Spaces:
Running
on
Zero
Running
on
Zero
File size: 8,436 Bytes
9d3cb0a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 |
import torch
from einops import rearrange
from torch import nn
class Pretransform(nn.Module):
def __init__(self, enable_grad, io_channels, is_discrete):
super().__init__()
self.is_discrete = is_discrete
self.io_channels = io_channels
self.encoded_channels = None
self.downsampling_ratio = None
self.enable_grad = enable_grad
def encode(self, x):
raise NotImplementedError
def decode(self, z):
raise NotImplementedError
def tokenize(self, x):
raise NotImplementedError
def decode_tokens(self, tokens):
raise NotImplementedError
class AutoencoderPretransform(Pretransform):
def __init__(self, model, scale=1.0, model_half=False, iterate_batch=False, chunked=False):
super().__init__(enable_grad=False, io_channels=model.io_channels, is_discrete=model.bottleneck is not None and model.bottleneck.is_discrete)
self.model = model
self.model.requires_grad_(False).eval()
self.scale=scale
self.downsampling_ratio = model.downsampling_ratio
self.io_channels = model.io_channels
self.sample_rate = model.sample_rate
self.model_half = model_half
self.iterate_batch = iterate_batch
self.encoded_channels = model.latent_dim
self.chunked = chunked
self.num_quantizers = model.bottleneck.num_quantizers if model.bottleneck is not None and model.bottleneck.is_discrete else None
self.codebook_size = model.bottleneck.codebook_size if model.bottleneck is not None and model.bottleneck.is_discrete else None
if self.model_half:
self.model.half()
def encode(self, x, **kwargs):
if self.model_half:
x = x.half()
self.model.to(torch.float16)
encoded = self.model.encode_audio(x, chunked=self.chunked, iterate_batch=self.iterate_batch, **kwargs)
if self.model_half:
encoded = encoded.float()
return encoded / self.scale
def decode(self, z, **kwargs):
z = z * self.scale
if self.model_half:
z = z.half()
self.model.to(torch.float16)
decoded = self.model.decode_audio(z, chunked=self.chunked, iterate_batch=self.iterate_batch, **kwargs)
if self.model_half:
decoded = decoded.float()
return decoded
def tokenize(self, x, **kwargs):
assert self.model.is_discrete, "Cannot tokenize with a continuous model"
_, info = self.model.encode(x, return_info = True, **kwargs)
return info[self.model.bottleneck.tokens_id]
def decode_tokens(self, tokens, **kwargs):
assert self.model.is_discrete, "Cannot decode tokens with a continuous model"
return self.model.decode_tokens(tokens, **kwargs)
def load_state_dict(self, state_dict, strict=True):
self.model.load_state_dict(state_dict, strict=strict)
class WaveletPretransform(Pretransform):
def __init__(self, channels, levels, wavelet):
super().__init__(enable_grad=False, io_channels=channels, is_discrete=False)
from .wavelets import WaveletEncode1d, WaveletDecode1d
self.encoder = WaveletEncode1d(channels, levels, wavelet)
self.decoder = WaveletDecode1d(channels, levels, wavelet)
self.downsampling_ratio = 2 ** levels
self.io_channels = channels
self.encoded_channels = channels * self.downsampling_ratio
def encode(self, x):
return self.encoder(x)
def decode(self, z):
return self.decoder(z)
class PQMFPretransform(Pretransform):
def __init__(self, attenuation=100, num_bands=16):
# TODO: Fix PQMF to take in in-channels
super().__init__(enable_grad=False, io_channels=1, is_discrete=False)
from .pqmf import PQMF
self.pqmf = PQMF(attenuation, num_bands)
def encode(self, x):
# x is (Batch x Channels x Time)
x = self.pqmf.forward(x)
# pqmf.forward returns (Batch x Channels x Bands x Time)
# but Pretransform needs Batch x Channels x Time
# so concatenate channels and bands into one axis
return rearrange(x, "b c n t -> b (c n) t")
def decode(self, x):
# x is (Batch x (Channels Bands) x Time), convert back to (Batch x Channels x Bands x Time)
x = rearrange(x, "b (c n) t -> b c n t", n=self.pqmf.num_bands)
# returns (Batch x Channels x Time)
return self.pqmf.inverse(x)
class PretrainedDACPretransform(Pretransform):
def __init__(self, model_type="44khz", model_bitrate="8kbps", scale=1.0, quantize_on_decode: bool = True, chunked=True):
super().__init__(enable_grad=False, io_channels=1, is_discrete=True)
import dac
model_path = dac.utils.download(model_type=model_type, model_bitrate=model_bitrate)
self.model = dac.DAC.load(model_path)
self.quantize_on_decode = quantize_on_decode
if model_type == "44khz":
self.downsampling_ratio = 512
else:
self.downsampling_ratio = 320
self.io_channels = 1
self.scale = scale
self.chunked = chunked
self.encoded_channels = self.model.latent_dim
self.num_quantizers = self.model.n_codebooks
self.codebook_size = self.model.codebook_size
def encode(self, x):
latents = self.model.encoder(x)
if self.quantize_on_decode:
output = latents
else:
z, _, _, _, _ = self.model.quantizer(latents, n_quantizers=self.model.n_codebooks)
output = z
if self.scale != 1.0:
output = output / self.scale
return output
def decode(self, z):
if self.scale != 1.0:
z = z * self.scale
if self.quantize_on_decode:
z, _, _, _, _ = self.model.quantizer(z, n_quantizers=self.model.n_codebooks)
return self.model.decode(z)
def tokenize(self, x):
return self.model.encode(x)[1]
def decode_tokens(self, tokens):
latents = self.model.quantizer.from_codes(tokens)
return self.model.decode(latents)
class AudiocraftCompressionPretransform(Pretransform):
def __init__(self, model_type="facebook/encodec_32khz", scale=1.0, quantize_on_decode: bool = True):
super().__init__(enable_grad=False, io_channels=1, is_discrete=True)
try:
from audiocraft.models import CompressionModel
except ImportError:
raise ImportError("Audiocraft is not installed. Please install audiocraft to use Audiocraft models.")
self.model = CompressionModel.get_pretrained(model_type)
self.quantize_on_decode = quantize_on_decode
self.downsampling_ratio = round(self.model.sample_rate / self.model.frame_rate)
self.sample_rate = self.model.sample_rate
self.io_channels = self.model.channels
self.scale = scale
#self.encoded_channels = self.model.latent_dim
self.num_quantizers = self.model.num_codebooks
self.codebook_size = self.model.cardinality
self.model.to(torch.float16).eval().requires_grad_(False)
def encode(self, x):
assert False, "Audiocraft compression models do not support continuous encoding"
# latents = self.model.encoder(x)
# if self.quantize_on_decode:
# output = latents
# else:
# z, _, _, _, _ = self.model.quantizer(latents, n_quantizers=self.model.n_codebooks)
# output = z
# if self.scale != 1.0:
# output = output / self.scale
# return output
def decode(self, z):
assert False, "Audiocraft compression models do not support continuous decoding"
# if self.scale != 1.0:
# z = z * self.scale
# if self.quantize_on_decode:
# z, _, _, _, _ = self.model.quantizer(z, n_quantizers=self.model.n_codebooks)
# return self.model.decode(z)
def tokenize(self, x):
with torch.cuda.amp.autocast(enabled=False):
return self.model.encode(x.to(torch.float16))[0]
def decode_tokens(self, tokens):
with torch.cuda.amp.autocast(enabled=False):
return self.model.decode(tokens)
|