File size: 8,417 Bytes
bcdb559 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 |
from abc import ABC, abstractmethod
from math import floor
from typing import Any, Callable, Optional, Sequence, Tuple, Union
import torch
from einops import pack, rearrange, unpack
from torch import Generator, Tensor, nn
from .components import AppendChannelsPlugin, MelSpectrogram
from .diffusion import ARVDiffusion, ARVSampler, VDiffusion, VSampler
from .utils import (
closest_power_2,
default,
downsample,
exists,
groupby,
randn_like,
upsample,
)
class DiffusionModel(nn.Module):
def __init__(
self,
net_t: Callable,
diffusion_t: Callable = VDiffusion,
sampler_t: Callable = VSampler,
loss_fn: Callable = torch.nn.functional.mse_loss,
dim: int = 1,
**kwargs,
):
super().__init__()
diffusion_kwargs, kwargs = groupby("diffusion_", kwargs)
sampler_kwargs, kwargs = groupby("sampler_", kwargs)
self.net = net_t(dim=dim, **kwargs)
self.diffusion = diffusion_t(net=self.net, loss_fn=loss_fn, **diffusion_kwargs)
self.sampler = sampler_t(net=self.net, **sampler_kwargs)
def forward(self, *args, **kwargs) -> Tensor:
return self.diffusion(*args, **kwargs)
@torch.no_grad()
def sample(self, *args, **kwargs) -> Tensor:
return self.sampler(*args, **kwargs)
class EncoderBase(nn.Module, ABC):
"""Abstract class for DiffusionAE encoder"""
@abstractmethod
def __init__(self):
super().__init__()
self.out_channels = None
self.downsample_factor = None
class AdapterBase(nn.Module, ABC):
"""Abstract class for DiffusionAE encoder"""
@abstractmethod
def encode(self, x: Tensor) -> Tensor:
pass
@abstractmethod
def decode(self, x: Tensor) -> Tensor:
pass
class DiffusionAE(DiffusionModel):
"""Diffusion Auto Encoder"""
def __init__(
self,
in_channels: int,
channels: Sequence[int],
encoder: EncoderBase,
inject_depth: int,
latent_factor: Optional[int] = None,
adapter: Optional[AdapterBase] = None,
**kwargs,
):
context_channels = [0] * len(channels)
context_channels[inject_depth] = encoder.out_channels
super().__init__(
in_channels=in_channels,
channels=channels,
context_channels=context_channels,
**kwargs,
)
self.in_channels = in_channels
self.encoder = encoder
self.inject_depth = inject_depth
# Optional custom latent factor and adapter
self.latent_factor = default(latent_factor, self.encoder.downsample_factor)
self.adapter = adapter.requires_grad_(False) if exists(adapter) else None
def forward( # type: ignore
self, x: Tensor, with_info: bool = False, **kwargs
) -> Union[Tensor, Tuple[Tensor, Any]]:
# Encode input to latent channels
latent, info = self.encode(x, with_info=True)
channels = [None] * self.inject_depth + [latent]
# Adapt input to diffusion if adapter provided
x = self.adapter.encode(x) if exists(self.adapter) else x
# Compute diffusion loss
loss = super().forward(x, channels=channels, **kwargs)
return (loss, info) if with_info else loss
def encode(self, *args, **kwargs):
return self.encoder(*args, **kwargs)
@torch.no_grad()
def decode(
self, latent: Tensor, generator: Optional[Generator] = None, **kwargs
) -> Tensor:
b = latent.shape[0]
noise_length = closest_power_2(latent.shape[2] * self.latent_factor)
# Compute noise by inferring shape from latent length
noise = torch.randn(
(b, self.in_channels, noise_length),
device=latent.device,
dtype=latent.dtype,
generator=generator,
)
# Compute context from latent
channels = [None] * self.inject_depth + [latent] # type: ignore
