Plachta's picture
Upload 116 files
56a1295 verified
"""
Lookup Free Quantization
Proposed in https://arxiv.org/abs/2310.05737
In the simplest setup, each dimension is quantized into {-1, 1}.
An entropy penalty is used to encourage utilization.
"""
from math import log2, ceil
from functools import partial, cache
from collections import namedtuple
from contextlib import nullcontext
import torch.distributed as dist
from torch.distributed import nn as dist_nn
import torch
from torch import nn, einsum
import torch.nn.functional as F
from torch.nn import Module
from torch.amp import autocast
from einops import rearrange, reduce, pack, unpack
# constants
Return = namedtuple('Return', ['quantized', 'indices', 'entropy_aux_loss'])
LossBreakdown = namedtuple('LossBreakdown', ['per_sample_entropy', 'batch_entropy', 'commitment'])
# distributed helpers
@cache
def is_distributed():
return dist.is_initialized() and dist.get_world_size() > 1
def maybe_distributed_mean(t):
if not is_distributed():
return t
dist_nn.all_reduce(t)
t = t / dist.get_world_size()
return t
# helper functions
def exists(v):
return v is not None
def identity(t):
return t
def default(*args):
for arg in args:
if exists(arg):
return arg() if callable(arg) else arg
return None
def pack_one(t, pattern):
return pack([t], pattern)
def unpack_one(t, ps, pattern):
return unpack(t, ps, pattern)[0]
def l2norm(t):
return F.normalize(t, dim = -1)
# entropy
def log(t, eps = 1e-5):
return t.clamp(min = eps).log()
def entropy(prob):
return (-prob * log(prob)).sum(dim=-1)
# cosine sim linear
class CosineSimLinear(Module):
def __init__(
self,
dim_in,
dim_out,
scale = 1.
):
super().__init__()
self.scale = scale
self.weight = nn.Parameter(torch.randn(dim_in, dim_out))
def forward(self, x):
x = F.normalize(x, dim = -1)
w = F.normalize(self.weight, dim = 0)
return (x @ w) * self.scale
def soft_entropy_loss(u, tau=1.0, gamma=1.0):
"""
Compute the soft entropy loss for Binary Spherical Quantization (BSQ).
Args:
u (torch.Tensor): Input latent embeddings of shape (batch_size, L).
tau (float): Temperature scaling factor.
gamma (float): Weight for the second entropy term.
Returns:
torch.Tensor: Soft entropy loss.
"""
# Binary quantization: Generate implicit codebook corners
L = u.size(1) # Dimensionality of codebook
corners = torch.tensor([-1.0, 1.0], device=u.device) / (L**0.5)
# Compute soft quantization probabilities for all dimensions
# q_hat(c|u) for each dimension
prob_matrix = torch.sigmoid(2 * tau * corners.unsqueeze(1) * u.unsqueeze(2)) # Shape: (batch_size, L, 2)
# Entropy of q_hat(c|u) (independent along each dimension)
entropy_per_dim = -torch.sum(prob_matrix * prob_matrix.log(), dim=-1) # Shape: (batch_size, L)
entropy_term1 = entropy_per_dim.mean()
# Expected probabilities for dataset entropy (approximation)
expected_probs = prob_matrix.mean(dim=0) # Mean across batch, shape: (L, 2)
entropy_term2 = -torch.sum(expected_probs * expected_probs.log(), dim=-1).mean()
# Final entropy loss
loss = entropy_term1 - gamma * entropy_term2
return loss
# class
class BinarySphericalQuantize(Module):
def __init__(
self,
*,
dim = None,
codebook_size = None,
entropy_loss_weight = 0.1,
commitment_loss_weight = 0.,
diversity_gamma = 1.,
straight_through_activation = nn.Identity(),
num_codebooks = 1,
keep_num_codebooks_dim = None,
codebook_scale = 1., # for residual LFQ, codebook scaled down by 2x at each layer
frac_per_sample_entropy = 0.25, # make less than 1. to only use a random fraction of the probs for per sample entropy
has_projections = None,
projection_has_bias = True,
soft_clamp_input_value = None,
cosine_sim_project_in = False,
cosine_sim_project_in_scale = None,
channel_first = None,
experimental_softplus_entropy_loss = False,
entropy_loss_offset = 5., # how much to shift the loss before softplus
spherical = True, # from https://arxiv.org/abs/2406.07548
force_quantization_f32 = True, # will force the quantization step to be full precision
enable_entropy_loss = True,
soft_entropy_loss = True,
):
super().__init__()
# some assert validations
assert exists(dim) or exists(codebook_size), 'either dim or codebook_size must be specified for LFQ'
