BryanW's picture
Upload folder using huggingface_hub
3d1c0e1 verified
"""
Binary Spherical Quantization
Proposed in https://arxiv.org/abs/2406.07548
In the simplest setup, each dimension is quantized into {-1, 1}.
An entropy penalty is used to encourage utilization.
"""
import json
import random
import copy
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
import numpy as np
from einops import rearrange, reduce, pack, unpack
# from einx import get_at
from infinity.models.videovae.utils.dynamic_resolution import predefined_HW_Scales_dynamic
from infinity.models.videovae.utils.dynamic_resolution_two_pyramid import dynamic_resolution_thw, total_pixels2scales
from infinity.models.videovae.modules.quantizer.finite_scalar_quantization import FSQ
# print(f"{dynamic_resolution_thw=}")
# 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 round_up_multiple(num, mult):
return ceil(num / mult) * mult
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 repeat_schedule(scale_schedule, repeat_scales_num, times):
new_scale_schedule = []
for i in range(repeat_scales_num):
new_scale_schedule.extend([scale_schedule[i] for _ in range(times)])
new_scale_schedule.extend(scale_schedule[repeat_scales_num:])
return new_scale_schedule
def get_latent2scale_schedule(T: int, H: int, W: int, mode="original", last_scale_repeat_n=0, args=None):
predefined_HW_Scales = {}
if mode.startswith("infinity_video_two_pyramid"):
if 'elegant' in mode:
base_scale_schedule = copy.deepcopy(dynamic_resolution_thw[(H, W)]['scales'])
image_scale_repetition = json.loads(args.image_scale_repetition)
video_scale_repetition = json.loads(args.video_scale_repetition)
# print(f'{image_scale_repetition=} {video_scale_repetition=}')
base_scale_schedule = copy.deepcopy(dynamic_resolution_thw[(H, W)]['scales'])
def repeat_scales(base_scale_schedule, scale_repetition):
scale_schedule = []
for i in range(len(base_scale_schedule)):
scale_schedule.extend([base_scale_schedule[i] for _ in range(scale_repetition[i])])
return scale_schedule
image_scale_schedule = repeat_scales(base_scale_schedule, image_scale_repetition)
spatial_time_schedule = []
spatial_time_schedule.extend(image_scale_schedule)
firstframe_scalecnt = len(image_scale_schedule)
if T > 1:
scale_schedule = repeat_scales(base_scale_schedule, video_scale_repetition)
spatial_time_schedule.extend([(T-1, h, w) for i, (_, h, w) in enumerate(scale_schedule)])
# double h and w
tower_split_index = firstframe_scalecnt
# print(f'{spatial_time_schedule=}')
return spatial_time_schedule, tower_split_index
if "motion_boost_v2" in mode:
times = 6
base_scale_schedule = copy.deepcopy(dynamic_resolution_thw[(H, W)]['scales'])
image_scale_schedule = repeat_schedule(base_scale_schedule, 3, times)
spatial_time_schedule = []
spatial_time_schedule.extend(image_scale_schedule)
firstframe_scalecnt = len(image_scale_schedule)
if T > 1:
scale_schedule = repeat_schedule(base_scale_schedule, 7, times)
predefined_t = [T - 1 for _ in range(len(scale_schedule))]
spatial_time_schedule.extend([(min(int(np.round(predefined_t[i])), T - 1), h, w) for i, (_, h, w) in enumerate(scale_schedule)])
# double h and w
spatial_time_schedule_double = [(t, 2*h, 2*w) for (t, h, w) in spatial_time_schedule]
tower_split_index = firstframe_scalecnt
return spatial_time_schedule_double, tower_split_index
spatial_time_schedule = copy.deepcopy(dynamic_resolution_thw[(H, W)]['scales'])
