Upload 3 files
Browse files- audiocraft/quantization/base.py +99 -0
- audiocraft/quantization/core_vq.py +405 -0
- audiocraft/quantization/vq.py +115 -0
audiocraft/quantization/base.py
ADDED
|
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
| 2 |
+
# All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# This source code is licensed under the license found in the
|
| 5 |
+
# LICENSE file in the root directory of this source tree.
|
| 6 |
+
|
| 7 |
+
"""
|
| 8 |
+
Base class for all quantizers.
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
from dataclasses import dataclass, field
|
| 12 |
+
import typing as tp
|
| 13 |
+
|
| 14 |
+
import torch
|
| 15 |
+
from torch import nn
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
@dataclass
|
| 19 |
+
class QuantizedResult:
|
| 20 |
+
x: torch.Tensor
|
| 21 |
+
codes: torch.Tensor
|
| 22 |
+
bandwidth: torch.Tensor # bandwidth in kb/s used, per batch item.
|
| 23 |
+
penalty: tp.Optional[torch.Tensor] = None
|
| 24 |
+
metrics: dict = field(default_factory=dict)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
class BaseQuantizer(nn.Module):
|
| 28 |
+
"""Base class for quantizers.
|
| 29 |
+
"""
|
| 30 |
+
|
| 31 |
+
def forward(self, x: torch.Tensor, frame_rate: int) -> QuantizedResult:
|
| 32 |
+
"""
|
| 33 |
+
Given input tensor x, returns first the quantized (or approximately quantized)
|
| 34 |
+
representation along with quantized codes, bandwidth, and any penalty term for the loss.
|
| 35 |
+
Finally, this returns a dict of metrics to update logging etc.
|
| 36 |
+
Frame rate must be passed so that the bandwidth is properly computed.
|
| 37 |
+
"""
|
| 38 |
+
raise NotImplementedError()
|
| 39 |
+
|
| 40 |
+
def encode(self, x: torch.Tensor) -> torch.Tensor:
|
| 41 |
+
"""Encode a given input tensor with the specified sample rate at the given bandwidth."""
|
| 42 |
+
raise NotImplementedError()
|
| 43 |
+
|
| 44 |
+
def decode(self, codes: torch.Tensor) -> torch.Tensor:
|
| 45 |
+
"""Decode the given codes to the quantized representation."""
|
| 46 |
+
raise NotImplementedError()
|
| 47 |
+
|
| 48 |
+
@property
|
| 49 |
+
def total_codebooks(self):
|
| 50 |
+
"""Total number of codebooks."""
|
| 51 |
+
raise NotImplementedError()
|
| 52 |
+
|
| 53 |
+
@property
|
| 54 |
+
def num_codebooks(self):
|
| 55 |
+
"""Number of active codebooks."""
|
| 56 |
+
raise NotImplementedError()
|
| 57 |
+
|
| 58 |
+
def set_num_codebooks(self, n: int):
|
| 59 |
+
"""Set the number of active codebooks."""
|
| 60 |
+
raise NotImplementedError()
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
class DummyQuantizer(BaseQuantizer):
|
| 64 |
+
"""Fake quantizer that actually does not perform any quantization.
|
| 65 |
+
"""
|
| 66 |
+
def __init__(self):
|
| 67 |
+
super().__init__()
|
| 68 |
+
|
| 69 |
+
def forward(self, x: torch.Tensor, frame_rate: int):
|
| 70 |
+
q = x.unsqueeze(1)
|
| 71 |
+
return QuantizedResult(x, q, torch.tensor(q.numel() * 32 * frame_rate / 1000 / len(x)).to(x))
|
| 72 |
+
|
| 73 |
+
def encode(self, x: torch.Tensor) -> torch.Tensor:
|
| 74 |
+
"""Encode a given input tensor with the specified sample rate at the given bandwidth.
|
| 75 |
+
In the case of the DummyQuantizer, the codes are actually identical
|
| 76 |
+
to the input and resulting quantized representation as no quantization is done.
|
| 77 |
+
"""
|
| 78 |
+
return x.unsqueeze(1)
|
| 79 |
+
|
| 80 |
+
def decode(self, codes: torch.Tensor) -> torch.Tensor:
|
| 81 |
+
"""Decode the given codes to the quantized representation.
|
| 82 |
+
In the case of the DummyQuantizer, the codes are actually identical
|
| 83 |
+
to the input and resulting quantized representation as no quantization is done.
