File size: 8,718 Bytes
814c8cf |
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 |
import numpy as np
import torch as t
import torch.nn as nn
import torch.nn.functional as F
import jukebox.utils.dist_adapter as dist
class BottleneckBlock(nn.Module):
def __init__(self, k_bins, emb_width, mu):
super().__init__()
self.k_bins = k_bins
self.emb_width = emb_width
self.mu = mu
self.reset_k()
self.threshold = 1.0
def reset_k(self):
self.init = False
self.k_sum = None
self.k_elem = None
self.register_buffer('k', t.zeros(self.k_bins, self.emb_width).cuda())
def _tile(self, x):
d, ew = x.shape
if d < self.k_bins:
n_repeats = (self.k_bins + d - 1) // d
std = 0.01 / np.sqrt(ew)
x = x.repeat(n_repeats, 1)
x = x + t.randn_like(x) * std
return x
def init_k(self, x):
mu, emb_width, k_bins = self.mu, self.emb_width, self.k_bins
self.init = True
# init k_w using random vectors from x
y = self._tile(x)
_k_rand = y[t.randperm(y.shape[0])][:k_bins]
dist.broadcast(_k_rand, 0)
self.k = _k_rand
assert self.k.shape == (k_bins, emb_width)
self.k_sum = self.k
self.k_elem = t.ones(k_bins, device=self.k.device)
def restore_k(self, num_tokens=None, threshold=1.0):
mu, emb_width, k_bins = self.mu, self.emb_width, self.k_bins
self.init = True
assert self.k.shape == (k_bins, emb_width)
self.k_sum = self.k.clone()
self.k_elem = t.ones(k_bins, device=self.k.device)
if num_tokens is not None:
expected_usage = num_tokens / k_bins
self.k_elem.data.mul_(expected_usage)
self.k_sum.data.mul_(expected_usage)
self.threshold = threshold
def update_k(self, x, x_l):
mu, emb_width, k_bins = self.mu, self.emb_width, self.k_bins
with t.no_grad():
# Calculate new centres
x_l_onehot = t.zeros(k_bins, x.shape[0], device=x.device) # k_bins, N * L
x_l_onehot.scatter_(0, x_l.view(1, x.shape[0]), 1)
_k_sum = t.matmul(x_l_onehot, x) # k_bins, w
_k_elem = x_l_onehot.sum(dim=-1) # k_bins
y = self._tile(x)
_k_rand = y[t.randperm(y.shape[0])][:k_bins]
dist.broadcast(_k_rand, 0)
dist.all_reduce(_k_sum)
dist.all_reduce(_k_elem)
# Update centres
old_k = self.k
self.k_sum = mu * self.k_sum + (1. - mu) * _k_sum # w, k_bins
self.k_elem = mu * self.k_elem + (1. - mu) * _k_elem # k_bins
usage = (self.k_elem.view(k_bins, 1) >= self.threshold).float()
self.k = usage * (self.k_sum.view(k_bins, emb_width) / self.k_elem.view(k_bins, 1)) \
+ (1 - usage) * _k_rand
_k_prob = _k_elem / t.sum(_k_elem) # x_l_onehot.mean(dim=-1) # prob of each bin
entropy = -t.sum(_k_prob * t.log(_k_prob + 1e-8)) # entropy ie how diverse
used_curr = (_k_elem >= self.threshold).sum()
usage = t.sum(usage)
dk = t.norm(self.k - old_k) / np.sqrt(np.prod(old_k.shape))
return dict(entropy=entropy,
used_curr=used_curr,
usage=usage,
dk=dk)
def preprocess(self, x):
# NCT -> NTC -> [NT, C]
x = x.permute(0, 2, 1).contiguous()
x = x.view(-1, x.shape[-1]) # x_en = (N * L, w), k_j = (w, k_bins)
if x.shape[-1] == self.emb_width:
prenorm = t.norm(x - t.mean(x)) / np.sqrt(np.prod(x.shape))
elif x.shape[-1] == 2 * self.emb_width:
x1, x2 = x[...,:self.emb_width], x[...,self.emb_width:]
prenorm = (t.norm(x1 - t.mean(x1)) / np.sqrt(np.prod(x1.shape))) + (t.norm(x2 - t.mean(x2)) / np.sqrt(np.prod(x2.shape)))
# Normalise
x = x1 + x2
else:
assert False, f"Expected {x.shape[-1]} to be (1 or 2) * {self.emb_width}"
return x, prenorm
def postprocess(self, x_l, x_d, x_shape):
# [NT, C] -> NTC -> NCT
N, T = x_shape
x_d = x_d.view(N, T, -1).permute(0, 2, 1).contiguous()
x_l = x_l.view(N, T)
return x_l, x_d
def quantise(self, x):
# Calculate latent code x_l
k_w = self.k.t()
distance = t.sum(x ** 2, dim=-1, keepdim=True) - 2 * t.matmul(x, k_w) + t.sum(k_w ** 2, dim=0,
keepdim=True) # (N * L, b)
min_distance, x_l = t.min(distance, dim=-1)
fit = t.mean(min_distance)
return x_l, fit
def dequantise(self, x_l):
x = F.embedding(x_l, self.k)
return x
def encode(self, x):