spatial_time_schedule.extend(spatial_time_schedule[-1:] * last_scale_repeat_n)
tower_split_index = dynamic_resolution_thw[(H, W)]['tower_split_index'] + last_scale_repeat_n
if T > 1:
# predefined_t = np.linspace(1, compressed_frames - 1, len(scale_schedule))
if mode == "infinity_video_two_pyramid_full_time":
spatial_time_schedule.extend([(T - 1, h, w) for i, (_, h, w) in enumerate(spatial_time_schedule)])
else:
predefined_t = np.linspace(1, T - 1, total_pixels2scales['0.06M']-3).tolist() + [T - 1] * (len(spatial_time_schedule)-total_pixels2scales['0.06M']+3)
spatial_time_schedule.extend([(min(int(np.round(predefined_t[i])), T - 1), h, w) for i, (_, h, w) in enumerate(spatial_time_schedule)])
spatial_time_schedule.extend(spatial_time_schedule[-1:] * last_scale_repeat_n)
# double h and w
spatial_time_schedule_double = [(t, 2*h, 2*w) for (t, h, w) in spatial_time_schedule]
return spatial_time_schedule_double, tower_split_index
if mode == "original":
predefined_HW_Scales = {
# 256x256
(16, 16): [(1, 1), (2, 2), (3, 3), (4, 4), (5, 5), (6, 6), (8, 8), (10, 10), (13, 13), (16, 16)],
(36, 64): [(1, 1), (2, 2), (3, 3), (4, 4), (6, 6), (9, 12), (13, 16), (18, 24), (24, 32), (32, 48), (36, 64)],
(18, 32): [(1, 1), (2, 2), (3, 3), (4, 4), (6, 8), (8, 10), (10, 14), (12, 18), (14, 22), (16, 26), (18, 32)],
(30, 53): [(1, 1), (2, 2), (3, 3), (4, 7), (6, 11), (8, 14), (12, 21), (16, 28), (20, 35), (22, 39), (24, 42), (26, 46), (28, 50), (30, 53)]
}
predefined_HW_Scales[(32, 32)] = predefined_HW_Scales[(16, 16)] + [(20, 20), (24, 24), (32, 32)]
predefined_HW_Scales[(64, 64)] = predefined_HW_Scales[(32, 32)] + [(40, 40), (48, 48), (64, 64)]
elif mode == "dynamic":
predefined_HW_Scales.update(predefined_HW_Scales_dynamic)
elif mode == "dense":
predefined_HW_Scales[(16, 16)] = [(x, x) for x in range(1, 16+1)]
predefined_HW_Scales[(32, 32)] = predefined_HW_Scales[(16, 16)] + [(20, 20), (24, 24), (28, 28), (32, 32)]
predefined_HW_Scales[(64, 64)] = predefined_HW_Scales[(32, 32)] + [(40, 40), (48, 48), (56, 56), (64, 64)]
elif mode == "dense_f8":
# predefined_HW_Scales[(16, 16)] = [(x, x) for x in range(1, 16+1)]
predefined_HW_Scales[(32, 32)] = [(x, x) for x in range(1, 16+1)] + [(20, 20), (24, 24), (28, 28), (32, 32)]
predefined_HW_Scales[(64, 64)] = predefined_HW_Scales[(32, 32)] + [(40, 40), (48, 48), (56, 56), (64, 64)]
predefined_HW_Scales[(128, 128)] = predefined_HW_Scales[(64, 64)] + [(80, 80), (96, 96), (112, 112), (128, 128)]
elif mode == "dense_f8_double":
# predefined_HW_Scales setting double from dense f16
predefined_HW_Scales[(32, 32)] = [(x, x) for x in range(1, 16+1)]
predefined_HW_Scales[(64, 64)] = predefined_HW_Scales[(32, 32)] + [(20, 20), (24, 24), (28, 28), (32, 32)]
predefined_HW_Scales[(96, 96)] = predefined_HW_Scales[(64, 64)] + [(40, 40), (48, 48)]
predefined_HW_Scales[(128, 128)] = predefined_HW_Scales[(64, 64)] + [(40, 40), (48, 48), (56, 56), (64, 64)]
predefined_HW_Scales[(24, 42)] = [(1, 1), (2, 2), (3, 3), (3, 4), (3, 5), (4, 6), (4, 7), (5, 8), (6, 9), (6, 10), (6, 11), (7, 12), (7, 13), (8, 14), (9, 15), (9, 16), (12, 21)]
predefined_HW_Scales[(36, 64)] = predefined_HW_Scales[(24, 42)] + [(14, 26), (18, 32)]
predefined_HW_Scales[(60, 108)] = predefined_HW_Scales[(36, 64)] + [(24, 42), (30, 54)]
predefined_HW_Scales[(90, 160)] = predefined_HW_Scales[(60, 108)] + [(38, 66),(45, 80)]
for k, v in predefined_HW_Scales.items():