|
| 84 |
+
"""
|
| 85 |
+
return codes.squeeze(1)
|
| 86 |
+
|
| 87 |
+
@property
|
| 88 |
+
def total_codebooks(self):
|
| 89 |
+
"""Total number of codebooks."""
|
| 90 |
+
return 1
|
| 91 |
+
|
| 92 |
+
@property
|
| 93 |
+
def num_codebooks(self):
|
| 94 |
+
"""Total number of codebooks."""
|
| 95 |
+
return self.total_codebooks
|
| 96 |
+
|
| 97 |
+
def set_num_codebooks(self, n: int):
|
| 98 |
+
"""Set the number of active codebooks."""
|
| 99 |
+
raise AttributeError("Cannot override the number of codebooks for the dummy quantizer")
|
audiocraft/quantization/core_vq.py
ADDED
|
@@ -0,0 +1,405 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
| 2 |
+
# All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# This source code is licensed under the license found in the
|
| 5 |
+
# LICENSE file in the root directory of this source tree.
|
| 6 |
+
|
| 7 |
+
import typing as tp
|
| 8 |
+
|
| 9 |
+
from einops import rearrange, repeat
|
| 10 |
+
import flashy
|
| 11 |
+
import torch
|
| 12 |
+
from torch import nn, einsum
|
| 13 |
+
import torch.nn.functional as F
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def exists(val: tp.Optional[tp.Any]) -> bool:
|
| 17 |
+
return val is not None
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def default(val: tp.Any, d: tp.Any) -> tp.Any:
|
| 21 |
+
return val if exists(val) else d
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def l2norm(t):
|
| 25 |
+
return F.normalize(t, p=2, dim=-1)
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
def ema_inplace(moving_avg, new, decay: float):
|
| 29 |
+
moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def laplace_smoothing(x, n_categories: int, epsilon: float = 1e-5):
|
| 33 |
+
return (x + epsilon) / (x.sum() + n_categories * epsilon)
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
def uniform_init(*shape: int):
|
| 37 |
+
t = torch.empty(shape)
|
| 38 |
+
nn.init.kaiming_uniform_(t)
|
| 39 |
+
return t
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
def sample_vectors(samples, num: int):
|
| 43 |
+
num_samples, device = samples.shape[0], samples.device
|
| 44 |
+
|
| 45 |
+
if num_samples >= num:
|
| 46 |
+
indices = torch.randperm(num_samples, device=device)[:num]
|
| 47 |
+
else:
|
| 48 |
+
indices = torch.randint(0, num_samples, (num,), device=device)
|
| 49 |
+
|
| 50 |
+
return samples[indices]
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def kmeans(samples, num_clusters: int, num_iters: int = 10):
|
| 54 |
+
dim, dtype = samples.shape[-1], samples.dtype
|
| 55 |
+
|
| 56 |
+
means = sample_vectors(samples, num_clusters)
|
| 57 |
+
|
| 58 |
+
for _ in range(num_iters):
|
| 59 |
+
diffs = rearrange(samples, "n d -> n () d") - rearrange(
|
| 60 |
+
means, "c d -> () c d"
|
| 61 |
+
)
|
| 62 |
+
dists = -(diffs ** 2).sum(dim=-1)
|
| 63 |
+
|
| 64 |
+
buckets = dists.max(dim=-1).indices
|
| 65 |
+
bins = torch.bincount(buckets, minlength=num_clusters)
|
| 66 |
+
zero_mask = bins == 0
|
| 67 |
+
bins_min_clamped = bins.masked_fill(zero_mask, 1)
|
| 68 |
+
|
| 69 |
+
new_means = buckets.new_zeros(num_clusters, dim, dtype=dtype)
|
| 70 |
+
new_means.scatter_add_(0, repeat(buckets, "n -> n d", d=dim), samples)
|
| 71 |
+
new_means = new_means / bins_min_clamped[..., None]
|
| 72 |
+
|
| 73 |
+
means = torch.where(zero_mask[..., None], means, new_means)
|
| 74 |
+
|
| 75 |
+
return means, bins
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
def orthogonal_loss_fn(t):
|
| 79 |
+
# eq (2) from https://arxiv.org/abs/2112.00384
|
| 80 |
+
n = t.shape[0]
|
| 81 |
+
normed_codes = l2norm(t)
|
| 82 |
+
identity = torch.eye(n, device=t.device)
|
| 83 |
+
cosine_sim = einsum("i d, j d -> i j", normed_codes, normed_codes)
|
| 84 |
+
return ((cosine_sim - identity) ** 2).sum() / (n ** 2)
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
class EuclideanCodebook(nn.Module):
|
| 88 |
+
"""Codebook with Euclidean distance.
