Push audiocraft

audiocraft/__init__.py

@@ -0,0 +1,26 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""
+AudioCraft is a general framework for training audio generative models.
+At the moment we provide the training code for:
+
+- [MusicGen](https://arxiv.org/abs/2306.05284), a state-of-the-art
+    text-to-music and melody+text autoregressive generative model.
+    For the solver, see `audiocraft.solvers.musicgen.MusicGenSolver`, and for the model,
+    `audiocraft.models.musicgen.MusicGen`.
+- [AudioGen](https://arxiv.org/abs/2209.15352), a state-of-the-art
+    text-to-general-audio generative model.
+- [EnCodec](https://arxiv.org/abs/2210.13438), efficient and high fidelity
+    neural audio codec which provides an excellent tokenizer for autoregressive language models.
+    See `audiocraft.solvers.compression.CompressionSolver`, and `audiocraft.models.encodec.EncodecModel`.
+- [MultiBandDiffusion](TODO), alternative diffusion-based decoder compatible with EnCodec that
+    improves the perceived quality and reduces the artifacts coming from adversarial decoders.
+"""
+
+# flake8: noqa
+from . import data, modules, models
+
+__version__ = '1.2.0a2'

audiocraft/adversarial/__init__.py

@@ -0,0 +1,22 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Adversarial losses and discriminator architectures."""
+
+# flake8: noqa
+from .discriminators import (
+    MultiPeriodDiscriminator,
+    MultiScaleDiscriminator,
+    MultiScaleSTFTDiscriminator
+)
+from .losses import (
+    AdversarialLoss,
+    AdvLossType,
+    get_adv_criterion,
+    get_fake_criterion,
+    get_real_criterion,
+    FeatLossType,
+    FeatureMatchingLoss
+)

audiocraft/adversarial/discriminators/__init__.py

@@ -0,0 +1,10 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# flake8: noqa
+from .mpd import MultiPeriodDiscriminator
+from .msd import MultiScaleDiscriminator
+from .msstftd import MultiScaleSTFTDiscriminator

audiocraft/adversarial/discriminators/base.py

@@ -0,0 +1,34 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from abc import ABC, abstractmethod
+import typing as tp
+
+import torch
+import torch.nn as nn
+
+
+FeatureMapType = tp.List[torch.Tensor]
+LogitsType = torch.Tensor
+MultiDiscriminatorOutputType = tp.Tuple[tp.List[LogitsType], tp.List[FeatureMapType]]
+
+
+class MultiDiscriminator(ABC, nn.Module):
+    """Base implementation for discriminators composed of sub-discriminators acting at different scales.
+    """
+    def __init__(self):
+        super().__init__()
+
+    @abstractmethod
+    def forward(self, x: torch.Tensor) -> MultiDiscriminatorOutputType:
+        ...
+
+    @property
+    @abstractmethod
+    def num_discriminators(self) -> int:
+        """Number of discriminators.
+        """
+        ...

audiocraft/adversarial/discriminators/mpd.py

@@ -0,0 +1,106 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import typing as tp
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ...modules import NormConv2d
+from .base import MultiDiscriminator, MultiDiscriminatorOutputType
+
+
+def get_padding(kernel_size: int, dilation: int = 1) -> int:
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+class PeriodDiscriminator(nn.Module):
+    """Period sub-discriminator.
+
+    Args:
+        period (int): Period between samples of audio.
+        in_channels (int): Number of input channels.
+        out_channels (int): Number of output channels.
+        n_layers (int): Number of convolutional layers.
+        kernel_sizes (list of int): Kernel sizes for convolutions.
+        stride (int): Stride for convolutions.
+        filters (int): Initial number of filters in convolutions.
+        filters_scale (int): Multiplier of number of filters as we increase depth.
+        max_filters (int): Maximum number of filters.
+        norm (str): Normalization method.
+        activation (str): Activation function.
+        activation_params (dict): Parameters to provide to the activation function.
+    """
+    def __init__(self, period: int, in_channels: int = 1, out_channels: int = 1,
+                 n_layers: int = 5, kernel_sizes: tp.List[int] = [5, 3], stride: int = 3,
+                 filters: int = 8, filters_scale: int = 4, max_filters: int = 1024,
+                 norm: str = 'weight_norm', activation: str = 'LeakyReLU',
+                 activation_params: dict = {'negative_slope': 0.2}):
+        super().__init__()
+        self.period = period
+        self.n_layers = n_layers
+        self.activation = getattr(torch.nn, activation)(**activation_params)
+        self.convs = nn.ModuleList()
+        in_chs = in_channels
+        for i in range(self.n_layers):
+            out_chs = min(filters * (filters_scale ** (i + 1)), max_filters)
+            eff_stride = 1 if i == self.n_layers - 1 else stride
+            self.convs.append(NormConv2d(in_chs, out_chs, kernel_size=(kernel_sizes[0], 1), stride=(eff_stride, 1),
+                                         padding=((kernel_sizes[0] - 1) // 2, 0), norm=norm))
+            in_chs = out_chs
+        self.conv_post = NormConv2d(in_chs, out_channels, kernel_size=(kernel_sizes[1], 1), stride=1,
+                                    padding=((kernel_sizes[1] - 1) // 2, 0), norm=norm)
+
+    def forward(self, x: torch.Tensor):
+        fmap = []
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), 'reflect')
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for conv in self.convs:
+            x = conv(x)
+            x = self.activation(x)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        # x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(MultiDiscriminator):
+    """Multi-Period (MPD) Discriminator.
+
+    Args:
+        in_channels (int): Number of input channels.
+        out_channels (int): Number of output channels.
+        periods (Sequence[int]): Periods between samples of audio for the sub-discriminators.
+        **kwargs: Additional args for `PeriodDiscriminator`
+    """
+    def __init__(self, in_channels: int = 1, out_channels: int = 1,
+                 periods: tp.Sequence[int] = [2, 3, 5, 7, 11], **kwargs):
+        super().__init__()
+        self.discriminators = nn.ModuleList([
+            PeriodDiscriminator(p, in_channels, out_channels, **kwargs) for p in periods
+        ])
+
+    @property
+    def num_discriminators(self):
+        return len(self.discriminators)
+
+    def forward(self, x: torch.Tensor) -> MultiDiscriminatorOutputType:
+        logits = []
+        fmaps = []
+        for disc in self.discriminators:
+            logit, fmap = disc(x)
+            logits.append(logit)
+            fmaps.append(fmap)
+        return logits, fmaps

audiocraft/adversarial/discriminators/msd.py

@@ -0,0 +1,126 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import typing as tp
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from ...modules import NormConv1d
+from .base import MultiDiscriminator, MultiDiscriminatorOutputType
+
+
+class ScaleDiscriminator(nn.Module):
+    """Waveform sub-discriminator.
+
+    Args:
+        in_channels (int): Number of input channels.
+        out_channels (int): Number of output channels.
+        kernel_sizes (Sequence[int]): Kernel sizes for first and last convolutions.
+        filters (int): Number of initial filters for convolutions.
+        max_filters (int): Maximum number of filters.
+        downsample_scales (Sequence[int]): Scale for downsampling implemented as strided convolutions.
+        inner_kernel_sizes (Sequence[int] or None): Kernel sizes for inner convolutions.
+        groups (Sequence[int] or None): Groups for inner convolutions.
+        strides (Sequence[int] or None): Strides for inner convolutions.
+        paddings (Sequence[int] or None): Paddings for inner convolutions.
+        norm (str): Normalization method.
+        activation (str): Activation function.
+        activation_params (dict): Parameters to provide to the activation function.
+        pad (str): Padding for initial convolution.
+        pad_params (dict): Parameters to provide to the padding module.
+    """
+    def __init__(self, in_channels=1, out_channels=1, kernel_sizes: tp.Sequence[int] = [5, 3],
+                 filters: int = 16, max_filters: int = 1024, downsample_scales: tp.Sequence[int] = [4, 4, 4, 4],
+                 inner_kernel_sizes: tp.Optional[tp.Sequence[int]] = None, groups: tp.Optional[tp.Sequence[int]] = None,
+                 strides: tp.Optional[tp.Sequence[int]] = None, paddings: tp.Optional[tp.Sequence[int]] = None,
+                 norm: str = 'weight_norm', activation: str = 'LeakyReLU',
+                 activation_params: dict = {'negative_slope': 0.2}, pad: str = 'ReflectionPad1d',
+                 pad_params: dict = {}):
+        super().__init__()
+        assert len(kernel_sizes) == 2
+        assert kernel_sizes[0] % 2 == 1
+        assert kernel_sizes[1] % 2 == 1
+        assert (inner_kernel_sizes is None or len(inner_kernel_sizes) == len(downsample_scales))
+        assert (groups is None or len(groups) == len(downsample_scales))
+        assert (strides is None or len(strides) == len(downsample_scales))
+        assert (paddings is None or len(paddings) == len(downsample_scales))
+        self.activation = getattr(torch.nn, activation)(**activation_params)
+        self.convs = nn.ModuleList()
+        self.convs.append(
+            nn.Sequential(
+                getattr(torch.nn, pad)((np.prod(kernel_sizes) - 1) // 2, **pad_params),
+                NormConv1d(in_channels, filters, kernel_size=np.prod(kernel_sizes), stride=1, norm=norm)
+            )
+        )
+
+        in_chs = filters
+        for i, downsample_scale in enumerate(downsample_scales):
+            out_chs = min(in_chs * downsample_scale, max_filters)
+            default_kernel_size = downsample_scale * 10 + 1
+            default_stride = downsample_scale
+            default_padding = (default_kernel_size - 1) // 2
+            default_groups = in_chs // 4
+            self.convs.append(
+                NormConv1d(in_chs, out_chs,
+                           kernel_size=inner_kernel_sizes[i] if inner_kernel_sizes else default_kernel_size,
+                           stride=strides[i] if strides else default_stride,
+                           groups=groups[i] if groups else default_groups,
+                           padding=paddings[i] if paddings else default_padding,
+                           norm=norm))
+            in_chs = out_chs
+
+        out_chs = min(in_chs * 2, max_filters)
+        self.convs.append(NormConv1d(in_chs, out_chs, kernel_size=kernel_sizes[0], stride=1,
+                                     padding=(kernel_sizes[0] - 1) // 2, norm=norm))
+        self.conv_post = NormConv1d(out_chs, out_channels, kernel_size=kernel_sizes[1], stride=1,
+                                    padding=(kernel_sizes[1] - 1) // 2, norm=norm)
+
+    def forward(self, x: torch.Tensor):
+        fmap = []
+        for layer in self.convs:
+            x = layer(x)
+            x = self.activation(x)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        # x = torch.flatten(x, 1, -1)
+        return x, fmap
+
+
+class MultiScaleDiscriminator(MultiDiscriminator):
+    """Multi-Scale (MSD) Discriminator,
+
+    Args:
+        in_channels (int): Number of input channels.
+        out_channels (int): Number of output channels.
+        downsample_factor (int): Downsampling factor between the different scales.
+        scale_norms (Sequence[str]): Normalization for each sub-discriminator.
+        **kwargs: Additional args for ScaleDiscriminator.
+    """
+    def __init__(self, in_channels: int = 1, out_channels: int = 1, downsample_factor: int = 2,
+                 scale_norms: tp.Sequence[str] = ['weight_norm', 'weight_norm', 'weight_norm'], **kwargs):
+        super().__init__()
+        self.discriminators = nn.ModuleList([
+            ScaleDiscriminator(in_channels, out_channels, norm=norm, **kwargs) for norm in scale_norms
+        ])
+        self.downsample = nn.AvgPool1d(downsample_factor * 2, downsample_factor, padding=downsample_factor)
+
+    @property
+    def num_discriminators(self):
+        return len(self.discriminators)
+
+    def forward(self, x: torch.Tensor) -> MultiDiscriminatorOutputType:
+        logits = []
+        fmaps = []
+        for i, disc in enumerate(self.discriminators):
+            if i != 0:
+                self.downsample(x)
+            logit, fmap = disc(x)
+            logits.append(logit)
+            fmaps.append(fmap)
+        return logits, fmaps