predefined_HW_Scales[k] = [(2*x, 2*y) for (x, y) in v]
elif mode.startswith("same"):
num_quant = int(mode[len("same"):])
predefined_HW_Scales[(16, 16)] = [(16, 16) for _ in range(num_quant)]
predefined_HW_Scales[(32, 32)] = [(32, 32) for _ in range(num_quant)]
predefined_HW_Scales[(64, 64)] = [(64, 64) for _ in range(num_quant)]
elif mode == "half":
predefined_HW_Scales[(32, 32)] = [(1, 1), (2, 2), (3, 3), (4, 4), (5, 5), (6, 6), (8, 8), (10, 10), (13, 13), (16, 16)]
predefined_HW_Scales[(64, 64)] = [(1,1),(2,2),(4,4),(6,6),(8,8),(12,12),(16,16)]
else:
raise NotImplementedError
# predefined_T_Scales = [1, 2, 3, 4, 5, 6, 7, 9, 11, 13, 17, 17, 17, 17, 17, 17]
# predefined_T_Scales = [1, 2, 3, 4, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27]
predefined_T_Scales = [1, 2, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29]
# predefined_T_Scales = [1, 2, 3, 5, 6, 8, 9, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25]
patch_THW_shape_per_scale = predefined_HW_Scales[(H, W)]
if len(predefined_T_Scales) < len(patch_THW_shape_per_scale):
# print("warning: the length of predefined_T_Scales is less than the length of patch_THW_shape_per_scale!")
predefined_T_Scales += [predefined_T_Scales[-1]] * (len(patch_THW_shape_per_scale) - len(predefined_T_Scales))
patch_THW_shape_per_scale = [(min(T, t), h, w ) for (h, w), t in zip(patch_THW_shape_per_scale, predefined_T_Scales[:len(patch_THW_shape_per_scale)])]
return patch_THW_shape_per_scale
def interpolate(tensor, size, mode):
"""
arguments:
tensor: (B,C,T,H,W)
size: (C1,T,H1,W1)
mode: str
return:
tensor: (B,*size)
"""
C1, T, H1, W1 = size
tensor = tensor.permute(0,2,1,3,4) # (B,C,T,H,W) -> (B,T,C,H,W)
tensor = F.interpolate(tensor, size=(C1, H1, W1), mode=mode)
tensor = tensor.permute(0,2,1,3,4) # (B,T,C1,H1,W1) -> (B,C1,T,H1,W1)
return tensor
# TP: Two Pyramid
class MultiScaleBSQTP(Module):
""" Follows Algorithm 1. in https://arxiv.org/pdf/2107.03312.pdf """
def __init__(
self,
*,
dim,
soft_clamp_input_value = None,
aux_loss = False, # intermediate auxiliary loss
use_stochastic_depth=False,
drop_rate=0.,
schedule_mode="original", # ["original", "dynamic", "dense"]
keep_first_quant=False,
keep_last_quant=False,
remove_residual_detach=False,
random_flip = False,
flip_prob = 0.5,
flip_mode = "stochastic", # "stochastic", "deterministic"
max_flip_lvl = 1,
random_flip_1lvl = False, # random flip one level each time
flip_lvl_idx = None,
drop_when_test=False,
drop_lvl_idx=None,
drop_lvl_num=0,
random_short_schedule = False, # randomly use short schedule (schedule for images of 256x256)
short_schedule_prob = 0.5,
disable_flip_prob = 0.0, # disable random flip in this image
casual_multi_scale = False, # causal multiscale
temporal_slicing = False,
last_scale_repeat_n = 0,
num_lvl_fsq = None,
other_args = None,
**kwargs
):
super().__init__()
codebook_dim = dim
self.use_stochastic_depth = use_stochastic_depth
self.drop_rate = drop_rate
self.remove_residual_detach = remove_residual_detach
self.random_flip = random_flip
self.flip_prob = flip_prob
self.flip_mode = flip_mode
self.max_flip_lvl = max_flip_lvl
self.random_flip_1lvl = random_flip_1lvl
self.flip_lvl_idx = flip_lvl_idx
assert (random_flip and random_flip_1lvl) == False
self.disable_flip_prob = disable_flip_prob
self.casual_multi_scale = casual_multi_scale
self.temporal_slicing = temporal_slicing
