
+        .gr-button:focus {
+            border-color: rgb(147 197 253 / var(--tw-border-opacity));
+            outline: none;
+            box-shadow: var(--tw-ring-offset-shadow), var(--tw-ring-shadow), var(--tw-shadow, 0 0 #0000);
+            --tw-border-opacity: 1;
+            --tw-ring-offset-shadow: var(--tw-ring-inset) 0 0 0 var(--tw-ring-offset-width) var(--tw-ring-offset-color);
+            --tw-ring-shadow: var(--tw-ring-inset) 0 0 0 calc(3px var(--tw-ring-offset-width)) var(--tw-ring-color);
+            --tw-ring-color: rgb(191 219 254 / var(--tw-ring-opacity));
+            --tw-ring-opacity: .5;
+        }
+        #advanced-btn {
+            font-size: .7rem !important;
+            line-height: 19px;
+            margin-top: 12px;
+            margin-bottom: 12px;
+            padding: 2px 8px;
+            border-radius: 14px !important;
+        }
+        #advanced-options {
+            margin-bottom: 20px;
+        }
+        .footer {
+            margin-bottom: 45px;
+            margin-top: 35px;
+            text-align: center;
+            border-bottom: 1px solid #e5e5e5;
+        }
+        .footer>p {
+            font-size: .8rem;
+            display: inline-block;
+            padding: 0 10px;
+            transform: translateY(10px);
+            background: white;
+        }
+        .dark .footer {
+            border-color: #303030;
+        }
+        .dark .footer>p {
+            background: #0b0f19;
+        }
+        .acknowledgments h4{
+            margin: 1.25em 0 .25em 0;
+            font-weight: bold;
+            font-size: 115%;
+        }
+        #container-advanced-btns{
+            display: flex;
+            flex-wrap: wrap;
+            justify-content: space-between;
+            align-items: center;
+        }
+        .animate-spin {
+            animation: spin 1s linear infinite;
+        }
+        @keyframes spin {
+            from {
+                transform: rotate(0deg);
+            }
+            to {
+                transform: rotate(360deg);
+            }
+        }
+        #share-btn-container {
+            display: flex; padding-left: 0.5rem !important; padding-right: 0.5rem !important; background-color: #000000; justify-content: center; align-items: center; border-radius: 9999px !important; width: 13rem;
+            margin-top: 10px;
+            margin-left: auto;
+        }
+        #share-btn {
+            all: initial; color: #ffffff;font-weight: 600; cursor:pointer; font-family: 'IBM Plex Sans', sans-serif; margin-left: 0.5rem !important; padding-top: 0.25rem !important; padding-bottom: 0.25rem !important;right:0;
+        }
+        #share-btn * {
+            all: unset;
+        }
+        #share-btn-container div:nth-child(-n+2){
+            width: auto !important;
+            min-height: 0px !important;
+        }
+        #share-btn-container .wrap {
+            display: none !important;
+        }
+        .gr-form{
+            flex: 1 1 50%; border-top-right-radius: 0; border-bottom-right-radius: 0;
+        }
+        #prompt-container{
+            gap: 0;
+        }
+        #generated_id{
+            min-height: 700px
+        }
+        #setting_id{
+          margin-bottom: 12px;
+          text-align: center;
+          font-weight: 900;
+        }
+"""
+
+# Script for loading the models
+
+LIST_MODEL_ID = [
+    "bigvgan_24khz_100band",
+    "bigvgan_base_24khz_100band",
+    "bigvgan_22khz_80band",
+    "bigvgan_base_22khz_80band",
+    "bigvgan_v2_22khz_80band_256x",
+    "bigvgan_v2_22khz_80band_fmax8k_256x",
+    "bigvgan_v2_24khz_100band_256x",
+    "bigvgan_v2_44khz_128band_256x",
+    "bigvgan_v2_44khz_128band_512x",
+]
+
+dict_model = {}
+dict_config = {}
+
+for model_name in LIST_MODEL_ID:
+
+    generator = BigVGAN.from_pretrained("nvidia/" + model_name)