audiocraft/adversarial/discriminators/msstftd.py

@@ -0,0 +1,134 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import typing as tp
+
+import torchaudio
+import torch
+from torch import nn
+from einops import rearrange
+
+from ...modules import NormConv2d
+from .base import MultiDiscriminator, MultiDiscriminatorOutputType
+
+
+def get_2d_padding(kernel_size: tp.Tuple[int, int], dilation: tp.Tuple[int, int] = (1, 1)):
+    return (((kernel_size[0] - 1) * dilation[0]) // 2, ((kernel_size[1] - 1) * dilation[1]) // 2)
+
+
+class DiscriminatorSTFT(nn.Module):
+    """STFT sub-discriminator.
+
+    Args:
+        filters (int): Number of filters in convolutions.
+        in_channels (int): Number of input channels.
+        out_channels (int): Number of output channels.
+        n_fft (int): Size of FFT for each scale.
+        hop_length (int): Length of hop between STFT windows for each scale.
+        kernel_size (tuple of int): Inner Conv2d kernel sizes.
+        stride (tuple of int): Inner Conv2d strides.
+        dilations (list of int): Inner Conv2d dilation on the time dimension.
+        win_length (int): Window size for each scale.
+        normalized (bool): Whether to normalize by magnitude after stft.
+        norm (str): Normalization method.
+        activation (str): Activation function.
+        activation_params (dict): Parameters to provide to the activation function.
+        growth (int): Growth factor for the filters.
+    """
+    def __init__(self, filters: int, in_channels: int = 1, out_channels: int = 1,
+                 n_fft: int = 1024, hop_length: int = 256, win_length: int = 1024, max_filters: int = 1024,
+                 filters_scale: int = 1, kernel_size: tp.Tuple[int, int] = (3, 9), dilations: tp.List = [1, 2, 4],
+                 stride: tp.Tuple[int, int] = (1, 2), normalized: bool = True, norm: str = 'weight_norm',
+                 activation: str = 'LeakyReLU', activation_params: dict = {'negative_slope': 0.2}):
+        super().__init__()
+        assert len(kernel_size) == 2
+        assert len(stride) == 2
+        self.filters = filters
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.n_fft = n_fft
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.normalized = normalized
+        self.activation = getattr(torch.nn, activation)(**activation_params)
+        self.spec_transform = torchaudio.transforms.Spectrogram(
+            n_fft=self.n_fft, hop_length=self.hop_length, win_length=self.win_length, window_fn=torch.hann_window,
+            normalized=self.normalized, center=False, pad_mode=None, power=None)
+        spec_channels = 2 * self.in_channels
+        self.convs = nn.ModuleList()
+        self.convs.append(
+            NormConv2d(spec_channels, self.filters, kernel_size=kernel_size, padding=get_2d_padding(kernel_size))
+        )
+        in_chs = min(filters_scale * self.filters, max_filters)
+        for i, dilation in enumerate(dilations):
+            out_chs = min((filters_scale ** (i + 1)) * self.filters, max_filters)
+            self.convs.append(NormConv2d(in_chs, out_chs, kernel_size=kernel_size, stride=stride,
+                                         dilation=(dilation, 1), padding=get_2d_padding(kernel_size, (dilation, 1)),
+                                         norm=norm))
+            in_chs = out_chs
+        out_chs = min((filters_scale ** (len(dilations) + 1)) * self.filters, max_filters)
+        self.convs.append(NormConv2d(in_chs, out_chs, kernel_size=(kernel_size[0], kernel_size[0]),
+                                     padding=get_2d_padding((kernel_size[0], kernel_size[0])),
+                                     norm=norm))
+        self.conv_post = NormConv2d(out_chs, self.out_channels,
+                                    kernel_size=(kernel_size[0], kernel_size[0]),
+                                    padding=get_2d_padding((kernel_size[0], kernel_size[0])),
+                                    norm=norm)
+
+    def forward(self, x: torch.Tensor):
+        fmap = []
+        z = self.spec_transform(x)  # [B, 2, Freq, Frames, 2]
+        z = torch.cat([z.real, z.imag], dim=1)
+        z = rearrange(z, 'b c w t -> b c t w')
+        for i, layer in enumerate(self.convs):
+            z = layer(z)
+            z = self.activation(z)
+            fmap.append(z)
+        z = self.conv_post(z)
+        return z, fmap
+
+
+class MultiScaleSTFTDiscriminator(MultiDiscriminator):
+    """Multi-Scale STFT (MS-STFT) discriminator.
+
+    Args:
+        filters (int): Number of filters in convolutions.
+        in_channels (int): Number of input channels.
+        out_channels (int): Number of output channels.
+        sep_channels (bool): Separate channels to distinct samples for stereo support.
+        n_ffts (Sequence[int]): Size of FFT for each scale.
+        hop_lengths (Sequence[int]): Length of hop between STFT windows for each scale.
+        win_lengths (Sequence[int]): Window size for each scale.
+        **kwargs: Additional args for STFTDiscriminator.
+    """
+    def __init__(self, filters: int, in_channels: int = 1, out_channels: int = 1, sep_channels: bool = False,
+                 n_ffts: tp.List[int] = [1024, 2048, 512], hop_lengths: tp.List[int] = [256, 512, 128],
+                 win_lengths: tp.List[int] = [1024, 2048, 512], **kwargs):
+        super().__init__()
+        assert len(n_ffts) == len(hop_lengths) == len(win_lengths)
+        self.sep_channels = sep_channels
+        self.discriminators = nn.ModuleList([
+            DiscriminatorSTFT(filters, in_channels=in_channels, out_channels=out_channels,
+                              n_fft=n_ffts[i], win_length=win_lengths[i], hop_length=hop_lengths[i], **kwargs)
+            for i in range(len(n_ffts))
+        ])
+
+    @property
+    def num_discriminators(self):
+        return len(self.discriminators)
+
+    def _separate_channels(self, x: torch.Tensor) -> torch.Tensor:
+        B, C, T = x.shape
+        return x.view(-1, 1, T)
+
+    def forward(self, x: torch.Tensor) -> MultiDiscriminatorOutputType:
+        logits = []
+        fmaps = []
+        for disc in self.discriminators:
+            logit, fmap = disc(x)
+            logits.append(logit)
+            fmaps.append(fmap)
+        return logits, fmaps

audiocraft/adversarial/losses.py

@@ -0,0 +1,228 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Utility module to handle adversarial losses without requiring to mess up the main training loop.
+"""
+
+import typing as tp
+
+import flashy
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+ADVERSARIAL_LOSSES = ['mse', 'hinge', 'hinge2']
+
+
+AdvLossType = tp.Union[nn.Module, tp.Callable[[torch.Tensor], torch.Tensor]]
+FeatLossType = tp.Union[nn.Module, tp.Callable[[torch.Tensor, torch.Tensor], torch.Tensor]]
+
+
+class AdversarialLoss(nn.Module):
+    """Adversary training wrapper.
+
+    Args:
+        adversary (nn.Module): The adversary module will be used to estimate the logits given the fake and real samples.
+            We assume here the adversary output is ``Tuple[List[torch.Tensor], List[List[torch.Tensor]]]``
+            where the first item is a list of logits and the second item is a list of feature maps.
+        optimizer (torch.optim.Optimizer): Optimizer used for training the given module.
+        loss (AdvLossType): Loss function for generator training.
+        loss_real (AdvLossType): Loss function for adversarial training on logits from real samples.
+        loss_fake (AdvLossType): Loss function for adversarial training on logits from fake samples.
+        loss_feat (FeatLossType): Feature matching loss function for generator training.
+        normalize (bool): Whether to normalize by number of sub-discriminators.
+
+    Example of usage:
+        adv_loss = AdversarialLoss(adversaries, optimizer, loss, loss_real, loss_fake)
+        for real in loader:
+            noise = torch.randn(...)
+            fake = model(noise)
+            adv_loss.train_adv(fake, real)
+            loss, _ = adv_loss(fake, real)
+            loss.backward()
+    """
+    def __init__(self,
+                 adversary: nn.Module,
+                 optimizer: torch.optim.Optimizer,
+                 loss: AdvLossType,
+                 loss_real: AdvLossType,
+                 loss_fake: AdvLossType,
+                 loss_feat: tp.Optional[FeatLossType] = None,
+                 normalize: bool = True):
+        super().__init__()
+        self.adversary: nn.Module = adversary
+        flashy.distrib.broadcast_model(self.adversary)
+        self.optimizer = optimizer
+        self.loss = loss
+        self.loss_real = loss_real
+        self.loss_fake = loss_fake
+        self.loss_feat = loss_feat
+        self.normalize = normalize
+
+    def _save_to_state_dict(self, destination, prefix, keep_vars):
+        # Add the optimizer state dict inside our own.
+        super()._save_to_state_dict(destination, prefix, keep_vars)
+        destination[prefix + 'optimizer'] = self.optimizer.state_dict()
+        return destination
+
+    def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs):
+        # Load optimizer state.
+        self.optimizer.load_state_dict(state_dict.pop(prefix + 'optimizer'))
+        super()._load_from_state_dict(state_dict, prefix, *args, **kwargs)
+
+    def get_adversary_pred(self, x):
+        """Run adversary model, validating expected output format."""
+        logits, fmaps = self.adversary(x)
+        assert isinstance(logits, list) and all([isinstance(t, torch.Tensor) for t in logits]), \
+            f'Expecting a list of tensors as logits but {type(logits)} found.'
+        assert isinstance(fmaps, list), f'Expecting a list of features maps but {type(fmaps)} found.'
+        for fmap in fmaps:
+            assert isinstance(fmap, list) and all([isinstance(f, torch.Tensor) for f in fmap]), \
+                f'Expecting a list of tensors as feature maps but {type(fmap)} found.'
+        return logits, fmaps
+
+    def train_adv(self, fake: torch.Tensor, real: torch.Tensor) -> torch.Tensor:
+        """Train the adversary with the given fake and real example.
+
+        We assume the adversary output is the following format: Tuple[List[torch.Tensor], List[List[torch.Tensor]]].
+        The first item being the logits and second item being a list of feature maps for each sub-discriminator.
+
+        This will automatically synchronize gradients (with `flashy.distrib.eager_sync_model`)
+        and call the optimizer.
+        """
+        loss = torch.tensor(0., device=fake.device)
+        all_logits_fake_is_fake, _ = self.get_adversary_pred(fake.detach())
+        all_logits_real_is_fake, _ = self.get_adversary_pred(real.detach())
+        n_sub_adversaries = len(all_logits_fake_is_fake)
+        for logit_fake_is_fake, logit_real_is_fake in zip(all_logits_fake_is_fake, all_logits_real_is_fake):
+            loss += self.loss_fake(logit_fake_is_fake) + self.loss_real(logit_real_is_fake)
+
+        if self.normalize:
+            loss /= n_sub_adversaries
+
+        self.optimizer.zero_grad()
+        with flashy.distrib.eager_sync_model(self.adversary):
+            loss.backward()
+        self.optimizer.step()
+
+        return loss
+
+    def forward(self, fake: torch.Tensor, real: torch.Tensor) -> tp.Tuple[torch.Tensor, torch.Tensor]:
+        """Return the loss for the generator, i.e. trying to fool the adversary,
+        and feature matching loss if provided.
+        """
+        adv = torch.tensor(0., device=fake.device)
+        feat = torch.tensor(0., device=fake.device)
+        with flashy.utils.readonly(self.adversary):
+            all_logits_fake_is_fake, all_fmap_fake = self.get_adversary_pred(fake)
+            all_logits_real_is_fake, all_fmap_real = self.get_adversary_pred(real)
+            n_sub_adversaries = len(all_logits_fake_is_fake)
+            for logit_fake_is_fake in all_logits_fake_is_fake:
+                adv += self.loss(logit_fake_is_fake)
+            if self.loss_feat:
+                for fmap_fake, fmap_real in zip(all_fmap_fake, all_fmap_real):
+                    feat += self.loss_feat(fmap_fake, fmap_real)
+
+        if self.normalize:
+            adv /= n_sub_adversaries
+            feat /= n_sub_adversaries
+
+        return adv, feat
+
+
+def get_adv_criterion(loss_type: str) -> tp.Callable:
+    assert loss_type in ADVERSARIAL_LOSSES
+    if loss_type == 'mse':
+        return mse_loss
+    elif loss_type == 'hinge':
+        return hinge_loss
+    elif loss_type == 'hinge2':
+        return hinge2_loss
+    raise ValueError('Unsupported loss')
+
+
+def get_fake_criterion(loss_type: str) -> tp.Callable:
+    assert loss_type in ADVERSARIAL_LOSSES
+    if loss_type == 'mse':
+        return mse_fake_loss
+    elif loss_type in ['hinge', 'hinge2']:
+        return hinge_fake_loss
+    raise ValueError('Unsupported loss')
+
+
+def get_real_criterion(loss_type: str) -> tp.Callable:
+    assert loss_type in ADVERSARIAL_LOSSES
+    if loss_type == 'mse':
+        return mse_real_loss
+    elif loss_type in ['hinge', 'hinge2']:
+        return hinge_real_loss
+    raise ValueError('Unsupported loss')
+
+
+def mse_real_loss(x: torch.Tensor) -> torch.Tensor:
+    return F.mse_loss(x, torch.tensor(1., device=x.device).expand_as(x))
+
+
+def mse_fake_loss(x: torch.Tensor) -> torch.Tensor:
+    return F.mse_loss(x, torch.tensor(0., device=x.device).expand_as(x))
+
+
+def hinge_real_loss(x: torch.Tensor) -> torch.Tensor:
+    return -torch.mean(torch.min(x - 1, torch.tensor(0., device=x.device).expand_as(x)))
+
+
+def hinge_fake_loss(x: torch.Tensor) -> torch.Tensor:
+    return -torch.mean(torch.min(-x - 1, torch.tensor(0., device=x.device).expand_as(x)))
+
+
+def mse_loss(x: torch.Tensor) -> torch.Tensor:
+    if x.numel() == 0:
+        return torch.tensor([0.0], device=x.device)
+    return F.mse_loss(x, torch.tensor(1., device=x.device).expand_as(x))
+
+
+def hinge_loss(x: torch.Tensor) -> torch.Tensor:
+    if x.numel() == 0:
+        return torch.tensor([0.0], device=x.device)
+    return -x.mean()
+
+
+def hinge2_loss(x: torch.Tensor) -> torch.Tensor:
+    if x.numel() == 0:
+        return torch.tensor([0.0])
+    return -torch.mean(torch.min(x - 1, torch.tensor(0., device=x.device).expand_as(x)))
+
+
+class FeatureMatchingLoss(nn.Module):
+    """Feature matching loss for adversarial training.
+
+    Args:
+        loss (nn.Module): Loss to use for feature matching (default=torch.nn.L1).
+        normalize (bool): Whether to normalize the loss.
+            by number of feature maps.
+    """
+    def __init__(self, loss: nn.Module = torch.nn.L1Loss(), normalize: bool = True):
+        super().__init__()
+        self.loss = loss
+        self.normalize = normalize
+
+    def forward(self, fmap_fake: tp.List[torch.Tensor], fmap_real: tp.List[torch.Tensor]) -> torch.Tensor:
+        assert len(fmap_fake) == len(fmap_real) and len(fmap_fake) > 0
+        feat_loss = torch.tensor(0., device=fmap_fake[0].device)
+        feat_scale = torch.tensor(0., device=fmap_fake[0].device)
+        n_fmaps = 0
+        for (feat_fake, feat_real) in zip(fmap_fake, fmap_real):
+            assert feat_fake.shape == feat_real.shape
+            n_fmaps += 1
+            feat_loss += self.loss(feat_fake, feat_real)
+            feat_scale += torch.mean(torch.abs(feat_real))
+
+        if self.normalize:
+            feat_loss /= n_fmaps
+
+        return feat_loss

