Spaces:
Runtime error
Runtime error
File size: 9,560 Bytes
e276be2 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from torch import einsum
from einops import rearrange
class VectorQuantizer2(nn.Module):
"""
Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly
avoids costly matrix multiplications and allows for post-hoc remapping of indices.
"""
# NOTE: due to a bug the beta term was applied to the wrong term. for
# backwards compatibility we use the buggy version by default, but you can
# specify legacy=False to fix it.
def __init__(self, n_e, e_dim, beta, remap=None, unknown_index="random", sane_index_shape=False, legacy=True):
super().__init__()
self.n_e = n_e
self.e_dim = e_dim
self.beta = beta
self.legacy = legacy
self.embedding = nn.Embedding(self.n_e, self.e_dim)
self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
self.remap = remap
if self.remap is not None:
self.register_buffer("used", torch.tensor(np.load(self.remap)))
self.re_embed = self.used.shape[0]
self.unknown_index = unknown_index # "random" or "extra" or integer
if self.unknown_index == "extra":
self.unknown_index = self.re_embed
self.re_embed = self.re_embed + 1
print(
f"Remapping {self.n_e} indices to {self.re_embed} indices. "
f"Using {self.unknown_index} for unknown indices."
)
else:
self.re_embed = n_e
self.sane_index_shape = sane_index_shape
def remap_to_used(self, inds):
ishape = inds.shape
assert len(ishape) > 1
inds = inds.reshape(ishape[0], -1)
used = self.used.to(inds)
match = (inds[:, :, None] == used[None, None, ...]).long()
new = match.argmax(-1)
unknown = match.sum(2) < 1
if self.unknown_index == "random":
new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to(device=new.device)
else:
new[unknown] = self.unknown_index
return new.reshape(ishape)
def unmap_to_all(self, inds):
ishape = inds.shape
assert len(ishape) > 1
inds = inds.reshape(ishape[0], -1)
used = self.used.to(inds)
if self.re_embed > self.used.shape[0]: # extra token
inds[inds >= self.used.shape[0]] = 0 # simply set to zero
back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
return back.reshape(ishape)
def forward(self, z, temp=None, rescale_logits=False, return_logits=False):
assert temp is None or temp == 1.0, "Only for interface compatible with Gumbel"
assert rescale_logits == False, "Only for interface compatible with Gumbel"
assert return_logits == False, "Only for interface compatible with Gumbel"
# reshape z -> (batch, height, width, channel) and flatten
z = rearrange(z, "b c h w -> b h w c").contiguous()
z_flattened = z.view(-1, self.e_dim)
# distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
d = (
torch.sum(z_flattened**2, dim=1, keepdim=True)
+ torch.sum(self.embedding.weight**2, dim=1)
- 2 * torch.einsum("bd,dn->bn", z_flattened, rearrange(self.embedding.weight, "n d -> d n"))
)
min_encoding_indices = torch.argmin(d, dim=1)
z_q = self.embedding(min_encoding_indices).view(z.shape)
perplexity = None
min_encodings = None
# compute loss for embedding
if not self.legacy:
loss = self.beta * torch.mean((z_q.detach() - z) ** 2) + torch.mean((z_q - z.detach()) ** 2)
else:
loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean((z_q - z.detach()) ** 2)
# preserve gradients
z_q = z + (z_q - z).detach()
# reshape back to match original input shape
z_q = rearrange(z_q, "b h w c -> b c h w").contiguous()
if self.remap is not None:
min_encoding_indices = min_encoding_indices.reshape(z.shape[0], -1) # add batch axis
min_encoding_indices = self.remap_to_used(min_encoding_indices)
min_encoding_indices = min_encoding_indices.reshape(-1, 1) # flatten
if self.sane_index_shape:
min_encoding_indices = min_encoding_indices.reshape(z_q.shape[0], z_q.shape[2], z_q.shape[3])
return z_q, loss, (perplexity, min_encodings, min_encoding_indices)
def get_codebook_entry(self, indices, shape):
# shape specifying (batch, height, width, channel)
if self.remap is not None:
indices = indices.reshape(shape[0], -1) # add batch axis
indices = self.unmap_to_all(indices)
indices = indices.reshape(-1) # flatten again
# get quantized latent vectors
z_q = self.embedding(indices)
if shape is not None:
z_q = z_q.view(shape)
# reshape back to match original input shape
z_q = z_q.permute(0, 3, 1, 2).contiguous()
return z_q
class GumbelQuantize(nn.Module):
"""
credit to @karpathy: https://github.com/karpathy/deep-vector-quantization/blob/main/model.py (thanks!)
