hma / magvit2 /modules /vqvae /lookup_free_quantize.py
LeroyWaa's picture
draft
246c106
raw
history blame
11.2 kB
"""
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.
Refer to
https://github.com/lucidrains/vector-quantize-pytorch/blob/master/vector_quantize_pytorch/lookup_free_quantization.py
https://github.com/theAdamColton/ijepa-enhanced/blob/7edef5f7288ae8f537f0db8a10044a2a487f70c9/ijepa_enhanced/lfq.py
"""
"""
Modified Open-MAGVIT2 code to use VQConfig.
"""
from math import log2, ceil
from collections import namedtuple
import torch
from torch import nn, einsum
import torch.nn.functional as F
from torch.nn import Module
from einops import rearrange, reduce, pack, unpack
from magvit2.config import VQConfig
# constants
LossBreakdown = namedtuple('LossBreakdown', ['per_sample_entropy', 'codebook_entropy', 'commitment', 'avg_probs'])
# helper functions
def exists(v):
return v is not None
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]
# entropy
def entropy(prob):
return (-prob * torch.log(prob + 1e-5)).sum(dim=-1)
# class
def mult_along_first_dims(x, y):
"""
returns x * y elementwise along the leading dimensions of y
"""
ndim_to_expand = x.ndim - y.ndim
for _ in range(ndim_to_expand):
y = y.unsqueeze(-1)
return x * y
def masked_mean(x, m):
"""
takes the mean of the elements of x that are not masked
the mean is taken along the shared leading dims of m
equivalent to: x[m].mean(tuple(range(m.ndim)))
The benefit of using masked_mean rather than using
tensor indexing is that masked_mean is much faster
for torch-compile on batches.
The drawback is larger floating point errors
"""
x = mult_along_first_dims(x, m)
x = x / m.sum()
return x.sum(tuple(range(m.ndim)))
def entropy_loss(
logits,
mask=None,
temperature=0.01,
sample_minimization_weight=1.0,
batch_maximization_weight=1.0,
eps=1e-5,
):
"""
Entropy loss of unnormalized logits
logits: Affinities are over the last dimension
https://github.com/google-research/magvit/blob/05e8cfd6559c47955793d70602d62a2f9b0bdef5/videogvt/train_lib/losses.py#L279
LANGUAGE MODEL BEATS DIFFUSION — TOKENIZER IS KEY TO VISUAL GENERATION (2024)
"""
probs = F.softmax(logits / temperature, -1)
log_probs = F.log_softmax(logits / temperature + eps, -1)
if mask is not None:
avg_probs = masked_mean(probs, mask)
else:
avg_probs = reduce(probs, "... D -> D", "mean")
avg_entropy = -torch.sum(avg_probs * torch.log(avg_probs + eps))
sample_entropy = -torch.sum(probs * log_probs, -1)
if mask is not None:
sample_entropy = masked_mean(sample_entropy, mask).mean()
else:
sample_entropy = torch.mean(sample_entropy)
loss = (sample_minimization_weight * sample_entropy) - (
batch_maximization_weight * avg_entropy
)
return sample_entropy, avg_entropy, loss
class LFQ(Module):
def __init__(self, config: VQConfig):
super().__init__()
# some assert validations
assert exists(config.z_channels) or exists(config.codebook_size), \
"either dim or codebook_size must be specified for LFQ"
assert not exists(config.codebook_size) or log2(config.codebook_size).is_integer(), \
f"your codebook size must be a power of 2 for lookup free quantization (suggested {2 ** ceil(log2(config.codebook_size))})"
self.codebook_size = default(config.codebook_size, lambda: 2 ** dim)
self.codebook_dim = int(log2(config.codebook_size))
codebook_dims = self.codebook_dim * config.num_codebooks
dim = default(config.z_channels, codebook_dims)
has_projections = dim != codebook_dims
self.has_projections = has_projections
self.dim = dim
self.codebook_dim = self.codebook_dim
self.num_codebooks = config.num_codebooks
# for entropy loss
self.sample_minimization_weight = config.sample_minimization_weight
self.batch_maximization_weight = config.batch_maximization_weight
# for no auxiliary loss, during inference
self.token_factorization = config.token_factorization # only utilized in second stage
if not self.token_factorization: # for first stage model
self.register_buffer('mask', 2 ** torch.arange(self.codebook_dim - 1, -1, -1), persistent=False)
else:
k = self.codebook_dim // 2
self.register_buffer("mask", 2 ** torch.arange(k - 1, -1, -1), persistent=False)
self.register_buffer('zero', torch.tensor(0.), persistent=False)
# codes
all_codes = torch.arange(config.codebook_size)
bits = self.indices_to_bits(all_codes)
codebook = bits * 2.0 - 1.0
self.register_buffer('codebook', codebook, persistent=False)
@property
def dtype(self):
return self.codebook.dtype
def indices_to_bits(self, x):
"""
x: long tensor of indices for constructing codebook, but actually not utilized in all the experiments.