+    generator.remove_weight_norm()
+    generator.eval()
+
+    dict_model[model_name] = generator
+    dict_config[model_name] = generator.h
+
+# Script for Gradio UI
+
+iface = gr.Blocks(css=css, title="BigVGAN - Demo")
+
+with iface:
+    gr.HTML(
+        """
+        <div style="text-align: center; max-width: 900px; margin: 0 auto;">
+            <div
+            style="
+                display: inline-flex;
+                align-items: center;
+                gap: 0.8rem;
+                font-size: 1.5rem;
+            "
+            >
+            <h1 style="font-weight: 700; margin-bottom: 7px; line-height: normal;">
+                BigVGAN: A Universal Neural Vocoder with Large-Scale Training
+            </h1>
+            </div>
+            <p style="margin-bottom: 10px; font-size: 125%">
+            <a href="https://arxiv.org/abs/2206.04658">[Paper]</a>  <a href="https://github.com/NVIDIA/BigVGAN">[Code]</a>  <a href="https://bigvgan-demo.github.io/">[Demo]</a>  <a href="https://research.nvidia.com/labs/adlr/projects/bigvgan/">[Project page]</a>
+            </p>
+        </div>
+        """
+    )
+    gr.HTML(
+        """
+        <div>
+        <h3>News</h3>
+        <p>[Jul 2024] We release BigVGAN-v2 along with pretrained checkpoints. Below are the highlights:</p>
+        <ul>
+            <li>Custom CUDA kernel for inference: we provide a fused upsampling + activation kernel written in CUDA for accelerated inference speed. Our test shows 1.5 - 3x faster speed on a single A100 GPU.</li>
+            <li>Improved discriminator and loss: BigVGAN-v2 is trained using a <a href="https://arxiv.org/abs/2311.14957" target="_blank">multi-scale sub-band CQT discriminator</a> and a <a href="https://arxiv.org/abs/2306.06546" target="_blank">multi-scale mel spectrogram loss</a>.</li>
+            <li>Larger training data: BigVGAN-v2 is trained using datasets containing diverse audio types, including speech in multiple languages, environmental sounds, and instruments.</li>
+            <li>We provide pretrained checkpoints of BigVGAN-v2 using diverse audio configurations, supporting up to 44 kHz sampling rate and 512x upsampling ratio. See the table below for the link.</li>
+        </ul>
+        </div>
+        """
+    )
+    gr.HTML(
+        """
+        <div>
+        <h3>Model Overview</h3>
+        BigVGAN is a universal neural vocoder model that generates audio waveforms using mel spectrogram as inputs.
+        <center><img src="https://user-images.githubusercontent.com/15963413/218609148-881e39df-33af-4af9-ab95-1427c4ebf062.png" width="800" style="margin-top: 20px; border-radius: 15px;"></center>
+        </div>
+        """
+    )
+    with gr.Accordion("Input"):
+
+        model_choice = gr.Dropdown(
+            label="Select the model to use",
+            info="The default model is bigvgan_v2_24khz_100band_256x",
+            value="bigvgan_v2_24khz_100band_256x",
+            choices=[m for m in LIST_MODEL_ID],
+            interactive=True,
+        )
+
+        audio_input = gr.Audio(
+            label="Input Audio", elem_id="input-audio", interactive=True
+        )
+
+    button = gr.Button("Submit")
+
+    with gr.Accordion("Output"):
+        with gr.Column():
+            output_audio = gr.Audio(label="Output Audio", elem_id="output-audio")
+            output_image = gr.Image(
+                label="Output Mel Spectrogram", elem_id="output-image-gen"
+            )
+
+    button.click(
+        inference_gradio,
+        inputs=[audio_input, model_choice],
+        outputs=[output_audio, output_image],
+        concurrency_limit=10,
+    )
+
+    gr.Examples(
+        [
+            [
+                os.path.join(os.path.dirname(__file__), "examples/jensen_24k.wav"),