# Decode by sampling while conditioning on latent channels
out = super().sample(noise, channels=channels, **kwargs)
# Decode output with adapter if provided
return self.adapter.decode(out) if exists(self.adapter) else out
class DiffusionUpsampler(DiffusionModel):
def __init__(
self,
in_channels: int,
upsample_factor: int,
net_t: Callable,
**kwargs,
):
self.upsample_factor = upsample_factor
super().__init__(
net_t=AppendChannelsPlugin(net_t, channels=in_channels),
in_channels=in_channels,
**kwargs,
)
def reupsample(self, x: Tensor) -> Tensor:
x = x.clone()
x = downsample(x, factor=self.upsample_factor)
x = upsample(x, factor=self.upsample_factor)
return x
def forward(self, x: Tensor, *args, **kwargs) -> Tensor: # type: ignore
reupsampled = self.reupsample(x)
return super().forward(x, *args, append_channels=reupsampled, **kwargs)
@torch.no_grad()
def sample( # type: ignore
self, downsampled: Tensor, generator: Optional[Generator] = None, **kwargs
) -> Tensor:
reupsampled = upsample(downsampled, factor=self.upsample_factor)
noise = randn_like(reupsampled, generator=generator)
return super().sample(noise, append_channels=reupsampled, **kwargs)
class DiffusionVocoder(DiffusionModel):
def __init__(
self,
net_t: Callable,
mel_channels: int,
mel_n_fft: int,
mel_hop_length: Optional[int] = None,
mel_win_length: Optional[int] = None,
in_channels: int = 1, # Ignored: channels are automatically batched.
**kwargs,
):
mel_hop_length = default(mel_hop_length, floor(mel_n_fft) // 4)
mel_win_length = default(mel_win_length, mel_n_fft)
mel_kwargs, kwargs = groupby("mel_", kwargs)
super().__init__(
net_t=AppendChannelsPlugin(net_t, channels=1),
in_channels=1,
**kwargs,
)
self.to_spectrogram = MelSpectrogram(
n_fft=mel_n_fft,
hop_length=mel_hop_length,
win_length=mel_win_length,
n_mel_channels=mel_channels,
**mel_kwargs,
)
self.to_flat = nn.ConvTranspose1d(
in_channels=mel_channels,
out_channels=1,
kernel_size=mel_win_length,
stride=mel_hop_length,
padding=(mel_win_length - mel_hop_length) // 2,
bias=False,
)
def forward(self, x: Tensor, *args, **kwargs) -> Tensor: # type: ignore
# Get spectrogram, pack channels and flatten
spectrogram = rearrange(self.to_spectrogram(x), "b c f l -> (b c) f l")
spectrogram_flat = self.to_flat(spectrogram)
# Pack wave channels
x = rearrange(x, "b c t -> (b c) 1 t")
return super().forward(x, *args, append_channels=spectrogram_flat, **kwargs)
@torch.no_grad()
def sample( # type: ignore
self, spectrogram: Tensor, generator: Optional[Generator] = None, **kwargs
) -> Tensor: # type: ignore
# Pack channels and flatten spectrogram
spectrogram, ps = pack([spectrogram], "* f l")
spectrogram_flat = self.to_flat(spectrogram)
# Get start noise and sample
noise = randn_like(spectrogram_flat, generator=generator)
waveform = super().sample(noise, append_channels=spectrogram_flat, **kwargs)
# Unpack wave channels
waveform = rearrange(waveform, "... 1 t -> ... t")
waveform = unpack(waveform, ps, "* t")[0]
return waveform
class DiffusionAR(DiffusionModel):
def __init__(
self,
in_channels: int,
length: int,
num_splits: int,
diffusion_t: Callable = ARVDiffusion,
sampler_t: Callable = ARVSampler,
**kwargs,
):
super().__init__(
in_channels=in_channels + 1,
out_channels=in_channels,
diffusion_t=diffusion_t,
diffusion_length=length,
diffusion_num_splits=num_splits,
sampler_t=sampler_t,
sampler_in_channels=in_channels,
sampler_length=length,
sampler_num_splits=num_splits,
use_time_conditioning=False,
use_modulation=False,
**kwargs,
)
|