Delete src/miocodec/module/fsq.py with huggingface_hub
Browse files- src/miocodec/module/fsq.py +0 -142
src/miocodec/module/fsq.py
DELETED
|
@@ -1,142 +0,0 @@
|
|
| 1 |
-
# Finite Scalar Quantization: https://arxiv.org/abs/2309.15505
|
| 2 |
-
|
| 3 |
-
import torch
|
| 4 |
-
from torch import nn
|
| 5 |
-
|
| 6 |
-
from ..util import get_logger
|
| 7 |
-
|
| 8 |
-
logger = get_logger()
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
def round_ste(z: torch.Tensor) -> torch.Tensor:
|
| 12 |
-
"""Round with straight through gradients."""
|
| 13 |
-
zhat = z.round()
|
| 14 |
-
return z + (zhat - z).detach()
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
def get_entropy(prob: torch.Tensor, eps: float = 1e-10) -> torch.Tensor:
|
| 18 |
-
return -torch.sum(prob * torch.log(prob + eps), dim=-1)
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
class FSQ(nn.Module):
|
| 22 |
-
def __init__(self, levels: list[int]):
|
| 23 |
-
super().__init__()
|
| 24 |
-
self.levels = levels
|
| 25 |
-
self.dim = len(levels)
|
| 26 |
-
|
| 27 |
-
_levels = torch.tensor(levels, dtype=torch.long)
|
| 28 |
-
self.register_buffer("_levels", _levels, persistent=False)
|
| 29 |
-
_basis = torch.cumprod(torch.tensor([1] + levels[:-1]), dim=0, dtype=torch.long)
|
| 30 |
-
self.register_buffer("_basis", _basis, persistent=False)
|
| 31 |
-
|
| 32 |
-
def bound(self, z: torch.Tensor, eps: float = 1e-3) -> torch.Tensor:
|
| 33 |
-
"""Bound `z`, an array of shape (..., d)."""
|
| 34 |
-
half_l = (self._levels - 1) * (1 - eps) / 2
|
| 35 |
-
offset = torch.where(self._levels % 2 == 0, 0.5, 0.0)
|
| 36 |
-
shift = (offset / half_l).tan()
|
| 37 |
-
return (z + shift).tanh() * half_l - offset
|
| 38 |
-
|
| 39 |
-
def quantize(self, z: torch.Tensor) -> torch.Tensor:
|
| 40 |
-
"""Quantizes z, returns quantized zhat, same shape as z."""
|
| 41 |
-
quantized = round_ste(self.bound(z))
|
| 42 |
-
half_width = self._levels // 2 # Renormalize to [-1, 1].
|
| 43 |
-
return quantized / half_width
|
| 44 |
-
|
| 45 |
-
def _scale_and_shift(self, zhat_normalized: torch.Tensor) -> torch.Tensor:
|
| 46 |
-
half_width = self._levels // 2
|
| 47 |
-
return (zhat_normalized * half_width) + half_width
|
| 48 |
-
|
| 49 |
-
def _scale_and_shift_inverse(self, zhat: torch.Tensor) -> torch.Tensor:
|
| 50 |
-
half_width = self._levels // 2
|
| 51 |
-
return (zhat - half_width) / half_width
|
| 52 |
-
|
| 53 |
-
def codes_to_indices(self, zhat: torch.Tensor) -> torch.Tensor:
|
| 54 |
-
"""Converts a `code` to an index in the codebook."""
|
| 55 |
-
# (B, T, C) -> (B, T)
|
| 56 |
-
assert zhat.shape[-1] == len(self.levels)
|
| 57 |
-
zhat = self._scale_and_shift(zhat)
|
| 58 |
-
# MPS doesn't support float64, fall back to float32
|
| 59 |
-
dtype = torch.float32 if zhat.device.type == "mps" else torch.float64
|
| 60 |
-
return (zhat * self._basis.to(dtype)).to(torch.long).sum(dim=-1)
|
| 61 |
-
|
| 62 |
-
def indices_to_codes(self, indices: torch.Tensor) -> torch.Tensor:
|
| 63 |
-
"""Inverse of `codes_to_indices`."""