audiocraft/data/__init__.py

@@ -0,0 +1,10 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Audio loading and writing support. Datasets for raw audio
+or also including some metadata."""
+
+# flake8: noqa
+from . import audio, audio_dataset, info_audio_dataset, music_dataset, sound_dataset

audiocraft/data/audio.py

@@ -0,0 +1,231 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+Audio IO methods are defined in this module (info, read, write),
+We rely on av library for faster read when possible, otherwise on torchaudio.
+"""
+
+from dataclasses import dataclass
+from pathlib import Path
+import logging
+import typing as tp
+
+import numpy as np
+import soundfile
+import torch
+from torch.nn import functional as F
+
+import av
+import subprocess as sp
+
+from .audio_utils import f32_pcm, normalize_audio
+
+
+_av_initialized = False
+
+
+def _init_av():
+    global _av_initialized
+    if _av_initialized:
+        return
+    logger = logging.getLogger('libav.mp3')
+    logger.setLevel(logging.ERROR)
+    _av_initialized = True
+
+
+@dataclass(frozen=True)
+class AudioFileInfo:
+    sample_rate: int
+    duration: float
+    channels: int
+
+
+def _av_info(filepath: tp.Union[str, Path]) -> AudioFileInfo:
+    _init_av()
+    with av.open(str(filepath)) as af:
+        stream = af.streams.audio[0]
+        sample_rate = stream.codec_context.sample_rate
+        duration = float(stream.duration * stream.time_base)
+        channels = stream.channels
+        return AudioFileInfo(sample_rate, duration, channels)
+
+
+def _soundfile_info(filepath: tp.Union[str, Path]) -> AudioFileInfo:
+    info = soundfile.info(filepath)
+    return AudioFileInfo(info.samplerate, info.duration, info.channels)
+
+
+def audio_info(filepath: tp.Union[str, Path]) -> AudioFileInfo:
+    # torchaudio no longer returns useful duration informations for some formats like mp3s.
+    filepath = Path(filepath)
+    if filepath.suffix in ['.flac', '.ogg']:  # TODO: Validate .ogg can be safely read with av_info
+        # ffmpeg has some weird issue with flac.
+        return _soundfile_info(filepath)
+    else:
+        return _av_info(filepath)
+
+
+def _av_read(filepath: tp.Union[str, Path], seek_time: float = 0, duration: float = -1.) -> tp.Tuple[torch.Tensor, int]:
+    """FFMPEG-based audio file reading using PyAV bindings.
+    Soundfile cannot read mp3 and av_read is more efficient than torchaudio.
+
+    Args:
+        filepath (str or Path): Path to audio file to read.
+        seek_time (float): Time at which to start reading in the file.
+        duration (float): Duration to read from the file. If set to -1, the whole file is read.
+    Returns:
+        tuple of torch.Tensor, int: Tuple containing audio data and sample rate
+    """
+    _init_av()
+    with av.open(str(filepath)) as af:
+        stream = af.streams.audio[0]
+        sr = stream.codec_context.sample_rate
+        num_frames = int(sr * duration) if duration >= 0 else -1
+        frame_offset = int(sr * seek_time)
+        # we need a small negative offset otherwise we get some edge artifact
+        # from the mp3 decoder.
+        af.seek(int(max(0, (seek_time - 0.1)) / stream.time_base), stream=stream)
+        frames = []
+        length = 0
+        for frame in af.decode(streams=stream.index):
+            current_offset = int(frame.rate * frame.pts * frame.time_base)
+            strip = max(0, frame_offset - current_offset)
+            buf = torch.from_numpy(frame.to_ndarray())
+            if buf.shape[0] != stream.channels:
+                buf = buf.view(-1, stream.channels).t()
+            buf = buf[:, strip:]
+            frames.append(buf)
+            length += buf.shape[1]
+            if num_frames > 0 and length >= num_frames:
+                break
+        assert frames
+        # If the above assert fails, it is likely because we seeked past the end of file point,
+        # in which case ffmpeg returns a single frame with only zeros, and a weird timestamp.
+        # This will need proper debugging, in due time.
+        wav = torch.cat(frames, dim=1)
+        assert wav.shape[0] == stream.channels
+        if num_frames > 0:
+            wav = wav[:, :num_frames]
+        return f32_pcm(wav), sr
+
+
+def audio_read(filepath: tp.Union[str, Path], seek_time: float = 0.,
+               duration: float = -1., pad: bool = False) -> tp.Tuple[torch.Tensor, int]:
+    """Read audio by picking the most appropriate backend tool based on the audio format.
+
+    Args:
+        filepath (str or Path): Path to audio file to read.
+        seek_time (float): Time at which to start reading in the file.
+        duration (float): Duration to read from the file. If set to -1, the whole file is read.
+        pad (bool): Pad output audio if not reaching expected duration.
+    Returns:
+        tuple of torch.Tensor, int: Tuple containing audio data and sample rate.
+    """
+    fp = Path(filepath)
+    if fp.suffix in ['.flac', '.ogg']:  # TODO: check if we can safely use av_read for .ogg
+        # There is some bug with ffmpeg and reading flac
+        info = _soundfile_info(filepath)
+        frames = -1 if duration <= 0 else int(duration * info.sample_rate)
+        frame_offset = int(seek_time * info.sample_rate)
+        wav, sr = soundfile.read(filepath, start=frame_offset, frames=frames, dtype=np.float32)
+        assert info.sample_rate == sr, f"Mismatch of sample rates {info.sample_rate} {sr}"
+        wav = torch.from_numpy(wav).t().contiguous()
+        if len(wav.shape) == 1:
+            wav = torch.unsqueeze(wav, 0)
+    else:
+        wav, sr = _av_read(filepath, seek_time, duration)
+    if pad and duration > 0:
+        expected_frames = int(duration * sr)
+        wav = F.pad(wav, (0, expected_frames - wav.shape[-1]))
+    return wav, sr
+
+
+def _piping_to_ffmpeg(out_path: tp.Union[str, Path], wav: torch.Tensor, sample_rate: int, flags: tp.List[str]):
+    # ffmpeg is always installed and torchaudio is a bit unstable lately, so let's bypass it entirely.
+    assert wav.dim() == 2, wav.shape
+    command = [
+        'ffmpeg',
+        '-loglevel', 'error',
+        '-y', '-f', 'f32le', '-ar', str(sample_rate), '-ac', str(wav.shape[0]),
+        '-i', '-'] + flags + [str(out_path)]
+    input_ = f32_pcm(wav).t().detach().cpu().numpy().tobytes()
+    sp.run(command, input=input_, check=True)
+
+
+def audio_write(stem_name: tp.Union[str, Path],
+                wav: torch.Tensor, sample_rate: int,
+                format: str = 'wav', mp3_rate: int = 320, ogg_rate: tp.Optional[int] = None,
+                normalize: bool = True, strategy: str = 'peak', peak_clip_headroom_db: float = 1,
+                rms_headroom_db: float = 18, loudness_headroom_db: float = 14,
+                loudness_compressor: bool = False,
+                log_clipping: bool = True, make_parent_dir: bool = True,
+                add_suffix: bool = True) -> Path:
+    """Convenience function for saving audio to disk. Returns the filename the audio was written to.
+
+    Args:
+        stem_name (str or Path): Filename without extension which will be added automatically.
+        wav (torch.Tensor): Audio data to save.
+        sample_rate (int): Sample rate of audio data.
+        format (str): Either "wav", "mp3", "ogg", or "flac".
+        mp3_rate (int): kbps when using mp3s.
+        ogg_rate (int): kbps when using ogg/vorbis. If not provided, let ffmpeg decide for itself.
+        normalize (bool): if `True` (default), normalizes according to the prescribed
+            strategy (see after). If `False`, the strategy is only used in case clipping
+            would happen.
+        strategy (str): Can be either 'clip', 'peak', or 'rms'. Default is 'peak',
+            i.e. audio is normalized by its largest value. RMS normalizes by root-mean-square
+            with extra headroom to avoid clipping. 'clip' just clips.
+        peak_clip_headroom_db (float): Headroom in dB when doing 'peak' or 'clip' strategy.
+        rms_headroom_db (float): Headroom in dB when doing 'rms' strategy. This must be much larger
+            than the `peak_clip` one to avoid further clipping.
+        loudness_headroom_db (float): Target loudness for loudness normalization.
+        loudness_compressor (bool): Uses tanh for soft clipping when strategy is 'loudness'.
+         when strategy is 'loudness' log_clipping (bool): If True, basic logging on stderr when clipping still
+            occurs despite strategy (only for 'rms').
+        make_parent_dir (bool): Make parent directory if it doesn't exist.
+    Returns:
+        Path: Path of the saved audio.
+    """
+    assert wav.dtype.is_floating_point, "wav is not floating point"
+    if wav.dim() == 1:
+        wav = wav[None]
+    elif wav.dim() > 2:
+        raise ValueError("Input wav should be at most 2 dimension.")
+    assert wav.isfinite().all()
+    wav = normalize_audio(wav, normalize, strategy, peak_clip_headroom_db,
+                          rms_headroom_db, loudness_headroom_db, loudness_compressor,
+                          log_clipping=log_clipping, sample_rate=sample_rate,
+                          stem_name=str(stem_name))
+    if format == 'mp3':
+        suffix = '.mp3'
+        flags = ['-f', 'mp3', '-c:a', 'libmp3lame', '-b:a', f'{mp3_rate}k']
+    elif format == 'wav':
+        suffix = '.wav'
+        flags = ['-f', 'wav', '-c:a', 'pcm_s16le']
+    elif format == 'ogg':
+        suffix = '.ogg'
+        flags = ['-f', 'ogg', '-c:a', 'libvorbis']
+        if ogg_rate is not None:
+            flags += ['-b:a', f'{ogg_rate}k']
+    elif format == 'flac':
+        suffix = '.flac'
+        flags = ['-f', 'flac']
+    else:
+        raise RuntimeError(f"Invalid format {format}. Only wav or mp3 are supported.")
+    if not add_suffix:
+        suffix = ''
+    path = Path(str(stem_name) + suffix)
+    if make_parent_dir:
+        path.parent.mkdir(exist_ok=True, parents=True)
+    try:
+        _piping_to_ffmpeg(path, wav, sample_rate, flags)
+    except Exception:
+        if path.exists():
+            # we do not want to leave half written files around.
+            path.unlink()
+        raise
+    return path