|
| 89 |
+
|
| 90 |
+
Args:
|
| 91 |
+
dim (int): Dimension.
|
| 92 |
+
codebook_size (int): Codebook size.
|
| 93 |
+
kmeans_init (bool): Whether to use k-means to initialize the codebooks.
|
| 94 |
+
If set to true, run the k-means algorithm on the first training batch and use
|
| 95 |
+
the learned centroids as initialization.
|
| 96 |
+
kmeans_iters (int): Number of iterations used for k-means algorithm at initialization.
|
| 97 |
+
decay (float): Decay for exponential moving average over the codebooks.
|
| 98 |
+
epsilon (float): Epsilon value for numerical stability.
|
| 99 |
+
threshold_ema_dead_code (int): Threshold for dead code expiration. Replace any codes
|
| 100 |
+
that have an exponential moving average cluster size less than the specified threshold with
|
| 101 |
+
randomly selected vector from the current batch.
|
| 102 |
+
"""
|
| 103 |
+
def __init__(
|
| 104 |
+
self,
|
| 105 |
+
dim: int,
|
| 106 |
+
codebook_size: int,
|
| 107 |
+
kmeans_init: int = False,
|
| 108 |
+
kmeans_iters: int = 10,
|
| 109 |
+
decay: float = 0.8,
|
| 110 |
+
epsilon: float = 1e-5,
|
| 111 |
+
threshold_ema_dead_code: int = 2,
|
| 112 |
+
):
|
| 113 |
+
super().__init__()
|
| 114 |
+
self.decay = decay
|
| 115 |
+
init_fn: tp.Union[tp.Callable[..., torch.Tensor], tp.Any] = uniform_init if not kmeans_init else torch.zeros
|
| 116 |
+
embed = init_fn(codebook_size, dim)
|
| 117 |
+
|
| 118 |
+
self.codebook_size = codebook_size
|
| 119 |
+
|
| 120 |
+
self.kmeans_iters = kmeans_iters
|
| 121 |
+
self.epsilon = epsilon
|
| 122 |
+
self.threshold_ema_dead_code = threshold_ema_dead_code
|
| 123 |
+
|
| 124 |
+
self.register_buffer("inited", torch.Tensor([not kmeans_init]))
|
| 125 |
+
self.register_buffer("cluster_size", torch.zeros(codebook_size))
|
| 126 |
+
self.register_buffer("embed", embed)
|
| 127 |
+
self.register_buffer("embed_avg", embed.clone())
|
| 128 |
+
|
| 129 |
+
@torch.jit.ignore
|
| 130 |
+
def init_embed_(self, data):
|
| 131 |
+
if self.inited:
|
| 132 |
+
return
|
| 133 |
+
|
| 134 |
+
embed, cluster_size = kmeans(data, self.codebook_size, self.kmeans_iters)
|
| 135 |
+
self.embed.data.copy_(embed)
|
| 136 |
+
self.embed_avg.data.copy_(embed.clone())
|
| 137 |
+
self.cluster_size.data.copy_(cluster_size)
|
| 138 |
+
self.inited.data.copy_(torch.Tensor([True]))
|
| 139 |
+
# Make sure all buffers across workers are in sync after initialization
|
| 140 |
+
flashy.distrib.broadcast_tensors(self.buffers())
|
| 141 |
+
|
| 142 |
+
def replace_(self, samples, mask):
|
| 143 |
+
modified_codebook = torch.where(
|
| 144 |
+
mask[..., None], sample_vectors(samples, self.codebook_size), self.embed
|
| 145 |
+
)
|
| 146 |
+
self.embed.data.copy_(modified_codebook)
|
| 147 |
+
|
| 148 |
+
def expire_codes_(self, batch_samples):
|
| 149 |
+
if self.threshold_ema_dead_code == 0:
|
| 150 |
+
return
|
| 151 |
+
|
| 152 |
+
expired_codes = self.cluster_size < self.threshold_ema_dead_code
|
| 153 |
+
if not torch.any(expired_codes):
|
| 154 |
+
return
|
| 155 |
+
|
| 156 |