|
| 64 |
-
# (B, T) -> (B, T, C)
|
| 65 |
-
indices = indices.unsqueeze(-1)
|
| 66 |
-
codes_non_centered = (indices // self._basis) % self._levels
|
| 67 |
-
return self._scale_and_shift_inverse(codes_non_centered)
|
| 68 |
-
|
| 69 |
-
def encode(self, z: torch.Tensor) -> torch.Tensor:
|
| 70 |
-
z_q = self.quantize(z)
|
| 71 |
-
indices = self.codes_to_indices(z_q) # (B, T)
|
| 72 |
-
return z_q, indices
|
| 73 |
-
|
| 74 |
-
def decode(self, indices: torch.Tensor) -> torch.Tensor:
|
| 75 |
-
z_q = self.indices_to_codes(indices) # (B, T, C)
|
| 76 |
-
return z_q
|
| 77 |
-
|
| 78 |
-
def forward(self, z: torch.Tensor):
|
| 79 |
-
z_q = self.quantize(z)
|
| 80 |
-
indices = self.codes_to_indices(z_q) # (B, T)
|
| 81 |
-
return z_q, indices
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
class FiniteScalarQuantizer(nn.Module):
|
| 85 |
-
def __init__(self, input_dim: int, output_dim: int, levels: list[int]) -> None:
|
| 86 |
-
super().__init__()
|
| 87 |
-
self.input_dim_ = input_dim
|
| 88 |
-
self.output_dim_ = output_dim
|
| 89 |
-
|
| 90 |
-
self.fsq = FSQ(levels)
|
| 91 |
-
logger.debug(
|
| 92 |
-
f"Finite Scalar Quantizer with levels: {levels}, input_dim: {input_dim}, output_dim: {output_dim}, codebook_size: {self.all_codebook_size}"
|
| 93 |
-
)
|
| 94 |
-
|
| 95 |
-
self.proj_in = nn.Linear(input_dim, len(levels)) if len(levels) != input_dim else nn.Identity()
|
| 96 |
-
self.proj_out = nn.Linear(len(levels), output_dim) if len(levels) != output_dim else nn.Identity()
|
| 97 |
-
|
| 98 |
-
def build_codebook(self) -> None:
|
| 99 |
-
pass
|
| 100 |
-
|
| 101 |
-
@property
|
| 102 |
-
def output_dim(self) -> int:
|
| 103 |
-
return self.output_dim_
|
| 104 |
-
|
| 105 |
-
@property
|
| 106 |
-
def all_codebook_size(self) -> int:
|
| 107 |
-
size = 1
|
| 108 |
-
for level in self.fsq.levels:
|
| 109 |
-
size *= level
|
| 110 |
-
return size
|
| 111 |
-
|
| 112 |
-
def forward(self, z: torch.Tensor) -> tuple[torch.Tensor, dict]:
|
| 113 |
-
latent = self.proj_in(z) # Latent projected by proj_in
|
| 114 |
-
quantized_latent, indices = self.fsq(latent) # Quantized latent before proj_out
|
| 115 |
-
z_q = self.proj_out(quantized_latent)
|
| 116 |
-
|
| 117 |
-
# Compute perplexity from used indices distribution
|
| 118 |
-
flat_indices = indices.view(-1)
|
| 119 |
-
unique_indices, counts = torch.unique(flat_indices, return_counts=True)
|
| 120 |
-
used_indices_probs = counts.float() / flat_indices.numel()
|
| 121 |
-
entropy = get_entropy(used_indices_probs)
|
| 122 |
-
perplexity = torch.exp(entropy)
|
| 123 |
-
|
| 124 |
-
info_dict = {
|
| 125 |
-
"latent": latent,
|
| 126 |
-
"quantized_latent": quantized_latent,
|
| 127 |
-
"indices": indices,
|
| 128 |
-
"perplexity": perplexity,
|
| 129 |
-
}
|
| 130 |
-
return z_q, info_dict
|
| 131 |
-
|
| 132 |
-
def encode(self, z: torch.Tensor, skip_proj: bool = False) -> tuple[torch.Tensor, torch.Tensor]:
|
| 133 |
-
z = self.proj_in(z)
|
| 134 |
-
z_q, indices = self.fsq.encode(z)
|
| 135 |
-
if not skip_proj:
|
| 136 |
-
z_q = self.proj_out(z_q)
|
| 137 |
-
return z_q, indices
|
| 138 |
-
|
| 139 |
-
def decode(self, indices: torch.Tensor) -> torch.Tensor:
|
| 140 |
-
z_q = self.fsq.decode(indices)
|
| 141 |
-
z_q = self.proj_out(z_q)
|
| 142 |
-
return z_q
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|