-
Notifications
You must be signed in to change notification settings - Fork 204
/
lookup_free_quantization.py
364 lines (252 loc) · 11.2 KB
/
lookup_free_quantization.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
"""
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
from collections import namedtuple
import torch
from torch import nn, einsum
import torch.nn.functional as F
from torch.nn import Module
from torch.cuda.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'])
# 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 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
# class
class LFQ(Module):
def __init__(
self,
*,
dim = None,
codebook_size = None,
entropy_loss_weight = 0.1,
commitment_loss_weight = 0.25,
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 = 1., # make less than 1. to only use a random fraction of the probs for per sample entropy
use_code_agnostic_commit_loss = False,
projection_has_bias = True,
soft_clamp_input_value = None,
cosine_sim_project_in = False,
cosine_sim_project_in_scale = None,
channel_first = None
):
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 = 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
# 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
self.use_code_agnostic_commit_loss = use_code_agnostic_commit_loss
# 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
# 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)
# codes
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, 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 = 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
@autocast(enabled = False)
def forward(
self,
x,
inv_temperature = 100.,
return_loss_breakdown = False,
mask = None,
):
"""
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 = x.float()
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)
# 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)
# 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
# calculate indices
indices = reduce((x > 0).int() * self.mask.int(), 'b n c d -> b n c', 'sum')
# entropy aux loss
if self.training:
# the same as euclidean distance up to a constant
distance = -2 * einsum('... i d, j d -> ... i j', original_input, self.codebook)
prob = (-distance * inv_temperature).softmax(dim = -1)
# account for mask
if exists(mask):
prob = prob[mask]
else:
prob = rearrange(prob, 'b n ... -> (b n) ...')
# whether to only use a fraction of probs, for reducing memory
if self.frac_per_sample_entropy < 1.:
num_tokens = prob.shape[0]
num_sampled_tokens = int(num_tokens * self.frac_per_sample_entropy)
rand_mask = torch.randn(num_tokens).argsort(dim = -1) < num_sampled_tokens
per_sample_probs = prob[rand_mask]
else:
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')
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
# commit loss
if self.training and self.commitment_loss_weight > 0.:
if self.use_code_agnostic_commit_loss:
# credit goes to @MattMcPartlon for sharing this in https://github.com/lucidrains/vector-quantize-pytorch/issues/120#issuecomment-2095089337
commit_loss = F.mse_loss(
original_input ** 2,
codebook_value ** 2,
reduction = 'none'
)
else:
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
# 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
ret = Return(x, indices, aux_loss)
if not return_loss_breakdown:
return ret
return ret, LossBreakdown(per_sample_entropy, codebook_entropy, commit_loss)