N, width, T = x.shape
# Preprocess.
x, prenorm = self.preprocess(x)
# Quantise
x_l, fit = self.quantise(x)
# Postprocess.
x_l = x_l.view(N, T)
return x_l
def decode(self, x_l):
N, T = x_l.shape
width = self.emb_width
# Dequantise
x_d = self.dequantise(x_l)
# Postprocess
x_d = x_d.view(N, T, width).permute(0, 2, 1).contiguous()
return x_d
def forward(self, x, update_k=True):
N, width, T = x.shape
# Preprocess
x, prenorm = self.preprocess(x)
# Init k if not inited
if update_k and not self.init:
self.init_k(x)
# Quantise and dequantise through bottleneck
x_l, fit = self.quantise(x)
x_d = self.dequantise(x_l)
# Update embeddings
if update_k:
update_metrics = self.update_k(x, x_l)
else:
update_metrics = {}
# Loss
commit_loss = t.norm(x_d.detach() - x) ** 2 / np.prod(x.shape)
# Passthrough
x_d = x + (x_d - x).detach()
# Postprocess
x_l, x_d = self.postprocess(x_l, x_d, (N,T))
return x_l, x_d, commit_loss, dict(fit=fit,
pn=prenorm,
**update_metrics)
class Bottleneck(nn.Module):
def __init__(self, l_bins, emb_width, mu, levels):
super().__init__()
self.levels = levels
level_block = lambda level: BottleneckBlock(l_bins, emb_width, mu)
self.level_blocks = nn.ModuleList()
for level in range(self.levels):
self.level_blocks.append(level_block(level))
def encode(self, xs):
zs = [level_block.encode(x) for (level_block, x) in zip(self.level_blocks, xs)]
return zs
def decode(self, zs, start_level=0, end_level=None):
if end_level is None:
end_level = self.levels
xs_quantised = [level_block.decode(z) for (level_block, z) in zip(self.level_blocks[start_level:end_level], zs)]
return xs_quantised
def forward(self, xs):
zs, xs_quantised, commit_losses, metrics = [], [], [], []
for level in range(self.levels):
level_block = self.level_blocks[level]
x = xs[level]
z, x_quantised, commit_loss, metric = level_block(x, update_k=self.training)
zs.append(z)
if not self.training:
# Be extra paranoid and make sure the encoder weights can't
# change from straight-through estimator
x_quantised = x_quantised.detach()
xs_quantised.append(x_quantised)
commit_losses.append(commit_loss)
if self.training:
metrics.append(metric)
return zs, xs_quantised, commit_losses, metrics
class NoBottleneckBlock(nn.Module):
def restore_k(self):
pass
class NoBottleneck(nn.Module):
def __init__(self, levels):
super().__init__()
self.level_blocks = nn.ModuleList()
self.levels = levels
for level in range(levels):
self.level_blocks.append(NoBottleneckBlock())
def encode(self, xs):
return xs
def decode(self, zs, start_level=0, end_level=None):
if end_level is None:
end_level = self.levels
return zs
def forward(self, xs):
zero = t.zeros(()).cuda()
commit_losses = [zero for _ in range(self.levels)]
metrics = [dict(entropy=zero, usage=zero, used_curr=zero, pn=zero, dk=zero) for _ in range(self.levels)]
return xs, xs, commit_losses, metrics
if __name__ == '__main__':
from jukebox.utils.dist_utils import setup_dist_from_mpi
rank, local_rank, device = setup_dist_from_mpi(port=29600)
bottleneck = Bottleneck(256, 64, 0.99, 2).to(device)
bottleneck.check()
|