audiocraft/data/audio_dataset.py

@@ -0,0 +1,587 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""AudioDataset support. In order to handle a larger number of files
+without having to scan again the folders, we precompute some metadata
+(filename, sample rate, duration), and use that to efficiently sample audio segments.
+"""
+import argparse
+import copy
+from concurrent.futures import ThreadPoolExecutor, Future
+from dataclasses import dataclass, fields
+from contextlib import ExitStack
+from functools import lru_cache
+import gzip
+import json
+import logging
+import os
+from pathlib import Path
+import random
+import sys
+import typing as tp
+
+import torch
+import torch.nn.functional as F
+
+from .audio import audio_read, audio_info
+from .audio_utils import convert_audio
+from .zip import PathInZip
+
+try:
+    import dora
+except ImportError:
+    dora = None  # type: ignore
+
+
+@dataclass(order=True)
+class BaseInfo:
+
+    @classmethod
+    def _dict2fields(cls, dictionary: dict):
+        return {
+            field.name: dictionary[field.name]
+            for field in fields(cls) if field.name in dictionary
+        }
+
+    @classmethod
+    def from_dict(cls, dictionary: dict):
+        _dictionary = cls._dict2fields(dictionary)
+        return cls(**_dictionary)
+
+    def to_dict(self):
+        return {
+            field.name: self.__getattribute__(field.name)
+            for field in fields(self)
+            }
+
+
+@dataclass(order=True)
+class AudioMeta(BaseInfo):
+    path: str
+    duration: float
+    sample_rate: int
+    amplitude: tp.Optional[float] = None
+    weight: tp.Optional[float] = None
+    # info_path is used to load additional information about the audio file that is stored in zip files.
+    info_path: tp.Optional[PathInZip] = None
+
+    @classmethod
+    def from_dict(cls, dictionary: dict):
+        base = cls._dict2fields(dictionary)
+        if 'info_path' in base and base['info_path'] is not None:
+            base['info_path'] = PathInZip(base['info_path'])
+        return cls(**base)
+
+    def to_dict(self):
+        d = super().to_dict()
+        if d['info_path'] is not None:
+            d['info_path'] = str(d['info_path'])
+        return d
+
+
+@dataclass(order=True)
+class SegmentInfo(BaseInfo):
+    meta: AudioMeta
+    seek_time: float
+    # The following values are given once the audio is processed, e.g.
+    # at the target sample rate and target number of channels.
+    n_frames: int      # actual number of frames without padding
+    total_frames: int  # total number of frames, padding included
+    sample_rate: int   # actual sample rate
+    channels: int      # number of audio channels.
+
+
+DEFAULT_EXTS = ['.wav', '.mp3', '.flac', '.ogg', '.m4a']
+
+logger = logging.getLogger(__name__)
+
+
+def _get_audio_meta(file_path: str, minimal: bool = True) -> AudioMeta:
+    """AudioMeta from a path to an audio file.
+
+    Args:
+        file_path (str): Resolved path of valid audio file.
+        minimal (bool): Whether to only load the minimal set of metadata (takes longer if not).
+    Returns:
+        AudioMeta: Audio file path and its metadata.
+    """
+    info = audio_info(file_path)
+    amplitude: tp.Optional[float] = None
+    if not minimal:
+        wav, sr = audio_read(file_path)
+        amplitude = wav.abs().max().item()
+    return AudioMeta(file_path, info.duration, info.sample_rate, amplitude)
+
+
+def _resolve_audio_meta(m: AudioMeta, fast: bool = True) -> AudioMeta:
+    """If Dora is available as a dependency, try to resolve potential relative paths
+    in list of AudioMeta. This method is expected to be used when loading meta from file.
+
+    Args:
+        m (AudioMeta): Audio meta to resolve.
+        fast (bool): If True, uses a really fast check for determining if a file
+            is already absolute or not. Only valid on Linux/Mac.
+    Returns:
+        AudioMeta: Audio meta with resolved path.
+    """
+    def is_abs(m):
+        if fast:
+            return str(m)[0] == '/'
+        else:
+            os.path.isabs(str(m))
+
+    if not dora:
+        return m
+
+    if not is_abs(m.path):
+        m.path = dora.git_save.to_absolute_path(m.path)
+    if m.info_path is not None and not is_abs(m.info_path.zip_path):
+        m.info_path.zip_path = dora.git_save.to_absolute_path(m.path)
+    return m
+
+
+def find_audio_files(path: tp.Union[Path, str],
+                     exts: tp.List[str] = DEFAULT_EXTS,
+                     resolve: bool = True,
+                     minimal: bool = True,
+                     progress: bool = False,
+                     workers: int = 0) -> tp.List[AudioMeta]:
+    """Build a list of AudioMeta from a given path,
+    collecting relevant audio files and fetching meta info.
+
+    Args:
+        path (str or Path): Path to folder containing audio files.
+        exts (list of str): List of file extensions to consider for audio files.
+        minimal (bool): Whether to only load the minimal set of metadata (takes longer if not).
+        progress (bool): Whether to log progress on audio files collection.
+        workers (int): number of parallel workers, if 0, use only the current thread.
+    Returns:
+        list of AudioMeta: List of audio file path and its metadata.
+    """
+    audio_files = []
+    futures: tp.List[Future] = []
+    pool: tp.Optional[ThreadPoolExecutor] = None
+    with ExitStack() as stack:
+        if workers > 0:
+            pool = ThreadPoolExecutor(workers)
+            stack.enter_context(pool)
+
+        if progress:
+            print("Finding audio files...")
+        for root, folders, files in os.walk(path, followlinks=True):
+            for file in files:
+                full_path = Path(root) / file
+                if full_path.suffix.lower() in exts:
+                    audio_files.append(full_path)
+                    if pool is not None:
+                        futures.append(pool.submit(_get_audio_meta, str(audio_files[-1]), minimal))
+                    if progress:
+                        print(format(len(audio_files), " 8d"), end='\r', file=sys.stderr)
+
+        if progress:
+            print("Getting audio metadata...")
+        meta: tp.List[AudioMeta] = []
+        for idx, file_path in enumerate(audio_files):
+            try:
+                if pool is None:
+                    m = _get_audio_meta(str(file_path), minimal)
+                else:
+                    m = futures[idx].result()
+                if resolve:
+                    m = _resolve_audio_meta(m)
+            except Exception as err:
+                print("Error with", str(file_path), err, file=sys.stderr)
+                continue
+            meta.append(m)
+            if progress:
+                print(format((1 + idx) / len(audio_files), " 3.1%"), end='\r', file=sys.stderr)
+    meta.sort()
+    return meta
+
+
+def load_audio_meta(path: tp.Union[str, Path],
+                    resolve: bool = True, fast: bool = True) -> tp.List[AudioMeta]:
+    """Load list of AudioMeta from an optionally compressed json file.
+
+    Args:
+        path (str or Path): Path to JSON file.
+        resolve (bool): Whether to resolve the path from AudioMeta (default=True).
+        fast (bool): activates some tricks to make things faster.
+    Returns:
+        list of AudioMeta: List of audio file path and its total duration.
+    """
+    open_fn = gzip.open if str(path).lower().endswith('.gz') else open
+    with open_fn(path, 'rb') as fp:  # type: ignore
+        lines = fp.readlines()
+    meta = []
+    for line in lines:
+        d = json.loads(line)
+        m = AudioMeta.from_dict(d)
+        if resolve:
+            m = _resolve_audio_meta(m, fast=fast)
+        meta.append(m)
+    return meta
+
+
+def save_audio_meta(path: tp.Union[str, Path], meta: tp.List[AudioMeta]):
+    """Save the audio metadata to the file pointer as json.
+
+    Args:
+        path (str or Path): Path to JSON file.
+        metadata (list of BaseAudioMeta): List of audio meta to save.
+    """
+    Path(path).parent.mkdir(exist_ok=True, parents=True)
+    open_fn = gzip.open if str(path).lower().endswith('.gz') else open
+    with open_fn(path, 'wb') as fp:  # type: ignore
+        for m in meta:
+            json_str = json.dumps(m.to_dict()) + '\n'
+            json_bytes = json_str.encode('utf-8')
+            fp.write(json_bytes)
+
+
+class AudioDataset:
+    """Base audio dataset.
+
+    The dataset takes a list of AudioMeta and create a dataset composed of segments of audio
+    and potentially additional information, by creating random segments from the list of audio
+    files referenced in the metadata and applying minimal data pre-processing such as resampling,
+    mixing of channels, padding, etc.
+
+    If no segment_duration value is provided, the AudioDataset will return the full wav for each
+    audio file. Otherwise, it will randomly sample audio files and create a segment of the specified
+    duration, applying padding if required.
+
+    By default, only the torch Tensor corresponding to the waveform is returned. Setting return_info=True
+    allows to return a tuple containing the torch Tensor and additional metadata on the segment and the
+    original audio meta.
+
+    Note that you can call `start_epoch(epoch)` in order to get
+    a deterministic "randomization" for `shuffle=True`.
+    For a given epoch and dataset index, this will always return the same extract.
+    You can get back some diversity by setting the `shuffle_seed` param.
+
+    Args:
+        meta (list of AudioMeta): List of audio files metadata.
+        segment_duration (float, optional): Optional segment duration of audio to load.
+            If not specified, the dataset will load the full audio segment from the file.
+        shuffle (bool): Set to `True` to have the data reshuffled at every epoch.
+        sample_rate (int): Target sample rate of the loaded audio samples.
+        channels (int): Target number of channels of the loaded audio samples.
+        sample_on_duration (bool): Set to `True` to sample segments with probability
+            dependent on audio file duration. This is only used if `segment_duration` is provided.
+        sample_on_weight (bool): Set to `True` to sample segments using the `weight` entry of
+            `AudioMeta`. If `sample_on_duration` is also True, the actual weight will be the product
+            of the file duration and file weight. This is only used if `segment_duration` is provided.
+        min_segment_ratio (float): Minimum segment ratio to use when the audio file
+            is shorter than the desired segment.
+        max_read_retry (int): Maximum number of retries to sample an audio segment from the dataset.
+        return_info (bool): Whether to return the wav only or return wav along with segment info and metadata.
+        min_audio_duration (float, optional): Minimum audio file duration, in seconds, if provided
+            audio shorter than this will be filtered out.
+        max_audio_duration (float, optional): Maximal audio file duration in seconds, if provided
+            audio longer than this will be filtered out.
+        shuffle_seed (int): can be used to further randomize
+        load_wav (bool): if False, skip loading the wav but returns a tensor of 0
+            with the expected segment_duration (which must be provided if load_wav is False).
+        permutation_on_files (bool): only if `sample_on_weight` and `sample_on_duration`
+            are False. Will ensure a permutation on files when going through the dataset.
+            In that case the epoch number must be provided in order for the model
+            to continue the permutation across epochs. In that case, it is assumed
+            that `num_samples = total_batch_size * num_updates_per_epoch`, with
+            `total_batch_size` the overall batch size accounting for all gpus.
+    """
+    def __init__(self,
+                 meta: tp.List[AudioMeta],
+                 segment_duration: tp.Optional[float] = None,
+                 shuffle: bool = True,
+                 num_samples: int = 10_000,
+                 sample_rate: int = 48_000,
+                 channels: int = 2,
+                 pad: bool = True,
+                 sample_on_duration: bool = True,
+                 sample_on_weight: bool = True,
+                 min_segment_ratio: float = 0.5,
+                 max_read_retry: int = 10,
+                 return_info: bool = False,
+                 min_audio_duration: tp.Optional[float] = None,
+                 max_audio_duration: tp.Optional[float] = None,
+                 shuffle_seed: int = 0,
+                 load_wav: bool = True,
+                 permutation_on_files: bool = False,
+                 ):
+        assert len(meta) > 0, "No audio meta provided to AudioDataset. Please check loading of audio meta."