self.last_scale_repeat_n = last_scale_repeat_n
# print(f"{casual_multi_scale=}")
self.drop_when_test = drop_when_test
self.drop_lvl_idx = drop_lvl_idx
self.drop_lvl_num = drop_lvl_num
if self.drop_when_test:
assert drop_lvl_idx is not None
assert drop_lvl_num > 0
self.random_short_schedule = random_short_schedule
self.short_schedule_prob = short_schedule_prob
self.z_interplote_up = 'trilinear'
self.z_interplote_down = 'area'
self.schedule_mode = schedule_mode
self.keep_first_quant = keep_first_quant
self.keep_last_quant = keep_last_quant
if self.use_stochastic_depth and self.drop_rate > 0:
assert self.keep_first_quant or self.keep_last_quant
self.full2short = {7:7, 10:7, 13:7, 16:16, 20:16, 24:16}
if self.schedule_mode == 'dense_f8':
self.full2short_f8 = {20:20, 24:24, 28:24}
elif self.schedule_mode == 'dense_f8_double':
self.full2short_f8 = {16: 14, 17: 14, 19: 14, 20:14, 21:14, 22:14, 24:14}
elif self.schedule_mode.startswith("infinity_video_two_pyramid"):
self.full2short_f8 = {11: 11, 13: 11, 14: 11, 16: 11, 29: 26, 28: 26, 26: 26}
self.other_args = other_args
print(f'{self.other_args=}')
self.origin_C = self.other_args.detail_scale_dim
self.detail_scale_dim, self.semantic_scale_dim = self.other_args.detail_scale_dim, self.other_args.semantic_scale_dim
self.semantic_scales = other_args.semantic_scales
if self.other_args.semantic_num_lvl == 2:
self.lfq_semantic = BSQ(
dim = self.semantic_scale_dim,
codebook_scale = 1,
soft_clamp_input_value = soft_clamp_input_value,
**kwargs,
)
else:
assert self.other_args.semantic_num_lvl >= 2, f'{self.other_args.semantic_num_lvl=} is not supported'
self.lfq_semantic = FSQ(
num_lvl = self.other_args.semantic_num_lvl,
dim = self.semantic_scale_dim,
)
if self.other_args.detail_num_lvl == 2:
self.lfq_detail = BSQ(
dim = self.detail_scale_dim,
codebook_scale = 1,
soft_clamp_input_value = soft_clamp_input_value,
**kwargs,
)
else:
assert self.other_args.detail_num_lvl >= 2, f'{self.other_args.semantic_num_lvl=} is not supported'
self.lfq_detail = FSQ(
num_lvl = self.other_args.detail_num_lvl,
dim = self.detail_scale_dim,
)
@property
def codebooks(self):
return self.lfq_detail.codebook
def get_codes_from_indices(self, indices_list):
all_codes = []
for indices in indices_list:
# indices: [B,t,h,w,d]
if indices.shape[-1] == self.origin_C:
codes = self.lfq.indices_to_codes(indices)
elif indices.shape[-1] == self.semantic_scale_dim:
codes = self.lfq_semantic.indices_to_codes(indices)
else:
raise NotImplementedError(f'indices shape {indices.shape} not supported')
all_codes.append(codes)
_, _, T, H, W = all_codes[-1].size()
summed_codes = 0
for code in all_codes:
summed_codes += F.interpolate(code, size=(T, H, W), mode=self.z_interplote_up)
return summed_codes
def get_output_from_indices(self, indices):
codes = self.get_codes_from_indices(indices)
codes_summed = reduce(codes, 'q ... -> ...', 'sum')
return codes_summed
def flip_quant(self, x):
# assert self.flip_mode in ['stochastic', 'stochastic_dynamic']
if self.flip_mode == 'stochastic':
flip_mask = torch.rand_like(x) < self.flip_prob
elif self.flip_mode == 'stochastic_dynamic':
flip_prob = random.uniform(0, self.flip_prob)
flip_mask = torch.rand_like(x) < flip_prob
else:
raise NotImplementedError
x = x.clone()
x[flip_mask] = -x[flip_mask]
return x
def forward(
self,
x_list,
mask = None,
return_all_codes = False,
):
assert len(x_list) <= 2
multi_scale = len(x_list) == 2
