audio diffusion pytorch

适用于 PyTorch 的功能齐全的音频扩散库。包括无条件音频生成、文本条件音频生成、扩散自动编码、上采样和声码的模型。提供的模型是基于波形的，但是，U-Net（使用a-unet构建）、 DiffusionModel 、扩散方法和扩散采样器对于任何维度都是通用的，并且可以高度定制以处理其他格式。注意：（1）这里没有提供预先训练的模型，（2）显示的配置是指示性的且未经测试，有关本文中使用的配置，请参阅 Moûsai。

安装

pip install audio-diffusion-pytorch

用法

无条件生成器

 from audio_diffusion_pytorch import DiffusionModel , UNetV0 , VDiffusion , VSampler

model = DiffusionModel (
    net_t = UNetV0 , # The model type used for diffusion (U-Net V0 in this case)
    in_channels = 2 , # U-Net: number of input/output (audio) channels
    channels = [ 8 , 32 , 64 , 128 , 256 , 512 , 512 , 1024 , 1024 ], # U-Net: channels at each layer
    factors = [ 1 , 4 , 4 , 4 , 2 , 2 , 2 , 2 , 2 ], # U-Net: downsampling and upsampling factors at each layer
    items = [ 1 , 2 , 2 , 2 , 2 , 2 , 2 , 4 , 4 ], # U-Net: number of repeating items at each layer
    attentions = [ 0 , 0 , 0 , 0 , 0 , 1 , 1 , 1 , 1 ], # U-Net: attention enabled/disabled at each layer
    attention_heads = 8 , # U-Net: number of attention heads per attention item
    attention_features = 64 , # U-Net: number of attention features per attention item
    diffusion_t = VDiffusion , # The diffusion method used
    sampler_t = VSampler , # The diffusion sampler used
)

# Train model with audio waveforms
audio = torch . randn ( 1 , 2 , 2 ** 18 ) # [batch_size, in_channels, length]
loss = model ( audio )
loss . backward ()

# Turn noise into new audio sample with diffusion
noise = torch . randn ( 1 , 2 , 2 ** 18 ) # [batch_size, in_channels, length]
sample = model . sample ( noise , num_steps = 10 ) # Suggested num_steps 10-100

文本条件生成器

使用t5-base文本嵌入来调节生成的文本到音频扩散模型需要pip install transformers 。

 from audio_diffusion_pytorch import DiffusionModel , UNetV0 , VDiffusion , VSampler

model = DiffusionModel (
    # ... same as unconditional model
    use_text_conditioning = True , # U-Net: enables text conditioning (default T5-base)
    use_embedding_cfg = True , # U-Net: enables classifier free guidance
    embedding_max_length = 64 , # U-Net: text embedding maximum length (default for T5-base)
    embedding_features = 768 , # U-Net: text mbedding features (default for T5-base)
    cross_attentions = [ 0 , 0 , 0 , 1 , 1 , 1 , 1 , 1 , 1 ], # U-Net: cross-attention enabled/disabled at each layer
)

# Train model with audio waveforms
audio_wave = torch . randn ( 1 , 2 , 2 ** 18 ) # [batch, in_channels, length]
loss = model (
    audio_wave ,
    text = [ 'The audio description' ], # Text conditioning, one element per batch
    embedding_mask_proba = 0.1 # Probability of masking text with learned embedding (Classifier-Free Guidance Mask)
)
loss . backward ()

# Turn noise into new audio sample with diffusion
noise = torch . randn ( 1 , 2 , 2 ** 18 )
sample = model . sample (
    noise ,
    text = [ 'The audio description' ],
    embedding_scale = 5.0 , # Higher for more text importance, suggested range: 1-15 (Classifier-Free Guidance Scale)
    num_steps = 2 # Higher for better quality, suggested num_steps: 10-100
)