+        assert segment_duration is None or segment_duration > 0
+        assert segment_duration is None or min_segment_ratio >= 0
+        self.segment_duration = segment_duration
+        self.min_segment_ratio = min_segment_ratio
+        self.max_audio_duration = max_audio_duration
+        self.min_audio_duration = min_audio_duration
+        if self.min_audio_duration is not None and self.max_audio_duration is not None:
+            assert self.min_audio_duration <= self.max_audio_duration
+        self.meta: tp.List[AudioMeta] = self._filter_duration(meta)
+        assert len(self.meta)  # Fail fast if all data has been filtered.
+        self.total_duration = sum(d.duration for d in self.meta)
+
+        if segment_duration is None:
+            num_samples = len(self.meta)
+        self.num_samples = num_samples
+        self.shuffle = shuffle
+        self.sample_rate = sample_rate
+        self.channels = channels
+        self.pad = pad
+        self.sample_on_weight = sample_on_weight
+        self.sample_on_duration = sample_on_duration
+        self.sampling_probabilities = self._get_sampling_probabilities()
+        self.max_read_retry = max_read_retry
+        self.return_info = return_info
+        self.shuffle_seed = shuffle_seed
+        self.current_epoch: tp.Optional[int] = None
+        self.load_wav = load_wav
+        if not load_wav:
+            assert segment_duration is not None
+        self.permutation_on_files = permutation_on_files
+        if permutation_on_files:
+            assert not self.sample_on_duration
+            assert not self.sample_on_weight
+            assert self.shuffle
+
+    def start_epoch(self, epoch: int):
+        self.current_epoch = epoch
+
+    def __len__(self):
+        return self.num_samples
+
+    def _get_sampling_probabilities(self, normalized: bool = True):
+        """Return the sampling probabilities for each file inside `self.meta`."""
+        scores: tp.List[float] = []
+        for file_meta in self.meta:
+            score = 1.
+            if self.sample_on_weight and file_meta.weight is not None:
+                score *= file_meta.weight
+            if self.sample_on_duration:
+                score *= file_meta.duration
+            scores.append(score)
+        probabilities = torch.tensor(scores)
+        if normalized:
+            probabilities /= probabilities.sum()
+        return probabilities
+
+    @staticmethod
+    @lru_cache(16)
+    def _get_file_permutation(num_files: int, permutation_index: int, base_seed: int):
+        # Used to keep the most recent files permutation in memory implicitely.
+        # will work unless someone is using a lot of Datasets in parallel.
+        rng = torch.Generator()
+        rng.manual_seed(base_seed + permutation_index)
+        return torch.randperm(num_files, generator=rng)
+
+    def sample_file(self, index: int, rng: torch.Generator) -> AudioMeta:
+        """Sample a given file from `self.meta`. Can be overridden in subclasses.
+        This is only called if `segment_duration` is not None.
+
+        You must use the provided random number generator `rng` for reproducibility.
+        You can further make use of the index accessed.
+        """
+        if self.permutation_on_files:
+            assert self.current_epoch is not None
+            total_index = self.current_epoch * len(self) + index
+            permutation_index = total_index // len(self.meta)
+            relative_index = total_index % len(self.meta)
+            permutation = AudioDataset._get_file_permutation(
+                len(self.meta), permutation_index, self.shuffle_seed)
+            file_index = permutation[relative_index]
+            return self.meta[file_index]
+
+        if not self.sample_on_weight and not self.sample_on_duration:
+            file_index = int(torch.randint(len(self.sampling_probabilities), (1,), generator=rng).item())
+        else:
+            file_index = int(torch.multinomial(self.sampling_probabilities, 1, generator=rng).item())
+
+        return self.meta[file_index]
+
+    def _audio_read(self, path: str, seek_time: float = 0, duration: float = -1):
+        # Override this method in subclass if needed.
+        if self.load_wav:
+            return audio_read(path, seek_time, duration, pad=False)
+        else:
+            assert self.segment_duration is not None
+            n_frames = int(self.sample_rate * self.segment_duration)
+            return torch.zeros(self.channels, n_frames), self.sample_rate
+
+    def __getitem__(self, index: int) -> tp.Union[torch.Tensor, tp.Tuple[torch.Tensor, SegmentInfo]]:
+        if self.segment_duration is None:
+            file_meta = self.meta[index]
+            out, sr = audio_read(file_meta.path)
+            out = convert_audio(out, sr, self.sample_rate, self.channels)
+            n_frames = out.shape[-1]
+            segment_info = SegmentInfo(file_meta, seek_time=0., n_frames=n_frames, total_frames=n_frames,
+                                       sample_rate=self.sample_rate, channels=out.shape[0])
+        else:
+            rng = torch.Generator()
+            if self.shuffle:
+                # We use index, plus extra randomness, either totally random if we don't know the epoch.
+                # otherwise we make use of the epoch number and optional shuffle_seed.
+                if self.current_epoch is None:
+                    rng.manual_seed(index + self.num_samples * random.randint(0, 2**24))
+                else:
+                    rng.manual_seed(index + self.num_samples * (self.current_epoch + self.shuffle_seed))
+            else:
+                # We only use index
+                rng.manual_seed(index)
+
+            for retry in range(self.max_read_retry):
+                file_meta = self.sample_file(index, rng)
+                # We add some variance in the file position even if audio file is smaller than segment
+                # without ending up with empty segments
+                max_seek = max(0, file_meta.duration - self.segment_duration * self.min_segment_ratio)
+                seek_time = torch.rand(1, generator=rng).item() * max_seek
+                try:
+                    out, sr = audio_read(file_meta.path, seek_time, self.segment_duration, pad=False)
+                    out = convert_audio(out, sr, self.sample_rate, self.channels)
+                    n_frames = out.shape[-1]
+                    target_frames = int(self.segment_duration * self.sample_rate)
+                    if self.pad:
+                        out = F.pad(out, (0, target_frames - n_frames))
+                    segment_info = SegmentInfo(file_meta, seek_time, n_frames=n_frames, total_frames=target_frames,
+                                               sample_rate=self.sample_rate, channels=out.shape[0])
+                except Exception as exc:
+                    logger.warning("Error opening file %s: %r", file_meta.path, exc)
+                    if retry == self.max_read_retry - 1:
+                        raise
+                else:
+                    break
+
+        if self.return_info:
+            # Returns the wav and additional information on the wave segment
+            return out, segment_info
+        else:
+            return out
+
+    def collater(self, samples):
+        """The collater function has to be provided to the dataloader
+        if AudioDataset has return_info=True in order to properly collate
+        the samples of a batch.
+        """
+        if self.segment_duration is None and len(samples) > 1:
+            assert self.pad, "Must allow padding when batching examples of different durations."
+
+        # In this case the audio reaching the collater is of variable length as segment_duration=None.
+        to_pad = self.segment_duration is None and self.pad
+        if to_pad:
+            max_len = max([wav.shape[-1] for wav, _ in samples])
+
+            def _pad_wav(wav):
+                return F.pad(wav, (0, max_len - wav.shape[-1]))
+
+        if self.return_info:
+            if len(samples) > 0:
+                assert len(samples[0]) == 2
+                assert isinstance(samples[0][0], torch.Tensor)
+                assert isinstance(samples[0][1], SegmentInfo)
+
+            wavs = [wav for wav, _ in samples]
+            segment_infos = [copy.deepcopy(info) for _, info in samples]
+
+            if to_pad:
+                # Each wav could be of a different duration as they are not segmented.
+                for i in range(len(samples)):
+                    # Determines the total length of the signal with padding, so we update here as we pad.
+                    segment_infos[i].total_frames = max_len
+                    wavs[i] = _pad_wav(wavs[i])
+
+            wav = torch.stack(wavs)
+            return wav, segment_infos
+        else:
+            assert isinstance(samples[0], torch.Tensor)
+            if to_pad:
+                samples = [_pad_wav(s) for s in samples]
+            return torch.stack(samples)
+
+    def _filter_duration(self, meta: tp.List[AudioMeta]) -> tp.List[AudioMeta]:
+        """Filters out audio files with audio durations that will not allow to sample examples from them."""
+        orig_len = len(meta)
+
+        # Filter data that is too short.
+        if self.min_audio_duration is not None:
+            meta = [m for m in meta if m.duration >= self.min_audio_duration]
+
+        # Filter data that is too long.
+        if self.max_audio_duration is not None:
+            meta = [m for m in meta if m.duration <= self.max_audio_duration]
+
+        filtered_len = len(meta)
+        removed_percentage = 100*(1-float(filtered_len)/orig_len)
+        msg = 'Removed %.2f percent of the data because it was too short or too long.' % removed_percentage
+        if removed_percentage < 10:
+            logging.debug(msg)
+        else:
+            logging.warning(msg)
+        return meta
+
+    @classmethod
+    def from_meta(cls, root: tp.Union[str, Path], **kwargs):
+        """Instantiate AudioDataset from a path to a directory containing a manifest as a jsonl file.
+
+        Args:
+            root (str or Path): Path to root folder containing audio files.
+            kwargs: Additional keyword arguments for the AudioDataset.
+        """
+        root = Path(root)
+        if root.is_dir():
+            if (root / 'data.jsonl').exists():
+                root = root / 'data.jsonl'
+            elif (root / 'data.jsonl.gz').exists():
+                root = root / 'data.jsonl.gz'
+            else:
+                raise ValueError("Don't know where to read metadata from in the dir. "
+                                 "Expecting either a data.jsonl or data.jsonl.gz file but none found.")
+        meta = load_audio_meta(root)
+        return cls(meta, **kwargs)
+
+    @classmethod
+    def from_path(cls, root: tp.Union[str, Path], minimal_meta: bool = True,
+                  exts: tp.List[str] = DEFAULT_EXTS, **kwargs):
+        """Instantiate AudioDataset from a path containing (possibly nested) audio files.
+
+        Args:
+            root (str or Path): Path to root folder containing audio files.
+            minimal_meta (bool): Whether to only load minimal metadata or not.
+            exts (list of str): Extensions for audio files.
+            kwargs: Additional keyword arguments for the AudioDataset.
+        """
+        root = Path(root)
+        if root.is_file():
+            meta = load_audio_meta(root, resolve=True)
+        else:
+            meta = find_audio_files(root, exts, minimal=minimal_meta, resolve=True)
+        return cls(meta, **kwargs)
+
+
+def main():
+    logging.basicConfig(stream=sys.stderr, level=logging.INFO)
+    parser = argparse.ArgumentParser(
+        prog='audio_dataset',
+        description='Generate .jsonl files by scanning a folder.')
+    parser.add_argument('root', help='Root folder with all the audio files')
+    parser.add_argument('output_meta_file',
+                        help='Output file to store the metadata, ')
+    parser.add_argument('--complete',
+                        action='store_false', dest='minimal', default=True,
+                        help='Retrieve all metadata, even the one that are expansive '
+                             'to compute (e.g. normalization).')
+    parser.add_argument('--resolve',
+                        action='store_true', default=False,
+                        help='Resolve the paths to be absolute and with no symlinks.')
+    parser.add_argument('--workers',
+                        default=10, type=int,
+                        help='Number of workers.')
+    args = parser.parse_args()
+    meta = find_audio_files(args.root, DEFAULT_EXTS, progress=True,
+                            resolve=args.resolve, minimal=args.minimal, workers=args.workers)
+    save_audio_meta(args.output_meta_file, meta)
+
+
+if __name__ == '__main__':
+    main()