returns big endian bits
"""
mask = 2 ** torch.arange(self.codebook_dim, device=x.device, dtype=torch.long)
# x is now big endian bits, the last dimension being the bits
x = (x.unsqueeze(-1) & mask) != 0
return x
def get_codebook_entry(self, x, bhwc):
if self.token_factorization:
k = self.codebook_dim // 2
mask = 2 ** torch.arange(k - 1, -1, -1, device=x.device, dtype=torch.long)
else:
mask = 2 ** torch.arange(self.codebook_dim-1, -1, -1, device=x.device, dtype=torch.long)
x = (x.unsqueeze(-1) & mask) != 0 # find its bit representation
x = x * 2.0 - 1.0 #back to the float
## scale back to the desired shape
b, h, w, c = bhwc
x = rearrange(x, "b (h w) c -> b h w c", h=h, w=w, c=c)
x = rearrange(x, "b h w c -> b c h w")
return x
def bits_to_indices(self, bits):
"""
bits: bool tensor of big endian bits, where the last dimension is the bit dimension
returns indices, which are long integers from 0 to self.codebook_size
"""
assert bits.shape[-1] == self.codebook_dim
indices = 2 ** torch.arange(
0,
self.codebook_dim,
1,
dtype=torch.long,
device=bits.device,
)
return (bits * indices).sum(-1)
def decode(self, x):
"""
x: ... NH
where NH is number of codebook heads
A longtensor of codebook indices, containing values from
0 to self.codebook_size
"""
x = self.indices_to_bits(x)
# to some sort of float
x = x.to(self.dtype)
# -1 or 1
x = x * 2 - 1
x = rearrange(x, "... NC Z-> ... (NC Z)")
return x
def forward(
self,
x,
return_loss_breakdown=False,
mask=None,
return_loss=True,
flip=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
"""
x = rearrange(x, 'b d ... -> b ... d')
x, ps = pack_one(x, 'b * d')
# split out number of codebooks
x = rearrange(x, 'b n (c d) -> b n c d', c=self.num_codebooks)
codebook_value = torch.Tensor([1.0]).to(device=x.device, dtype=x.dtype)
quantized = torch.where(x > 0, codebook_value, -codebook_value) # higher than 0 filled
# calculate indices
if self.token_factorization:
k = self.codebook_dim // 2
indices_pre = reduce((quantized[..., :k] > 0).int() * self.mask.int(), "b n c d -> b n c", "sum")
indices_post = reduce((quantized[..., k:] > 0).int() * self.mask.int(), "b n c d -> b n c", "sum")
# indices_post = 2**k + indices_post #shifter to the 1024
else:
if not flip:
indices = reduce((quantized > 0).int() * self.mask.int(), 'b n c d -> b n c', 'sum')
else:
# not sure why this is necessary
indices = reduce((quantized > 0).flip(-1).int() * self.mask.int(), 'b n c d -> b n c', 'sum')
# entropy aux loss
if self.training and return_loss:
logits = 2 * einsum('... i d, j d -> ... i j', x, self.codebook)
# the same as Euclidean distance up to a constant
per_sample_entropy, codebook_entropy, entropy_aux_loss = entropy_loss(
logits=logits,
sample_minimization_weight=self.sample_minimization_weight,
batch_maximization_weight=self.batch_maximization_weight
)
avg_probs = self.zero
else:
## calculate the codebook_entropy needed for one batch evaluation
#------------------------------------------------------------------
# logits = 2 * einsum('... i d, j d -> ... i j', x, self.codebook)
# probs = F.softmax(logits / 0.01, -1)
# avg_probs = reduce(probs, "b n c d -> b d", "mean")
# avg_probs = torch.sum(avg_probs, 0) #batch dimension
#-------------------------------------------------------------------
# if not training, just return dummy 0
per_sample_entropy = codebook_entropy = self.zero
entropy_aux_loss = self.zero
avg_probs = self.zero
# commit loss
if self.training:
commit_loss = F.mse_loss(x, quantized.detach(), reduction='none')
if exists(mask):
commit_loss = commit_loss[mask]
commit_loss = commit_loss.mean()
else:
commit_loss = self.zero
# use straight-through gradients (optionally with custom activation fn) if training
quantized = x + (quantized - x).detach() # transfer to quantized
# merge back codebook dim
quantized = rearrange(quantized, 'b n c d -> b n (c d)')
# reconstitute image or video dimensions
quantized = unpack_one(quantized, ps, 'b * d')
quantized = rearrange(quantized, 'b ... d -> b d ...')
if self.token_factorization:
indices_pre = unpack_one(indices_pre, ps, "b * c")
indices_post = unpack_one(indices_post, ps, "b * c")
indices_pre = indices_pre.flatten()
indices_post = indices_post.flatten()
indices = (indices_pre, indices_post)
else:
indices = unpack_one(indices, ps, 'b * c')
indices = indices.flatten()
ret = (quantized, entropy_aux_loss, indices)
if not return_loss_breakdown:
return ret
return ret, LossBreakdown(per_sample_entropy, codebook_entropy, commit_loss, avg_probs)