扩散上采样器

使用扩散将音频从较低采样率上采样到较高采样率，例如 3kHz 到 48kHz。

 from audio_diffusion_pytorch import DiffusionUpsampler , UNetV0 , VDiffusion , VSampler

upsampler = DiffusionUpsampler (
    net_t = UNetV0 , # The model type used for diffusion
    upsample_factor = 16 , # The upsample factor (e.g. 16 can be used for 3kHz to 48kHz)
    in_channels = 2 , # U-Net: number of input/output (audio) channels
    channels = [ 8 , 32 , 64 , 128 , 256 , 512 , 512 , 1024 , 1024 ], # U-Net: channels at each layer
    factors = [ 1 , 4 , 4 , 4 , 2 , 2 , 2 , 2 , 2 ], # U-Net: downsampling and upsampling factors at each layer
    items = [ 1 , 2 , 2 , 2 , 2 , 2 , 2 , 4 , 4 ], # U-Net: number of repeating items at each layer
    diffusion_t = VDiffusion , # The diffusion method used
    sampler_t = VSampler , # The diffusion sampler used
)

# Train model with high sample rate audio waveforms
audio = torch . randn ( 1 , 2 , 2 ** 18 ) # [batch, in_channels, length]
loss = upsampler ( audio )
loss . backward ()

# Turn low sample rate audio into high sample rate
downsampled_audio = torch . randn ( 1 , 2 , 2 ** 14 ) # [batch, in_channels, length]
sample = upsampler . sample ( downsampled_audio , num_steps = 10 ) # Output has shape: [1, 2, 2**18]

扩散声码器

使用扩散将梅尔谱图转换为波谱。

 from audio_diffusion_pytorch import DiffusionVocoder , UNetV0 , VDiffusion , VSampler

vocoder = DiffusionVocoder (
    mel_n_fft = 1024 , # Mel-spectrogram n_fft
    mel_channels = 80 , # Mel-spectrogram channels
    mel_sample_rate = 48000 , # Mel-spectrogram sample rate
    mel_normalize_log = True , # Mel-spectrogram log normalization (alternative is mel_normalize=True for [-1,1] power normalization)
    net_t = UNetV0 , # The model type used for diffusion vocoding
    channels = [ 8 , 32 , 64 , 128 , 256 , 512 , 512 , 1024 , 1024 ], # U-Net: channels at each layer
    factors = [ 1 , 4 , 4 , 4 , 2 , 2 , 2 , 2 , 2 ], # U-Net: downsampling and upsampling factors at each layer
    items = [ 1 , 2 , 2 , 2 , 2 , 2 , 2 , 4 , 4 ], # U-Net: number of repeating items at each layer
    diffusion_t = VDiffusion , # The diffusion method used
    sampler_t = VSampler , # The diffusion sampler used
)

# Train model on waveforms (automatically converted to mel internally)
audio = torch . randn ( 1 , 2 , 2 ** 18 ) # [batch, in_channels, length]
loss = vocoder ( audio )
loss . backward ()

# Turn mel spectrogram into waveform
mel_spectrogram = torch . randn ( 1 , 2 , 80 , 1024 ) # [batch, in_channels, mel_channels, mel_length]
sample = vocoder . sample ( mel_spectrogram , num_steps = 10 ) # Output has shape: [1, 2, 2**18]

扩散自动编码器

使用扩散将音频自动编码为压缩潜在音频。可以提供任何编码器，只要它是EncoderBase类的子类或包含out_channels和downsample_factor字段。

 from audio_diffusion_pytorch import DiffusionAE , UNetV0 , VDiffusion , VSampler
from audio_encoders_pytorch import MelE1d , TanhBottleneck

autoencoder = DiffusionAE (
    encoder = MelE1d ( # The encoder used, in this case a mel-spectrogram encoder
        in_channels = 2 ,
        channels = 512 ,
        multipliers = [ 1 , 1 ],
        factors = [ 2 ],
        num_blocks = [ 12 ],
        out_channels = 32 ,
        mel_channels = 80 ,
        mel_sample_rate = 48000 ,
        mel_normalize_log = True ,
        bottleneck = TanhBottleneck (),
    ),
    inject_depth = 6 ,
    net_t = UNetV0 , # The model type used for diffusion upsampling
    in_channels = 2 , # U-Net: number of input/output (audio) channels
    channels = [ 8 , 32 , 64 , 128 , 256 , 512 , 512 , 1024 , 1024 ], # U-Net: channels at each layer
    factors = [ 1 , 4 , 4 , 4 , 2 , 2 , 2 , 2 , 2 ], # U-Net: downsampling and upsampling factors at each layer
    items = [ 1 , 2 , 2 , 2 , 2 , 2 , 2 , 4 , 4 ], # U-Net: number of repeating items at each layer
    diffusion_t = VDiffusion , # The diffusion method used
    sampler_t = VSampler , # The diffusion sampler used
)