audiocraft/data/audio_utils.py

@@ -0,0 +1,176 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Various utilities for audio convertion (pcm format, sample rate and channels),
+and volume normalization."""
+import sys
+import typing as tp
+
+import julius
+import torch
+import torchaudio
+
+
+def convert_audio_channels(wav: torch.Tensor, channels: int = 2) -> torch.Tensor:
+    """Convert audio to the given number of channels.
+
+    Args:
+        wav (torch.Tensor): Audio wave of shape [B, C, T].
+        channels (int): Expected number of channels as output.
+    Returns:
+        torch.Tensor: Downmixed or unchanged audio wave [B, C, T].
+    """
+    *shape, src_channels, length = wav.shape
+    if src_channels == channels:
+        pass
+    elif channels == 1:
+        # Case 1:
+        # The caller asked 1-channel audio, and the stream has multiple
+        # channels, downmix all channels.
+        wav = wav.mean(dim=-2, keepdim=True)
+    elif src_channels == 1:
+        # Case 2:
+        # The caller asked for multiple channels, but the input file has
+        # a single channel, replicate the audio over all channels.
+        wav = wav.expand(*shape, channels, length)
+    elif src_channels >= channels:
+        # Case 3:
+        # The caller asked for multiple channels, and the input file has
+        # more channels than requested. In that case return the first channels.
+        wav = wav[..., :channels, :]
+    else:
+        # Case 4: What is a reasonable choice here?
+        raise ValueError('The audio file has less channels than requested but is not mono.')
+    return wav
+
+
+def convert_audio(wav: torch.Tensor, from_rate: float,
+                  to_rate: float, to_channels: int) -> torch.Tensor:
+    """Convert audio to new sample rate and number of audio channels."""
+    wav = julius.resample_frac(wav, int(from_rate), int(to_rate))
+    wav = convert_audio_channels(wav, to_channels)
+    return wav
+
+
+def normalize_loudness(wav: torch.Tensor, sample_rate: int, loudness_headroom_db: float = 14,
+                       loudness_compressor: bool = False, energy_floor: float = 2e-3):
+    """Normalize an input signal to a user loudness in dB LKFS.
+    Audio loudness is defined according to the ITU-R BS.1770-4 recommendation.
+
+    Args:
+        wav (torch.Tensor): Input multichannel audio data.
+        sample_rate (int): Sample rate.
+        loudness_headroom_db (float): Target loudness of the output in dB LUFS.
+        loudness_compressor (bool): Uses tanh for soft clipping.
+        energy_floor (float): anything below that RMS level will not be rescaled.
+    Returns:
+        torch.Tensor: Loudness normalized output data.
+    """
+    energy = wav.pow(2).mean().sqrt().item()
+    if energy < energy_floor:
+        return wav
+    transform = torchaudio.transforms.Loudness(sample_rate)
+    input_loudness_db = transform(wav).item()
+    # calculate the gain needed to scale to the desired loudness level
+    delta_loudness = -loudness_headroom_db - input_loudness_db
+    gain = 10.0 ** (delta_loudness / 20.0)
+    output = gain * wav
+    if loudness_compressor:
+        output = torch.tanh(output)
+    assert output.isfinite().all(), (input_loudness_db, wav.pow(2).mean().sqrt())
+    return output
+
+
+def _clip_wav(wav: torch.Tensor, log_clipping: bool = False, stem_name: tp.Optional[str] = None) -> None:
+    """Utility function to clip the audio with logging if specified."""
+    max_scale = wav.abs().max()
+    if log_clipping and max_scale > 1:
+        clamp_prob = (wav.abs() > 1).float().mean().item()
+        print(f"CLIPPING {stem_name or ''} happening with proba (a bit of clipping is okay):",
+              clamp_prob, "maximum scale: ", max_scale.item(), file=sys.stderr)
+    wav.clamp_(-1, 1)
+
+
+def normalize_audio(wav: torch.Tensor, normalize: bool = True,
+                    strategy: str = 'peak', peak_clip_headroom_db: float = 1,
+                    rms_headroom_db: float = 18, loudness_headroom_db: float = 14,
+                    loudness_compressor: bool = False, log_clipping: bool = False,
+                    sample_rate: tp.Optional[int] = None,
+                    stem_name: tp.Optional[str] = None) -> torch.Tensor:
+    """Normalize the audio according to the prescribed strategy (see after).
+
+    Args:
+        wav (torch.Tensor): Audio data.
+        normalize (bool): if `True` (default), normalizes according to the prescribed
+            strategy (see after). If `False`, the strategy is only used in case clipping
+            would happen.
+        strategy (str): Can be either 'clip', 'peak', or 'rms'. Default is 'peak',
+            i.e. audio is normalized by its largest value. RMS normalizes by root-mean-square
+            with extra headroom to avoid clipping. 'clip' just clips.
+        peak_clip_headroom_db (float): Headroom in dB when doing 'peak' or 'clip' strategy.
+        rms_headroom_db (float): Headroom in dB when doing 'rms' strategy. This must be much larger
+            than the `peak_clip` one to avoid further clipping.
+        loudness_headroom_db (float): Target loudness for loudness normalization.
+        loudness_compressor (bool): If True, uses tanh based soft clipping.
+        log_clipping (bool): If True, basic logging on stderr when clipping still
+            occurs despite strategy (only for 'rms').
+        sample_rate (int): Sample rate for the audio data (required for loudness).
+        stem_name (str, optional): Stem name for clipping logging.
+    Returns:
+        torch.Tensor: Normalized audio.
+    """
+    scale_peak = 10 ** (-peak_clip_headroom_db / 20)
+    scale_rms = 10 ** (-rms_headroom_db / 20)
+    if strategy == 'peak':
+        rescaling = (scale_peak / wav.abs().max())
+        if normalize or rescaling < 1:
+            wav = wav * rescaling
+    elif strategy == 'clip':
+        wav = wav.clamp(-scale_peak, scale_peak)
+    elif strategy == 'rms':
+        mono = wav.mean(dim=0)
+        rescaling = scale_rms / mono.pow(2).mean().sqrt()
+        if normalize or rescaling < 1:
+            wav = wav * rescaling
+        _clip_wav(wav, log_clipping=log_clipping, stem_name=stem_name)
+    elif strategy == 'loudness':
+        assert sample_rate is not None, "Loudness normalization requires sample rate."
+        wav = normalize_loudness(wav, sample_rate, loudness_headroom_db, loudness_compressor)
+        _clip_wav(wav, log_clipping=log_clipping, stem_name=stem_name)
+    else:
+        assert wav.abs().max() < 1
+        assert strategy == '' or strategy == 'none', f"Unexpected strategy: '{strategy}'"
+    return wav
+
+
+def f32_pcm(wav: torch.Tensor) -> torch.Tensor:
+    """Convert audio to float 32 bits PCM format.
+    """
+    if wav.dtype.is_floating_point:
+        return wav
+    elif wav.dtype == torch.int16:
+        return wav.float() / 2**15
+    elif wav.dtype == torch.int32:
+        return wav.float() / 2**31
+    raise ValueError(f"Unsupported wav dtype: {wav.dtype}")
+
+
+def i16_pcm(wav: torch.Tensor) -> torch.Tensor:
+    """Convert audio to int 16 bits PCM format.
+
+    ..Warning:: There exist many formula for doing this conversion. None are perfect
+    due to the asymmetry of the int16 range. One either have possible clipping, DC offset,
+    or inconsistencies with f32_pcm. If the given wav doesn't have enough headroom,
+    it is possible that `i16_pcm(f32_pcm)) != Identity`.
+    """
+    if wav.dtype.is_floating_point:
+        assert wav.abs().max() <= 1
+        candidate = (wav * 2 ** 15).round()
+        if candidate.max() >= 2 ** 15:  # clipping would occur
+            candidate = (wav * (2 ** 15 - 1)).round()
+        return candidate.short()
+    else:
+        assert wav.dtype == torch.int16
+        return wav

audiocraft/data/info_audio_dataset.py

@@ -0,0 +1,110 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Base classes for the datasets that also provide non-audio metadata,
+e.g. description, text transcription etc.
+"""
+from dataclasses import dataclass
+import logging
+import math
+import re
+import typing as tp
+
+import torch
+
+from .audio_dataset import AudioDataset, AudioMeta
+from ..environment import AudioCraftEnvironment
+from ..modules.conditioners import SegmentWithAttributes, ConditioningAttributes
+
+
+logger = logging.getLogger(__name__)
+
+
+def _clusterify_meta(meta: AudioMeta) -> AudioMeta:
+    """Monkey-patch meta to match cluster specificities."""
+    meta.path = AudioCraftEnvironment.apply_dataset_mappers(meta.path)
+    if meta.info_path is not None:
+        meta.info_path.zip_path = AudioCraftEnvironment.apply_dataset_mappers(meta.info_path.zip_path)
+    return meta
+
+
+def clusterify_all_meta(meta: tp.List[AudioMeta]) -> tp.List[AudioMeta]:
+    """Monkey-patch all meta to match cluster specificities."""
+    return [_clusterify_meta(m) for m in meta]
+
+
+@dataclass
+class AudioInfo(SegmentWithAttributes):
+    """Dummy SegmentInfo with empty attributes.
+
+    The InfoAudioDataset is expected to return metadata that inherits
+    from SegmentWithAttributes class and can return conditioning attributes.
+
+    This basically guarantees all datasets will be compatible with current
+    solver that contain conditioners requiring this.
+    """
+    audio_tokens: tp.Optional[torch.Tensor] = None  # populated when using cached batch for training a LM.
+
+    def to_condition_attributes(self) -> ConditioningAttributes:
+        return ConditioningAttributes()
+
+
+class InfoAudioDataset(AudioDataset):
+    """AudioDataset that always returns metadata as SegmentWithAttributes along with the audio waveform.
+
+    See `audiocraft.data.audio_dataset.AudioDataset` for initialization arguments.
+    """
+    def __init__(self, meta: tp.List[AudioMeta], **kwargs):
+        super().__init__(clusterify_all_meta(meta), **kwargs)
+
+    def __getitem__(self, index: int) -> tp.Union[torch.Tensor, tp.Tuple[torch.Tensor, SegmentWithAttributes]]:
+        if not self.return_info:
+            wav = super().__getitem__(index)
+            assert isinstance(wav, torch.Tensor)
+            return wav
+        wav, meta = super().__getitem__(index)
+        return wav, AudioInfo(**meta.to_dict())
+
+
+def get_keyword_or_keyword_list(value: tp.Optional[str]) -> tp.Union[tp.Optional[str], tp.Optional[tp.List[str]]]:
+    """Preprocess a single keyword or possible a list of keywords."""
+    if isinstance(value, list):
+        return get_keyword_list(value)
+    else:
+        return get_keyword(value)
+
+
+def get_string(value: tp.Optional[str]) -> tp.Optional[str]:
+    """Preprocess a single keyword."""
+    if value is None or (not isinstance(value, str)) or len(value) == 0 or value == 'None':
+        return None
+    else:
+        return value.strip()
+
+
+def get_keyword(value: tp.Optional[str]) -> tp.Optional[str]:
+    """Preprocess a single keyword."""
+    if value is None or (not isinstance(value, str)) or len(value) == 0 or value == 'None':
+        return None
+    else:
+        return value.strip().lower()
+
+
+def get_keyword_list(values: tp.Union[str, tp.List[str]]) -> tp.Optional[tp.List[str]]:
+    """Preprocess a list of keywords."""
+    if isinstance(values, str):
+        values = [v.strip() for v in re.split(r'[,\s]', values)]
+    elif isinstance(values, float) and math.isnan(values):
+        values = []
+    if not isinstance(values, list):
+        logger.debug(f"Unexpected keyword list {values}")
+        values = [str(values)]
+
+    kws = [get_keyword(v) for v in values]
+    kw_list = [k for k in kws if k is not None]
+    if len(kw_list) == 0:
+        return None
+    else:
+        return kw_list