for i in range(len(x_list)):
if x_list[i].ndim == 4:
x_list[i] = x_list[i].unsqueeze(2)
B, C, T, H, W = x_list[-1].size()
if self.schedule_mode.startswith("same"):
scale_num = int(self.schedule_mode[len("same"):])
assert T == 1
scale_schedule = [(1, H, W)] * scale_num
elif self.schedule_mode.startswith("infinity_video_two_pyramid"):
scale_schedule, tower_split_index = get_latent2scale_schedule(T, H, W, mode=self.schedule_mode, last_scale_repeat_n=self.last_scale_repeat_n, args=self.other_args)
scale_num = len(scale_schedule)
else:
scale_schedule = get_latent2scale_schedule(T, H, W, mode=self.schedule_mode, args=self.other_args)
scale_num = len(scale_schedule)
quantized_out = torch.zeros((B, C, 1, 1, 1), device=x_list[-1].device, dtype=x_list[-1].dtype)
quantized_out_firstframe = None
all_losses = []
all_indices = []
# cal scale_in_one_clip
unique_scale_schedule = [scale_schedule[0]]
scale_in_one_clip = 1
for si in range(1, len(scale_schedule)):
if np.array(scale_schedule[si]).prod() < np.array(scale_schedule[si-1]).prod():
break
if scale_schedule[si] != scale_schedule[si-1]:
unique_scale_schedule.append(scale_schedule[si])
scale_in_one_clip += 1
current_scale_in_one_clip = 0
must_preserve_scales = []
if self.other_args.quant_not_rely_256:
must_preserve_scales = [11]
with autocast('cuda', enabled = False):
for si, (pt, ph, pw) in enumerate(scale_schedule):
if si > 0 and scale_schedule[si] != scale_schedule[si-1]:
current_scale_in_one_clip += 1
current_scale_in_one_clip = current_scale_in_one_clip % scale_in_one_clip
last_step_in_one_scale = False
if si < len(scale_schedule)-1 and (scale_schedule[si] != scale_schedule[si+1]):
last_step_in_one_scale = True
if si == len(scale_schedule)-1:
last_step_in_one_scale = True
if si < tower_split_index:
ss, ee = 0, 1
else:
ss, ee = 1, T
if multi_scale and current_scale_in_one_clip < self.other_args.scales_256:
target = x_list[0][:,:,ss:ee]
else:
target = x_list[-1][:,:,ss:ee]
tgt_shape = target.shape[-4:]
skip_this_scale = False
if current_scale_in_one_clip < self.semantic_scales:
C1 = self.semantic_scale_dim
lfq = self.lfq_semantic
else:
C1 = self.detail_scale_dim
lfq = self.lfq_detail
if current_scale_in_one_clip not in must_preserve_scales:
skip_this_scale = random.random() < self.other_args.skip_detail_scales_prob
if not skip_this_scale:
quantized_out = interpolate(quantized_out, size=tgt_shape, mode=self.z_interplote_up)
interpolate_residual = interpolate(target-quantized_out, size=(C1, pt, ph, pw), mode=self.z_interplote_down)
quantized, indices, loss = lfq(interpolate_residual)
quantized = interpolate(quantized, size=tgt_shape, mode=self.z_interplote_up)
all_indices.append(indices)
all_losses.append(loss)
quantized_out = quantized_out + quantized
if si == tower_split_index - 1:
quantized_out_firstframe = quantized_out.clone()
quantized_out = quantized_out * 0. # set to zero
if multi_scale and si < tower_split_index and last_step_in_one_scale and current_scale_in_one_clip == self.other_args.scales_256-1:
quantized_out_firstframe_256 = quantized_out.clone()
if self.other_args.quant_not_rely_256:
quantized_out = quantized_out * 0.
if multi_scale and si >= tower_split_index and last_step_in_one_scale and current_scale_in_one_clip == self.other_args.scales_256-1:
quantized_out_256 = quantized_out.clone()
if self.other_args.quant_not_rely_256:
quantized_out = quantized_out * 0.