# Train autoencoder with audio samples
audio = torch . randn ( 1 , 2 , 2 ** 18 ) # [batch, in_channels, length]
loss = autoencoder ( audio )
loss . backward ()

# Encode/decode audio
audio = torch . randn ( 1 , 2 , 2 ** 18 ) # [batch, in_channels, length]
latent = autoencoder . encode ( audio ) # Encode
sample = autoencoder . decode ( latent , num_steps = 10 ) # Decode by sampling diffusion model conditioning on latent 

其他

修复

 from audio_diffusion_pytorch import UNetV0 , VInpainter

# The diffusion UNetV0 (this is an example, the net must be trained to work)
net = UNetV0 (
    dim = 1 ,
    in_channels = 2 , # U-Net: number of input/output (audio) channels
    channels = [ 8 , 32 , 64 , 128 , 256 , 512 , 512 , 1024 , 1024 ], # U-Net: channels at each layer
    factors = [ 1 , 4 , 4 , 4 , 2 , 2 , 2 , 2 , 2 ], # U-Net: downsampling and upsampling factors at each layer
    items = [ 1 , 2 , 2 , 2 , 2 , 2 , 2 , 4 , 4 ], # U-Net: number of repeating items at each layer
    attentions = [ 0 , 0 , 0 , 0 , 0 , 1 , 1 , 1 , 1 ], # U-Net: attention enabled/disabled at each layer
    attention_heads = 8 , # U-Net: number of attention heads per attention block
    attention_features = 64 , # U-Net: number of attention features per attention block,
)

# Instantiate inpainter with trained net
inpainter = VInpainter ( net = net )

# Inpaint source
y = inpainter (
    source = torch . randn ( 1 , 2 , 2 ** 18 ), # Start source
    mask = torch . randint ( 0 , 2 , ( 1 , 2 , 2 ** 18 ), dtype = torch . bool ),  # Set to `True` the parts you want to keep
    num_steps = 10 , # Number of inpainting steps
    num_resamples = 2 , # Number of resampling steps
    show_progress = True ,
) # [1, 2, 2 ** 18] 

欣赏

• StabilityAI 用于计算，Zach Evans 和 HarmonAI 的其他人进行了有趣的研究讨论。
• 苏黎世联邦理工学院提供资源，Zhijing Jin、Bernhard Schoelkopf 和 Mrinmaya Sachan 指导本论文。
• Phil Wang 在扩散和 Imagen 方面做出了精彩的开源贡献。
• 凯瑟琳·克劳森 (Katherine Crowson) 的 k 扩散实验和疯狂的采样器集合。

引文

DDPM扩散

 @misc { 2006.11239 ,
Author = { Jonathan Ho and Ajay Jain and Pieter Abbeel } ,
Title = { Denoising Diffusion Probabilistic Models } ,
Year = { 2020 } ,
Eprint = { arXiv:2006.11239 } ,
}

DDIM（V 采样器）

 @misc { 2010.02502 ,
Author = { Jiaming Song and Chenlin Meng and Stefano Ermon } ,
Title = { Denoising Diffusion Implicit Models } ,
Year = { 2020 } ,
Eprint = { arXiv:2010.02502 } ,
}

V-扩散

 @misc { 2202.00512 ,
Author = { Tim Salimans and Jonathan Ho } ,
Title = { Progressive Distillation for Fast Sampling of Diffusion Models } ,
Year = { 2022 } ,
Eprint = { arXiv:2202.00512 } ,
}

Imagen（T5 文本调节）

 @misc { 2205.11487 ,
Author = { Chitwan Saharia and William Chan and Saurabh Saxena and Lala Li and Jay Whang and Emily Denton and Seyed Kamyar Seyed Ghasemipour and Burcu Karagol Ayan and S. Sara Mahdavi and Rapha Gontijo Lopes and Tim Salimans and Jonathan Ho and David J Fleet and Mohammad Norouzi } ,
Title = { Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding } ,
Year = { 2022 } ,
Eprint = { arXiv:2205.11487 } ,
}