audiocraft/data/music_dataset.py

@@ -0,0 +1,270 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Dataset of music tracks with rich metadata.
+"""
+from dataclasses import dataclass, field, fields, replace
+import gzip
+import json
+import logging
+from pathlib import Path
+import random
+import typing as tp
+
+import torch
+
+from .info_audio_dataset import (
+    InfoAudioDataset,
+    AudioInfo,
+    get_keyword_list,
+    get_keyword,
+    get_string
+)
+from ..modules.conditioners import (
+    ConditioningAttributes,
+    JointEmbedCondition,
+    WavCondition,
+)
+from ..utils.utils import warn_once
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class MusicInfo(AudioInfo):
+    """Segment info augmented with music metadata.
+    """
+    # music-specific metadata
+    title: tp.Optional[str] = None
+    artist: tp.Optional[str] = None  # anonymized artist id, used to ensure no overlap between splits
+    key: tp.Optional[str] = None
+    bpm: tp.Optional[float] = None
+    genre: tp.Optional[str] = None
+    moods: tp.Optional[list] = None
+    keywords: tp.Optional[list] = None
+    description: tp.Optional[str] = None
+    name: tp.Optional[str] = None
+    instrument: tp.Optional[str] = None
+    # original wav accompanying the metadata
+    self_wav: tp.Optional[WavCondition] = None
+    # dict mapping attributes names to tuple of wav, text and metadata
+    joint_embed: tp.Dict[str, JointEmbedCondition] = field(default_factory=dict)
+
+    @property
+    def has_music_meta(self) -> bool:
+        return self.name is not None
+
+    def to_condition_attributes(self) -> ConditioningAttributes:
+        out = ConditioningAttributes()
+        for _field in fields(self):
+            key, value = _field.name, getattr(self, _field.name)
+            if key == 'self_wav':
+                out.wav[key] = value
+            elif key == 'joint_embed':
+                for embed_attribute, embed_cond in value.items():
+                    out.joint_embed[embed_attribute] = embed_cond
+            else:
+                if isinstance(value, list):
+                    value = ' '.join(value)
+                out.text[key] = value
+        return out
+
+    @staticmethod
+    def attribute_getter(attribute):
+        if attribute == 'bpm':
+            preprocess_func = get_bpm
+        elif attribute == 'key':
+            preprocess_func = get_musical_key
+        elif attribute in ['moods', 'keywords']:
+            preprocess_func = get_keyword_list
+        elif attribute in ['genre', 'name', 'instrument']:
+            preprocess_func = get_keyword
+        elif attribute in ['title', 'artist', 'description']:
+            preprocess_func = get_string
+        else:
+            preprocess_func = None
+        return preprocess_func
+
+    @classmethod
+    def from_dict(cls, dictionary: dict, fields_required: bool = False):
+        _dictionary: tp.Dict[str, tp.Any] = {}
+
+        # allow a subset of attributes to not be loaded from the dictionary
+        # these attributes may be populated later
+        post_init_attributes = ['self_wav', 'joint_embed']
+        optional_fields = ['keywords']
+
+        for _field in fields(cls):
+            if _field.name in post_init_attributes:
+                continue
+            elif _field.name not in dictionary:
+                if fields_required and _field.name not in optional_fields:
+                    raise KeyError(f"Unexpected missing key: {_field.name}")
+            else:
+                preprocess_func: tp.Optional[tp.Callable] = cls.attribute_getter(_field.name)
+                value = dictionary[_field.name]
+                if preprocess_func:
+                    value = preprocess_func(value)
+                _dictionary[_field.name] = value
+        return cls(**_dictionary)
+
+
+def augment_music_info_description(music_info: MusicInfo, merge_text_p: float = 0.,
+                                   drop_desc_p: float = 0., drop_other_p: float = 0.) -> MusicInfo:
+    """Augment MusicInfo description with additional metadata fields and potential dropout.
+    Additional textual attributes are added given probability 'merge_text_conditions_p' and
+    the original textual description is dropped from the augmented description given probability drop_desc_p.
+
+    Args:
+        music_info (MusicInfo): The music metadata to augment.
+        merge_text_p (float): Probability of merging additional metadata to the description.
+            If provided value is 0, then no merging is performed.
+        drop_desc_p (float): Probability of dropping the original description on text merge.
+            if provided value is 0, then no drop out is performed.
+        drop_other_p (float): Probability of dropping the other fields used for text augmentation.
+    Returns:
+        MusicInfo: The MusicInfo with augmented textual description.
+    """
+    def is_valid_field(field_name: str, field_value: tp.Any) -> bool:
+        valid_field_name = field_name in ['key', 'bpm', 'genre', 'moods', 'instrument', 'keywords']
+        valid_field_value = field_value is not None and isinstance(field_value, (int, float, str, list))
+        keep_field = random.uniform(0, 1) < drop_other_p
+        return valid_field_name and valid_field_value and keep_field
+
+    def process_value(v: tp.Any) -> str:
+        if isinstance(v, (int, float, str)):
+            return str(v)
+        if isinstance(v, list):
+            return ", ".join(v)
+        else:
+            raise ValueError(f"Unknown type for text value! ({type(v), v})")
+
+    description = music_info.description
+
+    metadata_text = ""
+    if random.uniform(0, 1) < merge_text_p:
+        meta_pairs = [f'{_field.name}: {process_value(getattr(music_info, _field.name))}'
+                      for _field in fields(music_info) if is_valid_field(_field.name, getattr(music_info, _field.name))]
+        random.shuffle(meta_pairs)
+        metadata_text = ". ".join(meta_pairs)
+        description = description if not random.uniform(0, 1) < drop_desc_p else None
+        logger.debug(f"Applying text augmentation on MMI info. description: {description}, metadata: {metadata_text}")
+
+    if description is None:
+        description = metadata_text if len(metadata_text) > 1 else None
+    else:
+        description = ". ".join([description.rstrip('.'), metadata_text])
+    description = description.strip() if description else None
+
+    music_info = replace(music_info)
+    music_info.description = description
+    return music_info
+
+
+class Paraphraser:
+    def __init__(self, paraphrase_source: tp.Union[str, Path], paraphrase_p: float = 0.):
+        self.paraphrase_p = paraphrase_p
+        open_fn = gzip.open if str(paraphrase_source).lower().endswith('.gz') else open
+        with open_fn(paraphrase_source, 'rb') as f:  # type: ignore
+            self.paraphrase_source = json.loads(f.read())
+        logger.info(f"loaded paraphrasing source from: {paraphrase_source}")
+
+    def sample_paraphrase(self, audio_path: str, description: str):
+        if random.random() >= self.paraphrase_p:
+            return description
+        info_path = Path(audio_path).with_suffix('.json')
+        if info_path not in self.paraphrase_source:
+            warn_once(logger, f"{info_path} not in paraphrase source!")
+            return description
+        new_desc = random.choice(self.paraphrase_source[info_path])
+        logger.debug(f"{description} -> {new_desc}")
+        return new_desc
+
+
+class MusicDataset(InfoAudioDataset):
+    """Music dataset is an AudioDataset with music-related metadata.
+
+    Args:
+        info_fields_required (bool): Whether to enforce having required fields.
+        merge_text_p (float): Probability of merging additional metadata to the description.
+        drop_desc_p (float): Probability of dropping the original description on text merge.
+        drop_other_p (float): Probability of dropping the other fields used for text augmentation.
+        joint_embed_attributes (list[str]): A list of attributes for which joint embedding metadata is returned.
+        paraphrase_source (str, optional): Path to the .json or .json.gz file containing the
+            paraphrases for the description. The json should be a dict with keys are the
+            original info path (e.g. track_path.json) and each value is a list of possible
+            paraphrased.
+        paraphrase_p (float): probability of taking a paraphrase.
+
+    See `audiocraft.data.info_audio_dataset.InfoAudioDataset` for full initialization arguments.
+    """
+    def __init__(self, *args, info_fields_required: bool = True,
+                 merge_text_p: float = 0., drop_desc_p: float = 0., drop_other_p: float = 0.,
+                 joint_embed_attributes: tp.List[str] = [],
+                 paraphrase_source: tp.Optional[str] = None, paraphrase_p: float = 0,
+                 **kwargs):
+        kwargs['return_info'] = True  # We require the info for each song of the dataset.
+        super().__init__(*args, **kwargs)
+        self.info_fields_required = info_fields_required
+        self.merge_text_p = merge_text_p
+        self.drop_desc_p = drop_desc_p
+        self.drop_other_p = drop_other_p
+        self.joint_embed_attributes = joint_embed_attributes
+        self.paraphraser = None
+        if paraphrase_source is not None:
+            self.paraphraser = Paraphraser(paraphrase_source, paraphrase_p)
+
+    def __getitem__(self, index):
+        wav, info = super().__getitem__(index)
+        info_data = info.to_dict()
+        music_info_path = Path(info.meta.path).with_suffix('.json')
+
+        if Path(music_info_path).exists():
+            with open(music_info_path, 'r') as json_file:
+                music_data = json.load(json_file)
+                music_data.update(info_data)
+                music_info = MusicInfo.from_dict(music_data, fields_required=self.info_fields_required)
+            if self.paraphraser is not None:
+                music_info.description = self.paraphraser.sample(music_info.meta.path, music_info.description)
+            if self.merge_text_p:
+                music_info = augment_music_info_description(
+                    music_info, self.merge_text_p, self.drop_desc_p, self.drop_other_p)
+        else:
+            music_info = MusicInfo.from_dict(info_data, fields_required=False)
+
+        music_info.self_wav = WavCondition(
+            wav=wav[None], length=torch.tensor([info.n_frames]),
+            sample_rate=[info.sample_rate], path=[info.meta.path], seek_time=[info.seek_time])
+
+        for att in self.joint_embed_attributes:
+            att_value = getattr(music_info, att)
+            joint_embed_cond = JointEmbedCondition(
+                wav[None], [att_value], torch.tensor([info.n_frames]),
+                sample_rate=[info.sample_rate], path=[info.meta.path], seek_time=[info.seek_time])
+            music_info.joint_embed[att] = joint_embed_cond
+
+        return wav, music_info
+
+
+def get_musical_key(value: tp.Optional[str]) -> tp.Optional[str]:
+    """Preprocess key keywords, discarding them if there are multiple key defined."""
+    if value is None or (not isinstance(value, str)) or len(value) == 0 or value == 'None':
+        return None
+    elif ',' in value:
+        # For now, we discard when multiple keys are defined separated with comas
+        return None
+    else:
+        return value.strip().lower()
+
+
+def get_bpm(value: tp.Optional[str]) -> tp.Optional[float]:
+    """Preprocess to a float."""
+    if value is None:
+        return None
+    try:
+        return float(value)
+    except ValueError:
+        return None