quantized_out_list = []
if T == 1:
if multi_scale:
quantized_out_list.append(quantized_out_firstframe_256)
quantized_out_list.append(quantized_out_firstframe)
else:
quantized_out_list.append(quantized_out_firstframe)
else:
if multi_scale:
quantized_out_256 = torch.cat([quantized_out_firstframe_256, quantized_out_256], dim=2)
quantized_out_list.append(quantized_out_256)
quantized_out = torch.cat([quantized_out_firstframe, quantized_out], dim=2)
quantized_out_list.append(quantized_out)
else:
quantized_out = torch.cat([quantized_out_firstframe, quantized_out], dim=2)
quantized_out_list.append(quantized_out)
all_losses = torch.stack(all_losses, dim = -1)
ret = (quantized_out_list, all_indices, all_losses)
if not return_all_codes:
return ret
# whether to return all codes from all codebooks across layers
all_codes = self.get_codes_from_indices(all_indices)
# will return all codes in shape (quantizer, batch, sequence length, codebook dimension)
return (*ret, all_codes)
class BSQ(Module):
def __init__(
self,
*,
dim = None,
entropy_loss_weight = 0.1,
commitment_loss_weight = 0.25,
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 = 1., # make less than 1. to only use a random fraction of the probs for per sample entropy
soft_clamp_input_value = 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
inv_temperature = 100.0,
gamma0=1.0, gamma=1.0, zeta=1.0,
use_out_phi = False, # use output phi network
use_out_phi_res = False, # residual out phi
use_bernoulli = False,
use_rot_trick = False,
):
super().__init__()
# some assert validations
assert exists(dim) , 'dim must be specified for BSQ'
codebook_dim = dim
codebook_dims = codebook_dim * num_codebooks
dim = default(dim, codebook_dims)
self.codebook_dims = codebook_dims
self.out_phi = nn.Linear(codebook_dims, codebook_dims) if use_out_phi else nn.Identity()
self.use_out_phi_res = use_out_phi_res
if self.use_out_phi_res:
self.out_phi_scale = nn.Parameter(torch.zeros(codebook_dims), requires_grad=True) # init as zero
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
# For BSQ (binary spherical quantization)
if not spherical:
raise ValueError("For BSQ, spherical must be True.")
self.persample_entropy_compute = 'analytical'
self.inv_temperature = inv_temperature
self.gamma0 = gamma0 # loss weight for entropy penalty
self.gamma = gamma # loss weight for entropy penalty
self.zeta = zeta # loss weight for entire entropy penalty
self.use_bernoulli = use_bernoulli
self.use_rot_trick = use_rot_trick
# entropy aux loss related weights
assert 0 < frac_per_sample_entropy <= 1.
self.frac_per_sample_entropy = frac_per_sample_entropy
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
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 = l2norm(codes) # must normalize when using BSQ
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)
# rearrange codes back to original shape
if should_transpose:
codes = rearrange(codes, 'b ... d -> b d ...')