Gumbel Softmax trick quantizer
Categorical Reparameterization with Gumbel-Softmax, Jang et al. 2016
https://arxiv.org/abs/1611.01144
"""
def __init__(
self,
num_hiddens,
embedding_dim,
n_embed,
straight_through=True,
kl_weight=5e-4,
temp_init=1.0,
use_vqinterface=True,
remap=None,
unknown_index="random",
):
super().__init__()
self.embedding_dim = embedding_dim
self.n_embed = n_embed
self.straight_through = straight_through
self.temperature = temp_init
self.kl_weight = kl_weight
self.proj = nn.Conv2d(num_hiddens, n_embed, 1)
self.embed = nn.Embedding(n_embed, embedding_dim)
self.use_vqinterface = use_vqinterface
self.remap = remap
if self.remap is not None:
self.register_buffer("used", torch.tensor(np.load(self.remap)))
self.re_embed = self.used.shape[0]
self.unknown_index = unknown_index # "random" or "extra" or integer
if self.unknown_index == "extra":
self.unknown_index = self.re_embed
self.re_embed = self.re_embed + 1
print(
f"Remapping {self.n_embed} indices to {self.re_embed} indices. "
f"Using {self.unknown_index} for unknown indices."
)
else:
self.re_embed = n_embed
def remap_to_used(self, inds):
ishape = inds.shape
assert len(ishape) > 1
inds = inds.reshape(ishape[0], -1)
used = self.used.to(inds)
match = (inds[:, :, None] == used[None, None, ...]).long()
new = match.argmax(-1)
unknown = match.sum(2) < 1
if self.unknown_index == "random":
new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to(device=new.device)
else:
new[unknown] = self.unknown_index
return new.reshape(ishape)
def unmap_to_all(self, inds):
ishape = inds.shape
assert len(ishape) > 1
inds = inds.reshape(ishape[0], -1)
used = self.used.to(inds)
if self.re_embed > self.used.shape[0]: # extra token
inds[inds >= self.used.shape[0]] = 0 # simply set to zero
back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
return back.reshape(ishape)
def forward(self, z, temp=None, return_logits=False):
# force hard = True when we are in eval mode, as we must quantize. actually, always true seems to work
hard = self.straight_through if self.training else True
temp = self.temperature if temp is None else temp
logits = self.proj(z)
if self.remap is not None:
# continue only with used logits
full_zeros = torch.zeros_like(logits)
logits = logits[:, self.used, ...]
soft_one_hot = F.gumbel_softmax(logits, tau=temp, dim=1, hard=hard)
if self.remap is not None:
# go back to all entries but unused set to zero
full_zeros[:, self.used, ...] = soft_one_hot
soft_one_hot = full_zeros
z_q = einsum("b n h w, n d -> b d h w", soft_one_hot, self.embed.weight)
# + kl divergence to the prior loss
qy = F.softmax(logits, dim=1)
diff = self.kl_weight * torch.sum(qy * torch.log(qy * self.n_embed + 1e-10), dim=1).mean()
ind = soft_one_hot.argmax(dim=1)
if self.remap is not None:
ind = self.remap_to_used(ind)
if self.use_vqinterface:
if return_logits:
return z_q, diff, (None, None, ind), logits
return z_q, diff, (None, None, ind)
return z_q, diff, ind
def get_codebook_entry(self, indices, shape):
b, h, w, c = shape
assert b * h * w == indices.shape[0]
indices = rearrange(indices, "(b h w) -> b h w", b=b, h=h, w=w)
if self.remap is not None:
indices = self.unmap_to_all(indices)
one_hot = F.one_hot(indices, num_classes=self.n_embed).permute(0, 3, 1, 2).float()
z_q = einsum("b n h w, n d -> b d h w", one_hot, self.embed.weight)
return z_q
|