audio diffusion pytorch
v0.1.3
PyTorch 用の完全な機能を備えたオーディオ拡散ライブラリ。無条件オーディオ生成、テキスト条件付きオーディオ生成、拡散自動エンコーディング、アップサンプリング、およびボコーディングのモデルが含まれています。提供されるモデルは波形ベースですが、U-Net ( a-unetを使用して構築)、 DiffusionModel 、拡散メソッド、および拡散サンプラーはいずれもあらゆる次元に汎用的であり、他の形式で動作するように高度にカスタマイズ可能です。注: (1) ここでは事前トレーニングされたモデルは提供されていません。(2) 示されている構成は参考であり、テストされていません。論文で使用されている構成については Moûsai を参照してください。
pip install audio-diffusion-pytorch 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 10-100t5-baseテキスト埋め込みによる生成を条件付けるテキストからオーディオへの拡散モデルには、 pip install transformersが必要です。
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
)拡散を使用して、オーディオを低いサンプル レートから高いサンプル レートにアップサンプリングします (例: 3kHz から 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]拡散を使用してメル スペクトログラムを wavefrom に変換します。
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]拡散を使用してオーディオを圧縮潜在に自動エンコードします。 EncoderBaseクラスをサブクラス化するか、 out_channelsおよびdownsample_factorフィールドを含む限り、任意のエンコーダーを提供できます。
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 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] DDPMの普及
@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サンプラー）
@misc { 2010.02502 ,
Author = { Jiaming Song and Chenlin Meng and Stefano Ermon } ,
Title = { Denoising Diffusion Implicit Models } ,
Year = { 2020 } ,
Eprint = { arXiv:2010.02502 } ,
}V-ディフュージョン
@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 テキスト コンディショニング)
@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 } ,
}