+
batch_samples = rearrange(batch_samples, "... d -> (...) d")
|
| 157 |
+
self.replace_(batch_samples, mask=expired_codes)
|
| 158 |
+
flashy.distrib.broadcast_tensors(self.buffers())
|
| 159 |
+
|
| 160 |
+
def preprocess(self, x):
|
| 161 |
+
x = rearrange(x, "... d -> (...) d")
|
| 162 |
+
return x
|
| 163 |
+
|
| 164 |
+
def quantize(self, x):
|
| 165 |
+
embed = self.embed.t()
|
| 166 |
+
dist = -(
|
| 167 |
+
x.pow(2).sum(1, keepdim=True)
|
| 168 |
+
- 2 * x @ embed
|
| 169 |
+
+ embed.pow(2).sum(0, keepdim=True)
|
| 170 |
+
)
|
| 171 |
+
embed_ind = dist.max(dim=-1).indices
|
| 172 |
+
return embed_ind
|
| 173 |
+
|
| 174 |
+
def postprocess_emb(self, embed_ind, shape):
|
| 175 |
+
return embed_ind.view(*shape[:-1])
|
| 176 |
+
|
| 177 |
+
def dequantize(self, embed_ind):
|
| 178 |
+
quantize = F.embedding(embed_ind, self.embed)
|
| 179 |
+
return quantize
|
| 180 |
+
|
| 181 |
+
def encode(self, x):
|
| 182 |
+
shape = x.shape
|
| 183 |
+
# pre-process
|
| 184 |
+
x = self.preprocess(x)
|
| 185 |
+
# quantize
|
| 186 |
+
embed_ind = self.quantize(x)
|
| 187 |
+
# post-process
|
| 188 |
+
embed_ind = self.postprocess_emb(embed_ind, shape)
|
| 189 |
+
return embed_ind
|
| 190 |
+
|
| 191 |
+
def decode(self, embed_ind):
|
| 192 |
+
quantize = self.dequantize(embed_ind)
|
| 193 |
+
return quantize
|
| 194 |
+
|
| 195 |
+
def forward(self, x):
|
| 196 |
+
shape, dtype = x.shape, x.dtype
|
| 197 |
+
x = self.preprocess(x)
|
| 198 |
+
self.init_embed_(x)
|
| 199 |
+
|
| 200 |
+
embed_ind = self.quantize(x)
|
| 201 |
+
embed_onehot = F.one_hot(embed_ind, self.codebook_size).type(dtype)
|
| 202 |
+
embed_ind = self.postprocess_emb(embed_ind, shape)
|
| 203 |
+
quantize = self.dequantize(embed_ind)
|
| 204 |
+
|
| 205 |
+
if self.training:
|
| 206 |
+
# We do the expiry of code at that point as buffers are in sync
|
| 207 |
+
# and all the workers will take the same decision.
|
| 208 |
+
self.expire_codes_(x)
|
| 209 |
+
ema_inplace(self.cluster_size, embed_onehot.sum(0), self.decay)
|
| 210 |
+
embed_sum = x.t() @ embed_onehot
|
| 211 |
+
ema_inplace(self.embed_avg, embed_sum.t(), self.decay)
|
| 212 |
+
cluster_size = (
|
| 213 |
+
laplace_smoothing(self.cluster_size, self.codebook_size, self.epsilon)
|
| 214 |
+
* self.cluster_size.sum()
|
| 215 |
+
)
|
| 216 |
+
embed_normalized = self.embed_avg / cluster_size.unsqueeze(1)
|
| 217 |
+
self.embed.data.copy_(embed_normalized)
|
| 218 |
+
|
| 219 |
+
return quantize, embed_ind
|
| 220 |
+
|
| 221 |
+
|
| 222 |
+
class VectorQuantization(nn.Module):
|
| 223 |
+
"""Vector quantization implementation.
|
| 224 |
+
Currently supports only euclidean distance.
|
| 225 |
+
|
| 226 |
+
Args:
|
| 227 |
+
dim (int): Dimension
|
| 228 |
+
codebook_size (int): Codebook size
|
| 229 |
+
codebook_dim (int): Codebook dimension. If not defined, uses the specified dimension in dim.
|
| 230 |
+
decay (float): Decay for exponential moving average over the codebooks.
|
| 231 |
+
epsilon (float): Epsilon value for numerical stability.