+                "bigvgan_v2_24khz_100band_256x",
+            ],
+            [
+                os.path.join(os.path.dirname(__file__), "examples/libritts_24k.wav"),
+                "bigvgan_v2_24khz_100band_256x",
+            ],
+            [
+                os.path.join(os.path.dirname(__file__), "examples/queen_24k.wav"),
+                "bigvgan_v2_24khz_100band_256x",
+            ],
+            [
+                os.path.join(os.path.dirname(__file__), "examples/dance_24k.wav"),
+                "bigvgan_v2_24khz_100band_256x",
+            ],
+            [
+                os.path.join(os.path.dirname(__file__), "examples/megalovania_24k.wav"),
+                "bigvgan_v2_24khz_100band_256x",
+            ],
+            [
+                os.path.join(os.path.dirname(__file__), "examples/hifitts_44k.wav"),
+                "bigvgan_v2_44khz_128band_256x",
+            ],
+            [
+                os.path.join(os.path.dirname(__file__), "examples/musdbhq_44k.wav"),
+                "bigvgan_v2_44khz_128band_256x",
+            ],
+            [
+                os.path.join(os.path.dirname(__file__), "examples/musiccaps1_44k.wav"),
+                "bigvgan_v2_44khz_128band_256x",
+            ],
+            [
+                os.path.join(os.path.dirname(__file__), "examples/musiccaps2_44k.wav"),
+                "bigvgan_v2_44khz_128band_256x",
+            ],
+        ],
+        fn=inference_gradio,
+        inputs=[audio_input, model_choice],
+        outputs=[output_audio, output_image],
+    )
+
+    # Define the data for the table
+    data = {
+        "Model Name": [
+            "bigvgan_v2_44khz_128band_512x",
+            "bigvgan_v2_44khz_128band_256x",
+            "bigvgan_v2_24khz_100band_256x",
+            "bigvgan_v2_22khz_80band_256x",
+            "bigvgan_v2_22khz_80band_fmax8k_256x",
+            "bigvgan_24khz_100band",
+            "bigvgan_base_24khz_100band",
+            "bigvgan_22khz_80band",
+            "bigvgan_base_22khz_80band",
+        ],
+        "Sampling Rate": [
+            "44 kHz",
+            "44 kHz",
+            "24 kHz",
+            "22 kHz",
+            "22 kHz",
+            "24 kHz",
+            "24 kHz",
+            "22 kHz",
+            "22 kHz",
+        ],
+        "Mel band": [128, 128, 100, 80, 80, 100, 100, 80, 80],
+        "fmax": [22050, 22050, 12000, 11025, 8000, 12000, 12000, 8000, 8000],
+        "Upsampling Ratio": [512, 256, 256, 256, 256, 256, 256, 256, 256],
+        "Parameters": [
+            "122M",
+            "112M",
+            "112M",
+            "112M",
+            "112M",
+            "112M",
+            "14M",
+            "112M",
+            "14M",
+        ],
+        "Dataset": [
+            "Large-scale Compilation",
+            "Large-scale Compilation",
+            "Large-scale Compilation",
+            "Large-scale Compilation",
+            "Large-scale Compilation",
+            "LibriTTS",
+            "LibriTTS",
+            "LibriTTS + VCTK + LJSpeech",
+            "LibriTTS + VCTK + LJSpeech",
+        ],
+        "Fine-Tuned": ["No", "No", "No", "No", "No", "No", "No", "No", "No"],
+    }
+
+    base_url = "https://huggingface.co/nvidia/"
+
+    df = pd.DataFrame(data)
+    df["Model Name"] = df["Model Name"].apply(
+        lambda x: f'<a href="{base_url}{x}">{x}</a>'
+    )
+
+    html_table = gr.HTML(
+        f"""
+        <div style="text-align: center;">
+            {df.to_html(index=False, escape=False, classes='border="1" cellspacing="0" cellpadding="5" style="margin-left: auto; margin-right: auto;')}
+            <p><b>NOTE: The v1 models are trained using speech audio datasets ONLY! (24kHz models: LibriTTS, 22kHz models: LibriTTS + VCTK + LJSpeech).</b></p>
+        </div>
+        """
+    )
+
+iface.queue()
+iface.launch()