return codes
def quantize(self, z):
assert z.shape[-1] == self.codebook_dims, f"Expected {self.codebook_dims} dimensions, got {z.shape[-1]}"
zhat = torch.where(z > 0,
torch.tensor(1, dtype=z.dtype, device=z.device),
torch.tensor(-1, dtype=z.dtype, device=z.device))
q_scale = 1. / (self.codebook_dims ** 0.5)
zhat = q_scale * zhat # on unit sphere
return z + (zhat - z).detach()
def quantize_new_bernoulli(self, z, prob_z):
assert z.shape[-1] == self.codebook_dims, f"Expected {self.codebook_dims} dimensions, got {z.shape[-1]}"
zhat = (torch.bernoulli(prob_z) - 0.5) * 2.0
q_scale = 1. / (self.codebook_dims ** 0.5)
zhat = q_scale * zhat # on unit sphere
return z + (zhat - z).detach()
def rot_quantize(self, z, inference=False):
assert z.shape[-1] == self.codebook_dims, f"Expected {self.codebook_dims} dimensions, got {z.shape[-1]}"
q_scale = 1. / (self.codebook_dims ** 0.5)
zhat = torch.where(z > 0,
torch.tensor(1, dtype=z.dtype, device=z.device),
torch.tensor(-1, dtype=z.dtype, device=z.device)) * q_scale
if inference:
return zhat
w = ((z + zhat) / torch.norm(z + zhat, dim=-1, keepdim=True)).detach()
z = z.unsqueeze(1) - 2*torch.bmm(torch.bmm(z.unsqueeze(1), w.unsqueeze(-1)), w.unsqueeze(1)) + 2 * torch.bmm(
torch.bmm(z.unsqueeze(1), z.unsqueeze(-1).detach()), zhat.unsqueeze(1).detach())
return z.squeeze()
def soft_entropy_loss(self, z):
if self.persample_entropy_compute == 'analytical':
# if self.l2_norm:
p = torch.sigmoid(-4 * z / (self.codebook_dims ** 0.5) * self.inv_temperature)
# else:
# p = torch.sigmoid(-4 * z * self.inv_temperature)
prob = torch.stack([p, 1-p], dim=-1) # (b, h, w, 18, 2)
per_sample_entropy = self.get_entropy(prob, dim=-1, normalize=False).sum(dim=-1).mean() # (b,h,w,18)->(b,h,w)->scalar
else:
per_sample_entropy = self.get_entropy(prob, dim=-1, normalize=False).sum(dim=-1).mean()
# macro average of the probability of each subgroup
avg_prob = reduce(prob, '... g d ->g d', 'mean') # (18, 2)
codebook_entropy = self.get_entropy(avg_prob, dim=-1, normalize=False)
# the approximation of the entropy is the sum of the entropy of each subgroup
return per_sample_entropy, codebook_entropy.sum(), avg_prob
def get_entropy(self, count, dim=-1, eps=1e-4, normalize=True):
if normalize: # False
probs = (count + eps) / (count + eps).sum(dim=dim, keepdim =True)
else: # True
probs = count
H = -(probs * torch.log(probs + 1e-8)).sum(dim=dim)
return H
def forward(
self,
x,
return_loss_breakdown = False,
mask = None,
entropy_weight=0.1
):
"""
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') # x.shape [b, hwt, c]
assert x.shape[-1] == self.dim, f'expected dimension of {self.dim} but received {x.shape[-1]}'
# split out number of codebooks
x = rearrange(x, 'b n (c d) -> b n c d', c = self.num_codebooks)
if self.use_bernoulli: # False
prob_x = torch.sigmoid(x)
x = 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()
# use straight-through gradients
if self.use_rot_trick: # False
x_f = x.flatten(end_dim=-2) # (b, hwt, 1, d) -> (bhwt, d)
q_f = self.rot_quantize(x_f, inference= not self.training)
quantized = q_f.reshape(x.shape)
elif self.use_bernoulli: # False
quantized = self.quantize_new_bernoulli(x, prob_x)
else:
quantized = self.quantize(x)
# calculate indices
indices = reduce((quantized > 0).int() * self.mask.int(), 'b n c d -> b n c', 'sum')
# entropy aux loss
if self.training:
persample_entropy, cb_entropy, avg_prob = self.soft_entropy_loss(x) # compute entropy
entropy_penalty = self.gamma0 * persample_entropy - self.gamma * cb_entropy
else:
# if not training, just return dummy 0
entropy_penalty = persample_entropy = cb_entropy = self.zero
# commit loss
if self.training and self.commitment_loss_weight > 0.:
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
# input back to original dtype if needed
if force_f32:
x = x.type(orig_dtype)
# merge back codebook dim
x = quantized # rename quantized to x for output
if self.use_out_phi_res: # False
x = x + self.out_phi_scale * self.out_phi(x) # apply out_phi on quant output as residual
else:
x = self.out_phi(x) # apply out_phi on quant output
x = rearrange(x, 'b n c d -> b n (c d)')
# 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 = commit_loss * self.commitment_loss_weight + (self.zeta * entropy_penalty / self.inv_temperature)*entropy_weight
# returns
ret = Return(x, indices, aux_loss)
if not return_loss_breakdown:
return ret
return ret, LossBreakdown(persample_entropy, cb_entropy, commit_loss)
# Copyright (c) 2025 FoundationVision
# SPDX-License-Identifier: MIT