|
| 232 |
+
kmeans_init (bool): Whether to use kmeans to initialize the codebooks.
|
| 233 |
+
kmeans_iters (int): Number of iterations used for kmeans initialization.
|
| 234 |
+
threshold_ema_dead_code (int):
|
| 235 |
+
channels_last (bool): Channels are the last dimension in the input tensors.
|
| 236 |
+
commitment_weight (float): Weight for commitment loss.
|
| 237 |
+
orthogonal_reg_weight (float): Orthogonal regularization weights.
|
| 238 |
+
orthogonal_reg_active_codes_only (bool): Apply orthogonal regularization only on active codes.
|
| 239 |
+
orthogonal_reg_max_codes (optional int): Maximum number of codes to consider
|
| 240 |
+
for orthogonal regularization.
|
| 241 |
+
threshold_ema_dead_code (int): Threshold for dead code expiration. Replace any codes
|
| 242 |
+
that have an exponential moving average cluster size less than the specified threshold with
|
| 243 |
+
randomly selected vector from the current batch.
|
| 244 |
+
"""
|
| 245 |
+
def __init__(
|
| 246 |
+
self,
|
| 247 |
+
dim: int,
|
| 248 |
+
codebook_size: int,
|
| 249 |
+
codebook_dim: tp.Optional[int] = None,
|
| 250 |
+
decay: float = 0.8,
|
| 251 |
+
epsilon: float = 1e-5,
|
| 252 |
+
kmeans_init: bool = False,
|
| 253 |
+
kmeans_iters: int = 10,
|
| 254 |
+
threshold_ema_dead_code: int = 2,
|
| 255 |
+
channels_last: bool = False,
|
| 256 |
+
commitment_weight: float = 1.,
|
| 257 |
+
orthogonal_reg_weight: float = 0.0,
|
| 258 |
+
orthogonal_reg_active_codes_only: bool = False,
|
| 259 |
+
orthogonal_reg_max_codes: tp.Optional[int] = None,
|
| 260 |
+
):
|
| 261 |
+
super().__init__()
|
| 262 |
+
_codebook_dim: int = default(codebook_dim, dim)
|
| 263 |
+
|
| 264 |
+
requires_projection = _codebook_dim != dim
|
| 265 |
+
self.project_in = (nn.Linear(dim, _codebook_dim) if requires_projection else nn.Identity())
|
| 266 |
+
self.project_out = (nn.Linear(_codebook_dim, dim) if requires_projection else nn.Identity())
|
| 267 |
+
|
| 268 |
+
self.epsilon = epsilon
|
| 269 |
+
self.commitment_weight = commitment_weight
|
| 270 |
+
|
| 271 |
+
self.orthogonal_reg_weight = orthogonal_reg_weight
|
| 272 |
+
self.orthogonal_reg_active_codes_only = orthogonal_reg_active_codes_only
|
| 273 |
+
self.orthogonal_reg_max_codes = orthogonal_reg_max_codes
|
| 274 |
+
|
| 275 |
+
self._codebook = EuclideanCodebook(dim=_codebook_dim, codebook_size=codebook_size,
|
| 276 |
+
kmeans_init=kmeans_init, kmeans_iters=kmeans_iters,
|
| 277 |
+
decay=decay, epsilon=epsilon,
|
| 278 |
+
threshold_ema_dead_code=threshold_ema_dead_code)
|
| 279 |
+
self.codebook_size = codebook_size
|
| 280 |
+
|
| 281 |
+
self.channels_last = channels_last
|
| 282 |
+
|
| 283 |
+
@property
|
| 284 |
+
def codebook(self):
|
| 285 |
+
return self._codebook.embed
|
| 286 |
+
|
| 287 |
+
@property
|
| 288 |
+
def inited(self):
|
| 289 |
+
return self._codebook.inited
|
| 290 |
+
|
| 291 |
+
def _preprocess(self, x):
|
| 292 |
+
if not self.channels_last:
|
| 293 |
+
x = rearrange(x, "b d n -> b n d")
|
| 294 |
+
return x
|
| 295 |
+
|
| 296 |
+
def _postprocess(self, quantize):
|
| 297 |