assert not exists(codebook_size) or log2(codebook_size).is_integer(), f'your codebook size must be a power of 2 for lookup free quantization (suggested {2 ** ceil(log2(codebook_size))})'
codebook_size = default(codebook_size, lambda: 2 ** dim)
self.codebook_size = codebook_size
codebook_dim = int(log2(codebook_size))
codebook_dims = codebook_dim * num_codebooks
dim = default(dim, codebook_dims)
has_projections = default(has_projections, dim != codebook_dims)
if cosine_sim_project_in:
cosine_sim_project_in = default(cosine_sim_project_in_scale, codebook_scale)
project_in_klass = partial(CosineSimLinear, scale = cosine_sim_project_in)
else:
project_in_klass = partial(nn.Linear, bias = projection_has_bias)
self.project_in = project_in_klass(dim, codebook_dims) if has_projections else nn.Identity()
self.project_out = nn.Linear(codebook_dims, dim, bias = projection_has_bias) if has_projections else nn.Identity()
self.has_projections = has_projections
self.dim = dim
self.codebook_dim = codebook_dim
self.num_codebooks = num_codebooks
keep_num_codebooks_dim = default(keep_num_codebooks_dim, num_codebooks > 1)
assert not (num_codebooks > 1 and not keep_num_codebooks_dim)
self.keep_num_codebooks_dim = keep_num_codebooks_dim
# channel first
self.channel_first = channel_first
# straight through activation
self.activation = straight_through_activation
# whether to use BSQ (binary spherical quantization)
self.spherical = spherical
self.maybe_l2norm = (lambda t: l2norm(t) * self.codebook_scale) if spherical else identity
# entropy aux loss related weights
assert 0 < frac_per_sample_entropy <= 1.
self.frac_per_sample_entropy = frac_per_sample_entropy
self.diversity_gamma = diversity_gamma
self.entropy_loss_weight = entropy_loss_weight
# codebook scale
self.codebook_scale = codebook_scale
# commitment loss
self.commitment_loss_weight = commitment_loss_weight
# whether to soft clamp the input value from -value to value
self.soft_clamp_input_value = soft_clamp_input_value
assert not exists(soft_clamp_input_value) or soft_clamp_input_value >= codebook_scale
# whether to make the entropy loss positive through a softplus (experimental, please report if this worked or not in discussions)
self.entropy_loss_offset = entropy_loss_offset
self.experimental_softplus_entropy_loss = experimental_softplus_entropy_loss
# for no auxiliary loss, during inference
self.register_buffer('mask', 2 ** torch.arange(codebook_dim - 1, -1, -1))
self.register_buffer('zero', torch.tensor(0.), persistent = False)
# whether to force quantization step to be f32
self.force_quantization_f32 = force_quantization_f32
# codes
self.enable_entropy_loss = enable_entropy_loss
self.soft_entropy_loss = soft_entropy_loss
if codebook_size <= 100000:
all_codes = torch.arange(codebook_size)
bits = ((all_codes[..., None].int() & self.mask) != 0).float()
codebook = self.bits_to_codes(bits)
self.register_buffer('codebook', codebook.float(), persistent = False)
else:
all_codes = torch.arange(pow(2, 16))
mask = 2 ** torch.arange(16 - 1, -1, -1)
bits = ((all_codes[..., None].int() & mask) != 0).float()
codebook = self.bits_to_codes(bits)
self.register_buffer('codebook', codebook.float(), persistent = False)
def bits_to_codes(self, bits):
return bits * self.codebook_scale * 2 - self.codebook_scale
@property
def dtype(self):
return self.codebook.dtype
def indices_to_codes(
self,
indices,
project_out = True
):
is_img_or_video = indices.ndim >= (3 + int(self.keep_num_codebooks_dim))
should_transpose = default(self.channel_first, is_img_or_video)
if not self.keep_num_codebooks_dim:
indices = rearrange(indices, '... -> ... 1')
# indices to codes, which are bits of either -1 or 1
bits = ((indices[..., None].int() & self.mask) != 0).to(self.dtype)
codes = self.bits_to_codes(bits)
codes = self.maybe_l2norm(codes)
codes = rearrange(codes, '... c d -> ... (c d)')
# whether to project codes out to original dimensions
# if the input feature dimensions were not log2(codebook size)
if project_out:
codes = self.project_out(codes)
# rearrange codes back to original shape
if should_transpose:
codes = rearrange(codes, 'b ... d -> b d ...')