audiocraft/data/sound_dataset.py

@@ -0,0 +1,330 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Dataset of audio with a simple description.
+"""
+
+from dataclasses import dataclass, fields, replace
+import json
+from pathlib import Path
+import random
+import typing as tp
+
+import numpy as np
+import torch
+
+from .info_audio_dataset import (
+    InfoAudioDataset,
+    get_keyword_or_keyword_list
+)
+from ..modules.conditioners import (
+    ConditioningAttributes,
+    SegmentWithAttributes,
+    WavCondition,
+)
+
+
+EPS = torch.finfo(torch.float32).eps
+TARGET_LEVEL_LOWER = -35
+TARGET_LEVEL_UPPER = -15
+
+
+@dataclass
+class SoundInfo(SegmentWithAttributes):
+    """Segment info augmented with Sound metadata.
+    """
+    description: tp.Optional[str] = None
+    self_wav: tp.Optional[torch.Tensor] = None
+
+    @property
+    def has_sound_meta(self) -> bool:
+        return self.description is not None
+
+    def to_condition_attributes(self) -> ConditioningAttributes:
+        out = ConditioningAttributes()
+
+        for _field in fields(self):
+            key, value = _field.name, getattr(self, _field.name)
+            if key == 'self_wav':
+                out.wav[key] = value
+            else:
+                out.text[key] = value
+        return out
+
+    @staticmethod
+    def attribute_getter(attribute):
+        if attribute == 'description':
+            preprocess_func = get_keyword_or_keyword_list
+        else:
+            preprocess_func = None
+        return preprocess_func
+
+    @classmethod
+    def from_dict(cls, dictionary: dict, fields_required: bool = False):
+        _dictionary: tp.Dict[str, tp.Any] = {}
+
+        # allow a subset of attributes to not be loaded from the dictionary
+        # these attributes may be populated later
+        post_init_attributes = ['self_wav']
+
+        for _field in fields(cls):
+            if _field.name in post_init_attributes:
+                continue
+            elif _field.name not in dictionary:
+                if fields_required:
+                    raise KeyError(f"Unexpected missing key: {_field.name}")
+            else:
+                preprocess_func: tp.Optional[tp.Callable] = cls.attribute_getter(_field.name)
+                value = dictionary[_field.name]
+                if preprocess_func:
+                    value = preprocess_func(value)
+                _dictionary[_field.name] = value
+        return cls(**_dictionary)
+
+
+class SoundDataset(InfoAudioDataset):
+    """Sound audio dataset: Audio dataset with environmental sound-specific metadata.
+
+    Args:
+        info_fields_required (bool): Whether all the mandatory metadata fields should be in the loaded metadata.
+        external_metadata_source (tp.Optional[str]): Folder containing JSON metadata for the corresponding dataset.
+            The metadata files contained in this folder are expected to match the stem of the audio file with
+            a json extension.
+        aug_p (float): Probability of performing audio mixing augmentation on the batch.
+        mix_p (float): Proportion of batch items that are mixed together when applying audio mixing augmentation.
+        mix_snr_low (int): Lowerbound for SNR value sampled for mixing augmentation.
+        mix_snr_high (int): Upperbound for SNR value sampled for mixing augmentation.
+        mix_min_overlap (float): Minimum overlap between audio files when performing mixing augmentation.
+        kwargs: Additional arguments for AudioDataset.
+
+    See `audiocraft.data.info_audio_dataset.InfoAudioDataset` for full initialization arguments.
+    """
+    def __init__(
+        self,
+        *args,
+        info_fields_required: bool = True,
+        external_metadata_source: tp.Optional[str] = None,
+        aug_p: float = 0.,
+        mix_p: float = 0.,
+        mix_snr_low: int = -5,
+        mix_snr_high: int = 5,
+        mix_min_overlap: float = 0.5,
+        **kwargs
+    ):
+        kwargs['return_info'] = True  # We require the info for each song of the dataset.
+        super().__init__(*args, **kwargs)
+        self.info_fields_required = info_fields_required
+        self.external_metadata_source = external_metadata_source
+        self.aug_p = aug_p
+        self.mix_p = mix_p
+        if self.aug_p > 0:
+            assert self.mix_p > 0, "Expecting some mixing proportion mix_p if aug_p > 0"
+            assert self.channels == 1, "SoundDataset with audio mixing considers only monophonic audio"
+        self.mix_snr_low = mix_snr_low
+        self.mix_snr_high = mix_snr_high
+        self.mix_min_overlap = mix_min_overlap
+
+    def _get_info_path(self, path: tp.Union[str, Path]) -> Path:
+        """Get path of JSON with metadata (description, etc.).
+        If there exists a JSON with the same name as 'path.name', then it will be used.
+        Else, such JSON will be searched for in an external json source folder if it exists.
+        """
+        info_path = Path(path).with_suffix('.json')
+        if Path(info_path).exists():
+            return info_path
+        elif self.external_metadata_source and (Path(self.external_metadata_source) / info_path.name).exists():
+            return Path(self.external_metadata_source) / info_path.name
+        else:
+            raise Exception(f"Unable to find a metadata JSON for path: {path}")
+
+    def __getitem__(self, index):
+        wav, info = super().__getitem__(index)
+        info_data = info.to_dict()
+        info_path = self._get_info_path(info.meta.path)
+        if Path(info_path).exists():
+            with open(info_path, 'r') as json_file:
+                sound_data = json.load(json_file)
+                sound_data.update(info_data)
+                sound_info = SoundInfo.from_dict(sound_data, fields_required=self.info_fields_required)
+                # if there are multiple descriptions, sample one randomly
+                if isinstance(sound_info.description, list):
+                    sound_info.description = random.choice(sound_info.description)
+        else:
+            sound_info = SoundInfo.from_dict(info_data, fields_required=False)
+
+        sound_info.self_wav = WavCondition(
+            wav=wav[None], length=torch.tensor([info.n_frames]),
+            sample_rate=[sound_info.sample_rate], path=[info.meta.path], seek_time=[info.seek_time])
+
+        return wav, sound_info
+
+    def collater(self, samples):
+        # when training, audio mixing is performed in the collate function
+        wav, sound_info = super().collater(samples)  # SoundDataset always returns infos
+        if self.aug_p > 0:
+            wav, sound_info = mix_samples(wav, sound_info, self.aug_p, self.mix_p,
+                                          snr_low=self.mix_snr_low, snr_high=self.mix_snr_high,
+                                          min_overlap=self.mix_min_overlap)
+        return wav, sound_info
+
+
+def rms_f(x: torch.Tensor) -> torch.Tensor:
+    return (x ** 2).mean(1).pow(0.5)
+
+
+def normalize(audio: torch.Tensor, target_level: int = -25) -> torch.Tensor:
+    """Normalize the signal to the target level."""
+    rms = rms_f(audio)
+    scalar = 10 ** (target_level / 20) / (rms + EPS)
+    audio = audio * scalar.unsqueeze(1)
+    return audio
+
+
+def is_clipped(audio: torch.Tensor, clipping_threshold: float = 0.99) -> torch.Tensor:
+    return (abs(audio) > clipping_threshold).any(1)
+
+
+def mix_pair(src: torch.Tensor, dst: torch.Tensor, min_overlap: float) -> torch.Tensor:
+    start = random.randint(0, int(src.shape[1] * (1 - min_overlap)))
+    remainder = src.shape[1] - start
+    if dst.shape[1] > remainder:
+        src[:, start:] = src[:, start:] + dst[:, :remainder]
+    else:
+        src[:, start:start+dst.shape[1]] = src[:, start:start+dst.shape[1]] + dst
+    return src
+
+
+def snr_mixer(clean: torch.Tensor, noise: torch.Tensor, snr: int, min_overlap: float,
+              target_level: int = -25, clipping_threshold: float = 0.99) -> torch.Tensor:
+    """Function to mix clean speech and noise at various SNR levels.
+
+    Args:
+        clean (torch.Tensor): Clean audio source to mix, of shape [B, T].
+        noise (torch.Tensor): Noise audio source to mix, of shape [B, T].
+        snr (int): SNR level when mixing.
+        min_overlap (float): Minimum overlap between the two mixed sources.
+        target_level (int): Gain level in dB.
+        clipping_threshold (float): Threshold for clipping the audio.
+    Returns:
+        torch.Tensor: The mixed audio, of shape [B, T].
+    """
+    if clean.shape[1] > noise.shape[1]:
+        noise = torch.nn.functional.pad(noise, (0, clean.shape[1] - noise.shape[1]))
+    else:
+        noise = noise[:, :clean.shape[1]]
+
+    # normalizing to -25 dB FS
+    clean = clean / (clean.max(1)[0].abs().unsqueeze(1) + EPS)
+    clean = normalize(clean, target_level)
+    rmsclean = rms_f(clean)
+
+    noise = noise / (noise.max(1)[0].abs().unsqueeze(1) + EPS)
+    noise = normalize(noise, target_level)
+    rmsnoise = rms_f(noise)
+
+    # set the noise level for a given SNR
+    noisescalar = (rmsclean / (10 ** (snr / 20)) / (rmsnoise + EPS)).unsqueeze(1)
+    noisenewlevel = noise * noisescalar
+
+    # mix noise and clean speech
+    noisyspeech = mix_pair(clean, noisenewlevel, min_overlap)
+
+    # randomly select RMS value between -15 dBFS and -35 dBFS and normalize noisyspeech with that value
+    # there is a chance of clipping that might happen with very less probability, which is not a major issue.
+    noisy_rms_level = np.random.randint(TARGET_LEVEL_LOWER, TARGET_LEVEL_UPPER)
+    rmsnoisy = rms_f(noisyspeech)
+    scalarnoisy = (10 ** (noisy_rms_level / 20) / (rmsnoisy + EPS)).unsqueeze(1)
+    noisyspeech = noisyspeech * scalarnoisy
+    clean = clean * scalarnoisy
+    noisenewlevel = noisenewlevel * scalarnoisy
+
+    # final check to see if there are any amplitudes exceeding +/- 1. If so, normalize all the signals accordingly
+    clipped = is_clipped(noisyspeech)
+    if clipped.any():
+        noisyspeech_maxamplevel = noisyspeech[clipped].max(1)[0].abs().unsqueeze(1) / (clipping_threshold - EPS)
+        noisyspeech[clipped] = noisyspeech[clipped] / noisyspeech_maxamplevel
+
+    return noisyspeech
+
+
+def snr_mix(src: torch.Tensor, dst: torch.Tensor, snr_low: int, snr_high: int, min_overlap: float):
+    if snr_low == snr_high:
+        snr = snr_low
+    else:
+        snr = np.random.randint(snr_low, snr_high)
+    mix = snr_mixer(src, dst, snr, min_overlap)
+    return mix
+
+
+def mix_text(src_text: str, dst_text: str):
+    """Mix text from different sources by concatenating them."""
+    if src_text == dst_text:
+        return src_text
+    return src_text + " " + dst_text
+
+
+def mix_samples(wavs: torch.Tensor, infos: tp.List[SoundInfo], aug_p: float, mix_p: float,
+                snr_low: int, snr_high: int, min_overlap: float):
+    """Mix samples within a batch, summing the waveforms and concatenating the text infos.
+
+    Args:
+        wavs (torch.Tensor): Audio tensors of shape [B, C, T].
+        infos (list[SoundInfo]): List of SoundInfo items corresponding to the audio.
+        aug_p (float): Augmentation probability.
+        mix_p (float): Proportion of items in the batch to mix (and merge) together.
+        snr_low (int): Lowerbound for sampling SNR.
+        snr_high (int): Upperbound for sampling SNR.
+        min_overlap (float): Minimum overlap between mixed samples.
+    Returns:
+        tuple[torch.Tensor, list[SoundInfo]]: A tuple containing the mixed wavs
+            and mixed SoundInfo for the given batch.
+    """
+    # no mixing to perform within the batch
+    if mix_p == 0:
+        return wavs, infos
+
+    if random.uniform(0, 1) < aug_p:
+        # perform all augmentations on waveforms as [B, T]
+        # randomly picking pairs of audio to mix
+        assert wavs.size(1) == 1, f"Mix samples requires monophonic audio but C={wavs.size(1)}"
+        wavs = wavs.mean(dim=1, keepdim=False)
+        B, T = wavs.shape
+        k = int(mix_p * B)
+        mixed_sources_idx = torch.randperm(B)[:k]
+        mixed_targets_idx = torch.randperm(B)[:k]
+        aug_wavs = snr_mix(
+            wavs[mixed_sources_idx],
+            wavs[mixed_targets_idx],
+            snr_low,
+            snr_high,
+            min_overlap,
+        )
+        # mixing textual descriptions in metadata
+        descriptions = [info.description for info in infos]
+        aug_infos = []
+        for i, j in zip(mixed_sources_idx, mixed_targets_idx):
+            text = mix_text(descriptions[i], descriptions[j])
+            m = replace(infos[i])
+            m.description = text
+            aug_infos.append(m)
+
+        # back to [B, C, T]
+        aug_wavs = aug_wavs.unsqueeze(1)
+        assert aug_wavs.shape[0] > 0, "Samples mixing returned empty batch."
+        assert aug_wavs.dim() == 3, f"Returned wav should be [B, C, T] but dim = {aug_wavs.dim()}"
+        assert aug_wavs.shape[0] == len(aug_infos), "Mismatch between number of wavs and infos in the batch"
+
+        return aug_wavs, aug_infos  # [B, C, T]
+    else:
+        # randomly pick samples in the batch to match
+        # the batch size when performing audio mixing
+        B, C, T = wavs.shape
+        k = int(mix_p * B)
+        wav_idx = torch.randperm(B)[:k]
+        wavs = wavs[wav_idx]
+        infos = [infos[i] for i in wav_idx]
+        assert wavs.shape[0] == len(infos), "Mismatch between number of wavs and infos in the batch"
+
+        return wavs, infos  # [B, C, T]

audiocraft/data/zip.py

@@ -0,0 +1,76 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Utility for reading some info from inside a zip file.
+"""
+
+import typing
+import zipfile
+
+from dataclasses import dataclass
+from functools import lru_cache
+from typing_extensions import Literal
+
+
+DEFAULT_SIZE = 32
+MODE = Literal['r', 'w', 'x', 'a']
+
+
+@dataclass(order=True)
+class PathInZip:
+    """Hold a path of file within a zip file.
+
+    Args:
+        path (str): The convention is <path_to_zip>:<relative_path_inside_zip>.
+            Let's assume there is a zip file /some/location/foo.zip
+            and inside of it is a json file located at /data/file1.json,
+            Then we expect path = "/some/location/foo.zip:/data/file1.json".
+    """
+
+    INFO_PATH_SEP = ':'
+    zip_path: str
+    file_path: str
+
+    def __init__(self, path: str) -> None:
+        split_path = path.split(self.INFO_PATH_SEP)
+        assert len(split_path) == 2
+        self.zip_path, self.file_path = split_path
+
+    @classmethod
+    def from_paths(cls, zip_path: str, file_path: str):
+        return cls(zip_path + cls.INFO_PATH_SEP + file_path)
+
+    def __str__(self) -> str:
+        return self.zip_path + self.INFO_PATH_SEP + self.file_path
+
+
+def _open_zip(path: str, mode: MODE = 'r'):
+    return zipfile.ZipFile(path, mode)
+
+
+_cached_open_zip = lru_cache(DEFAULT_SIZE)(_open_zip)
+
+
+def set_zip_cache_size(max_size: int):
+    """Sets the maximal LRU caching for zip file opening.
+
+    Args:
+        max_size (int): the maximal LRU cache.
+    """
+    global _cached_open_zip
+    _cached_open_zip = lru_cache(max_size)(_open_zip)
+
+
+def open_file_in_zip(path_in_zip: PathInZip, mode: str = 'r') -> typing.IO:
+    """Opens a file stored inside a zip and returns a file-like object.
+
+    Args:
+        path_in_zip (PathInZip): A PathInZip object representing the file to return a file-like object of.
+        mode (str): The mode in which to open the file with.
+    Returns:
+        A file-like object for PathInZip.
+    """
+    zf = _cached_open_zip(path_in_zip.zip_path)
+    return zf.open(path_in_zip.file_path)