+
if not self.channels_last:
|
| 298 |
+
quantize = rearrange(quantize, "b n d -> b d n")
|
| 299 |
+
return quantize
|
| 300 |
+
|
| 301 |
+
def encode(self, x):
|
| 302 |
+
x = self._preprocess(x)
|
| 303 |
+
x = self.project_in(x)
|
| 304 |
+
embed_in = self._codebook.encode(x)
|
| 305 |
+
return embed_in
|
| 306 |
+
|
| 307 |
+
def decode(self, embed_ind):
|
| 308 |
+
quantize = self._codebook.decode(embed_ind)
|
| 309 |
+
quantize = self.project_out(quantize)
|
| 310 |
+
quantize = self._postprocess(quantize)
|
| 311 |
+
return quantize
|
| 312 |
+
|
| 313 |
+
def forward(self, x):
|
| 314 |
+
device = x.device
|
| 315 |
+
x = self._preprocess(x)
|
| 316 |
+
|
| 317 |
+
x = self.project_in(x)
|
| 318 |
+
quantize, embed_ind = self._codebook(x)
|
| 319 |
+
|
| 320 |
+
if self.training:
|
| 321 |
+
quantize = x + (quantize - x).detach()
|
| 322 |
+
|
| 323 |
+
loss = torch.tensor([0.0], device=device, requires_grad=self.training)
|
| 324 |
+
|
| 325 |
+
if self.training:
|
| 326 |
+
if self.commitment_weight > 0:
|
| 327 |
+
commit_loss = F.mse_loss(quantize.detach(), x)
|
| 328 |
+
loss = loss + commit_loss * self.commitment_weight
|
| 329 |
+
|
| 330 |
+
if self.orthogonal_reg_weight > 0:
|
| 331 |
+
codebook = self.codebook
|
| 332 |
+
|
| 333 |
+
if self.orthogonal_reg_active_codes_only:
|
| 334 |
+
# only calculate orthogonal loss for the activated codes for this batch
|
| 335 |
+
unique_code_ids = torch.unique(embed_ind)
|
| 336 |
+
codebook = codebook[unique_code_ids]
|
| 337 |
+
|
| 338 |
+
num_codes = codebook.shape[0]
|
| 339 |
+
if exists(self.orthogonal_reg_max_codes) and num_codes > self.orthogonal_reg_max_codes:
|
| 340 |
+
rand_ids = torch.randperm(num_codes, device=device)[:self.orthogonal_reg_max_codes]
|
| 341 |
+
codebook = codebook[rand_ids]
|
| 342 |
+
|
| 343 |
+
orthogonal_reg_loss = orthogonal_loss_fn(codebook)
|
| 344 |
+
loss = loss + orthogonal_reg_loss * self.orthogonal_reg_weight
|
| 345 |
+
|
| 346 |
+
quantize = self.project_out(quantize)
|
| 347 |
+
quantize = self._postprocess(quantize)
|
| 348 |
+
|
| 349 |
+
return quantize, embed_ind, loss
|
| 350 |
+
|
| 351 |
+
|
| 352 |
+
class ResidualVectorQuantization(nn.Module):
|
| 353 |
+
"""Residual vector quantization implementation.
|
| 354 |
+
|
| 355 |
+
Follows Algorithm 1. in https://arxiv.org/pdf/2107.03312.pdf
|
| 356 |
+
"""
|
| 357 |
+
def __init__(self, *, num_quantizers, **kwargs):
|
| 358 |
+
super().__init__()
|
| 359 |
+
self.layers = nn.ModuleList(
|
| 360 |
+
[VectorQuantization(**kwargs) for _ in range(num_quantizers)]
|
| 361 |
+
)
|
| 362 |
+
|
| 363 |
+
def forward(self, x, n_q: tp.Optional[int] = None):
|
| 364 |
+
quantized_out = 0.0
|
| 365 |
+
residual = x
|
| 366 |
+
|
| 367 |
+
all_losses = []
|
| 368 |
+
all_indices = []
|
| 369 |
+
|
| 370 |
+
n_q = n_q or len(self.layers)
|
| 371 |
+
|
| 372 |
+
for i, layer in enumerate(self.layers[:n_q]):
|
| 373 |
+
quantized, indices, loss = layer(residual)
|
| 374 |
+
quantized = quantized.detach()
|
| 375 |
+
residual = residual - quantized
|
| 376 |
+
quantized_out = quantized_out + quantized