return codes
def bits_to_z(self, bits):
# assert bits must contain only -1 and 1
assert torch.all(bits.abs() == 1)
quantized = bits.float()
quantized = self.maybe_l2norm(quantized)
z = self.project_out(quantized)
return z
def forward(
self,
x,
inv_temperature = 100.,
return_loss_breakdown = False,
mask = None,
return_bits = False
):
"""
einstein notation
b - batch
n - sequence (or flattened spatial dimensions)
d - feature dimension, which is also log2(codebook size)
c - number of codebook dim
"""
is_img_or_video = x.ndim >= 4
should_transpose = default(self.channel_first, is_img_or_video)
# standardize image or video into (batch, seq, dimension)
if should_transpose:
x = rearrange(x, 'b d ... -> b ... d')
x, ps = pack_one(x, 'b * d')
assert x.shape[-1] == self.dim, f'expected dimension of {self.dim} but received {x.shape[-1]}'
x = self.project_in(x)
# maybe soft clamp
if exists(self.soft_clamp_input_value):
clamp_value = self.soft_clamp_input_value
x = (x / clamp_value).tanh() * clamp_value
# split out number of codebooks
x = rearrange(x, 'b n (c d) -> b n c d', c = self.num_codebooks)
# maybe l2norm
x = self.maybe_l2norm(x)
# whether to force quantization step to be full precision or not
force_f32 = self.force_quantization_f32
quantization_context = partial(autocast, 'cuda', enabled = False) if force_f32 else nullcontext
with quantization_context():
if force_f32:
orig_dtype = x.dtype
x = x.float()
# quantize by eq 3.
original_input = x
codebook_value = torch.ones_like(x) * self.codebook_scale
quantized = torch.where(x > 0, codebook_value, -codebook_value)
if return_bits:
return quantized
# calculate indices
indices = reduce((quantized > 0).int() * self.mask.int(), 'b n c d -> b n c', 'sum')
# maybe l2norm
quantized = self.maybe_l2norm(quantized)
# use straight-through gradients (optionally with custom activation fn) if training
if self.training:
x = self.activation(x)
x = x + (quantized - x).detach()
else:
x = quantized
# entropy aux loss
if self.soft_entropy_loss:
entropy_aux_loss = soft_entropy_loss(x, tau=1.0, gamma=1.0)
elif self.training and self.enable_entropy_loss:
if force_f32:
codebook = self.codebook.float()
codebook = self.maybe_l2norm(codebook)
# whether to only use a fraction of probs, for reducing memory
if self.frac_per_sample_entropy < 1.:
# account for mask
if exists(mask):
original_input = original_input[mask]
original_input = rearrange(original_input, 'b n ... -> (b n) ...')
rand_mask = torch.randn(self.codebook_dim).argsort(dim = -1) < 16
sampled_input = original_input[..., rand_mask]
sampled_distance = -2 * einsum('... i d, j d -> ... i j', sampled_input, codebook)
sampled_prob = (-sampled_distance * inv_temperature).softmax(dim = -1)
per_sample_probs = sampled_prob
else:
if exists(mask):
original_input = original_input[mask]
original_input = rearrange(original_input, 'b n ... -> (b n) ...')
