Spaces:
Running
on
Zero
Running
on
Zero
File size: 4,860 Bytes
c8448bc |
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 |
from einops import rearrange
from torch import nn
class Pretransform(nn.Module):
def __init__(self, enable_grad=False, io_channels=2, ):
super().__init__()
self.io_channels = io_channels
self.encoded_channels = None
self.downsampling_ratio = None
self.enable_grad = enable_grad
def encode(self, x):
return x
def decode(self, z):
return z
class AutoencoderPretransform(Pretransform):
def __init__(self, model, scale=1.0, model_half=False, iterate_batch=False):
super().__init__()
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
if self.model_half:
self.model.half()
def encode(self, x, **kwargs):
# print(f'encoder takes input {x.shape}')
if self.model_half:
x = x.half()
encoded = self.model.encode(x, iterate_batch=self.iterate_batch, **kwargs)
if self.model_half:
encoded = encoded.float()
out = encoded / self.scale
# print(f'encoder out: {out.shape}')
return encoded / self.scale
def decode(self, z, **kwargs):
z = z * self.scale
if self.model_half:
z = z.half()
decoded = self.model.decode(z, iterate_batch=self.iterate_batch, **kwargs)
if self.model_half:
decoded = decoded.float()
return decoded
def load_state_dict(self, state_dict, strict=True):
# print(f'load state dict {state_dict}')
self.model.load_state_dict(state_dict, strict=strict)
class WaveletPretransform(Pretransform):
def __init__(self, channels, levels, wavelet):
super().__init__()
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):
super().__init__()
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__()
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
def encode(self, x):
# print(f"Input to DAC encoder shape: {x.shape}, type: {x.dtype}")
latents = self.model.encoder(x)
# print(f"Latents shape after DAC encoder: {latents.shape}")
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
# print(f'output from DAC encoder: {x.shape}')
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)
|