|
| 377 |
+
all_indices.append(indices)
|
| 378 |
+
all_losses.append(loss)
|
| 379 |
+
|
| 380 |
+
if self.training:
|
| 381 |
+
# Solving subtle bug with STE and RVQ: https://github.com/facebookresearch/encodec/issues/25
|
| 382 |
+
quantized_out = x + (quantized_out - x).detach()
|
| 383 |
+
|
| 384 |
+
out_losses, out_indices = map(torch.stack, (all_losses, all_indices))
|
| 385 |
+
return quantized_out, out_indices, out_losses
|
| 386 |
+
|
| 387 |
+
def encode(self, x: torch.Tensor, n_q: tp.Optional[int] = None) -> torch.Tensor:
|
| 388 |
+
residual = x
|
| 389 |
+
all_indices = []
|
| 390 |
+
n_q = n_q or len(self.layers)
|
| 391 |
+
for layer in self.layers[:n_q]:
|
| 392 |
+
indices = layer.encode(residual)
|
| 393 |
+
quantized = layer.decode(indices)
|
| 394 |
+
residual = residual - quantized
|
| 395 |
+
all_indices.append(indices)
|
| 396 |
+
out_indices = torch.stack(all_indices)
|
| 397 |
+
return out_indices
|
| 398 |
+
|
| 399 |
+
def decode(self, q_indices: torch.Tensor) -> torch.Tensor:
|
| 400 |
+
quantized_out = torch.tensor(0.0, device=q_indices.device)
|
| 401 |
+
for i, indices in enumerate(q_indices):
|
| 402 |
+
layer = self.layers[i]
|
| 403 |
+
quantized = layer.decode(indices)
|
| 404 |
+
quantized_out = quantized_out + quantized
|
| 405 |
+
return quantized_out
|
audiocraft/quantization/vq.py
ADDED
|
@@ -0,0 +1,115 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
| 2 |
+
# All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# This source code is licensed under the license found in the
|
| 5 |
+
# LICENSE file in the root directory of this source tree.
|
| 6 |
+
|
| 7 |
+
import math
|
| 8 |
+
import typing as tp
|
| 9 |
+
|
| 10 |
+
import torch
|
| 11 |
+
|
| 12 |
+
from .base import BaseQuantizer, QuantizedResult
|
| 13 |
+
from .core_vq import ResidualVectorQuantization
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
class ResidualVectorQuantizer(BaseQuantizer):
|
| 17 |
+
"""Residual Vector Quantizer.
|
| 18 |
+
|
| 19 |
+
Args:
|
| 20 |
+
dimension (int): Dimension of the codebooks.
|
| 21 |
+
n_q (int): Number of residual vector quantizers used.
|
| 22 |
+
q_dropout (bool): Random quantizer drop out at train time.
|
| 23 |
+
bins (int): Codebook size.
|
| 24 |
+
decay (float): Decay for exponential moving average over the codebooks.
|
| 25 |
+
kmeans_init (bool): Whether to use kmeans to initialize the codebooks.
|
| 26 |
+
kmeans_iters (int): Number of iterations used for kmeans initialization.
|
| 27 |
+
threshold_ema_dead_code (int): Threshold for dead code expiration. Replace any codes
|
| 28 |
+
that have an exponential moving average cluster size less than the specified threshold with
|
| 29 |
+
randomly selected vector from the current batch.
|
| 30 |
+
orthogonal_reg_weight (float): Orthogonal regularization weights.
|
| 31 |
+
orthogonal_reg_active_codes_only (bool): Apply orthogonal regularization only on active codes.
|
| 32 |
+
orthogonal_reg_max_codes (optional int): Maximum number of codes to consider.
|
| 33 |
+
for orthogonal regularization.