# the same as euclidean distance up to a constant
distance = -2 * einsum('... i d, j d -> ... i j', original_input, codebook)
prob = (-distance * inv_temperature).softmax(dim = -1)
per_sample_probs = prob
# calculate per sample entropy
per_sample_entropy = entropy(per_sample_probs).mean()
# distribution over all available tokens in the batch
avg_prob = reduce(per_sample_probs, '... c d -> c d', 'mean')
avg_prob = maybe_distributed_mean(avg_prob)
codebook_entropy = entropy(avg_prob).mean()
# 1. entropy will be nudged to be low for each code, to encourage the network to output confident predictions
# 2. codebook entropy will be nudged to be high, to encourage all codes to be uniformly used within the batch
entropy_aux_loss = per_sample_entropy - self.diversity_gamma * codebook_entropy
else:
# if not training, just return dummy 0
entropy_aux_loss = per_sample_entropy = codebook_entropy = self.zero
# whether to make the entropy loss positive or not through a (shifted) softplus
if self.training and self.experimental_softplus_entropy_loss:
entropy_aux_loss = F.softplus(entropy_aux_loss + self.entropy_loss_offset)
# commit loss
if self.training and self.commitment_loss_weight > 0.:
commit_loss = F.mse_loss(original_input, quantized.detach(), reduction = 'none')
if exists(mask):
commit_loss = commit_loss[mask]
commit_loss = commit_loss.mean()
else:
commit_loss = self.zero
# input back to original dtype if needed
if force_f32:
x = x.type(orig_dtype)
# merge back codebook dim
x = rearrange(x, 'b n c d -> b n (c d)')
# project out to feature dimension if needed
x = self.project_out(x)
# reconstitute image or video dimensions
if should_transpose:
x = unpack_one(x, ps, 'b * d')
x = rearrange(x, 'b ... d -> b d ...')
indices = unpack_one(indices, ps, 'b * c')
# whether to remove single codebook dim
if not self.keep_num_codebooks_dim:
indices = rearrange(indices, '... 1 -> ...')
# complete aux loss
aux_loss = entropy_aux_loss * self.entropy_loss_weight + commit_loss * self.commitment_loss_weight
# returns
ret = Return(x, indices, aux_loss)
if not return_loss_breakdown:
return ret
return ret, LossBreakdown(per_sample_entropy, codebook_entropy, commit_loss)
class GroupedResidualBSQ(Module):
def __init__(
self,
*,
dim,
groups = 1,
accept_image_fmap = False,
**kwargs
):
super().__init__()
self.dim = dim
self.groups = groups
assert (dim % groups) == 0
dim_per_group = dim // groups
self.accept_image_fmap = accept_image_fmap
self.rvqs = nn.ModuleList([])
for _ in range(groups):
self.rvqs.append(LFQ(
dim = dim_per_group,
**kwargs
))
self.codebook_size = self.rvqs[0].codebook_size
@property
def codebooks(self):
return torch.stack(tuple(rvq.codebooks for rvq in self.rvqs))
@property
def split_dim(self):
return 1 if self.accept_image_fmap else -1
def get_codes_from_indices(self, indices):
codes = tuple(rvq.get_codes_from_indices(chunk_indices) for rvq, chunk_indices in zip(self.rvqs, indices))
return torch.stack(codes)
def get_output_from_indices(self, indices):
outputs = tuple(rvq.get_output_from_indices(chunk_indices) for rvq, chunk_indices in zip(self.rvqs, indices))
return torch.cat(outputs, dim = self.split_dim)
def forward(
self,
x,
return_all_codes = False
):
shape, split_dim = x.shape, self.split_dim
assert shape[split_dim] == self.dim
# split the feature dimension into groups
x = x.chunk(self.groups, dim = split_dim)
forward_kwargs = dict(
)
# invoke residual vq on each group
out = tuple(rvq(chunk, **forward_kwargs) for rvq, chunk in zip(self.rvqs, x))
out = tuple(zip(*out))
# otherwise, get all the zipped outputs and combine them
quantized, all_indices, *maybe_aux_loss = out
quantized = torch.cat(quantized, dim = split_dim)
all_indices = torch.stack(all_indices)
ret = (quantized, all_indices, *maybe_aux_loss)
return ret