|
| 34 |
+
"""
|
| 35 |
+
def __init__(
|
| 36 |
+
self,
|
| 37 |
+
dimension: int = 256,
|
| 38 |
+
n_q: int = 8,
|
| 39 |
+
q_dropout: bool = False,
|
| 40 |
+
bins: int = 1024,
|
| 41 |
+
decay: float = 0.99,
|
| 42 |
+
kmeans_init: bool = True,
|
| 43 |
+
kmeans_iters: int = 10,
|
| 44 |
+
threshold_ema_dead_code: int = 2,
|
| 45 |
+
orthogonal_reg_weight: float = 0.0,
|
| 46 |
+
orthogonal_reg_active_codes_only: bool = False,
|
| 47 |
+
orthogonal_reg_max_codes: tp.Optional[int] = None,
|
| 48 |
+
):
|
| 49 |
+
super().__init__()
|
| 50 |
+
self.max_n_q = n_q
|
| 51 |
+
self.n_q = n_q
|
| 52 |
+
self.q_dropout = q_dropout
|
| 53 |
+
self.dimension = dimension
|
| 54 |
+
self.bins = bins
|
| 55 |
+
self.decay = decay
|
| 56 |
+
self.kmeans_init = kmeans_init
|
| 57 |
+
self.kmeans_iters = kmeans_iters
|
| 58 |
+
self.threshold_ema_dead_code = threshold_ema_dead_code
|
| 59 |
+
self.orthogonal_reg_weight = orthogonal_reg_weight
|
| 60 |
+
self.orthogonal_reg_active_codes_only = orthogonal_reg_active_codes_only
|
| 61 |
+
self.orthogonal_reg_max_codes = orthogonal_reg_max_codes
|
| 62 |
+
self.vq = ResidualVectorQuantization(
|
| 63 |
+
dim=self.dimension,
|
| 64 |
+
codebook_size=self.bins,
|
| 65 |
+
num_quantizers=self.n_q,
|
| 66 |
+
decay=self.decay,
|
| 67 |
+
kmeans_init=self.kmeans_init,
|
| 68 |
+
kmeans_iters=self.kmeans_iters,
|
| 69 |
+
threshold_ema_dead_code=self.threshold_ema_dead_code,
|
| 70 |
+
orthogonal_reg_weight=self.orthogonal_reg_weight,
|
| 71 |
+
orthogonal_reg_active_codes_only=self.orthogonal_reg_active_codes_only,
|
| 72 |
+
orthogonal_reg_max_codes=self.orthogonal_reg_max_codes,
|
| 73 |
+
channels_last=False
|
| 74 |
+
)
|
| 75 |
+
|
| 76 |
+
def forward(self, x: torch.Tensor, frame_rate: int):
|
| 77 |
+
n_q = self.n_q
|
| 78 |
+
if self.training and self.q_dropout:
|
| 79 |
+
n_q = int(torch.randint(1, self.n_q + 1, (1,)).item())
|
| 80 |
+
bw_per_q = math.log2(self.bins) * frame_rate / 1000
|
| 81 |
+
quantized, codes, commit_loss = self.vq(x, n_q=n_q)
|
| 82 |
+
codes = codes.transpose(0, 1)
|
| 83 |
+
# codes is [B, K, T], with T frames, K nb of codebooks.
|
| 84 |
+
bw = torch.tensor(n_q * bw_per_q).to(x)
|
| 85 |
+
return QuantizedResult(quantized, codes, bw, penalty=torch.mean(commit_loss))
|
| 86 |
+
|
| 87 |
+
def encode(self, x: torch.Tensor) -> torch.Tensor:
|
| 88 |
+
"""Encode a given input tensor with the specified frame rate at the given bandwidth.
|
| 89 |
+
The RVQ encode method sets the appropriate number of quantizer to use
|
| 90 |
+
and returns indices for each quantizer.
|
| 91 |
+
"""
|
| 92 |
+
n_q = self.n_q
|
| 93 |
+
codes = self.vq.encode(x, n_q=n_q)
|
| 94 |
+
codes = codes.transpose(0, 1)
|
| 95 |
+
# codes is [B, K, T], with T frames, K nb of codebooks.
|
| 96 |
+
return codes
|
| 97 |
+
|
| 98 |
+
def decode(self, codes: torch.Tensor) -> torch.Tensor:
|
| 99 |
+
"""Decode the given codes to the quantized representation."""
|
| 100 |
+
# codes is [B, K, T], with T frames, K nb of codebooks, vq.decode expects [K, B, T].
|
| 101 |
+
codes = codes.transpose(0, 1)
|
| 102 |
+
quantized = self.vq.decode(codes)
|
| 103 |
+
return quantized
|
| 104 |
+
|
| 105 |
+
@property
|
| 106 |
+
def total_codebooks(self):
|
| 107 |
+
return self.max_n_q
|
| 108 |
+
|
| 109 |
+
@property
|
| 110 |
+
def num_codebooks(self):
|
| 111 |
+
return self.n_q
|
| 112 |
+
|
| 113 |
+
def set_num_codebooks(self, n: int):
|
| 114 |
+
assert n > 0 and n <= self.max_n_q
|
| 115 |
+
self.n_q = n
|