File size: 13,862 Bytes
786f6a6 |
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 |
""" EvoNorm in PyTorch
Based on `Evolving Normalization-Activation Layers` - https://arxiv.org/abs/2004.02967
@inproceedings{NEURIPS2020,
author = {Liu, Hanxiao and Brock, Andy and Simonyan, Karen and Le, Quoc},
booktitle = {Advances in Neural Information Processing Systems},
editor = {H. Larochelle and M. Ranzato and R. Hadsell and M. F. Balcan and H. Lin},
pages = {13539--13550},
publisher = {Curran Associates, Inc.},
title = {Evolving Normalization-Activation Layers},
url = {https://proceedings.neurips.cc/paper/2020/file/9d4c03631b8b0c85ae08bf05eda37d0f-Paper.pdf},
volume = {33},
year = {2020}
}
An attempt at getting decent performing EvoNorms running in PyTorch.
While faster than other PyTorch impl, still quite a ways off the built-in BatchNorm
in terms of memory usage and throughput on GPUs.
I'm testing these modules on TPU w/ PyTorch XLA. Promising start but
currently working around some issues with builtin torch/tensor.var/std. Unlike
GPU, similar train speeds for EvoNormS variants and BatchNorm.
Hacked together by / Copyright 2020 Ross Wightman
"""
from typing import Sequence, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from .create_act import create_act_layer
from .trace_utils import _assert
def instance_std(x, eps: float = 1e-5):
std = x.float().var(dim=(2, 3), unbiased=False, keepdim=True).add(eps).sqrt().to(x.dtype)
return std.expand(x.shape)
def instance_std_tpu(x, eps: float = 1e-5):
std = manual_var(x, dim=(2, 3)).add(eps).sqrt()
return std.expand(x.shape)
# instance_std = instance_std_tpu
def instance_rms(x, eps: float = 1e-5):
rms = x.float().square().mean(dim=(2, 3), keepdim=True).add(eps).sqrt().to(x.dtype)
return rms.expand(x.shape)
def manual_var(x, dim: Union[int, Sequence[int]], diff_sqm: bool = False):
xm = x.mean(dim=dim, keepdim=True)
if diff_sqm:
# difference of squared mean and mean squared, faster on TPU can be less stable
var = ((x * x).mean(dim=dim, keepdim=True) - (xm * xm)).clamp(0)
else:
var = ((x - xm) * (x - xm)).mean(dim=dim, keepdim=True)
return var
def group_std(x, groups: int = 32, eps: float = 1e-5, flatten: bool = False):
B, C, H, W = x.shape
x_dtype = x.dtype
_assert(C % groups == 0, '')
if flatten:
x = x.reshape(B, groups, -1) # FIXME simpler shape causing TPU / XLA issues
std = x.float().var(dim=2, unbiased=False, keepdim=True).add(eps).sqrt().to(x_dtype)
else:
x = x.reshape(B, groups, C // groups, H, W)
std = x.float().var(dim=(2, 3, 4), unbiased=False, keepdim=True).add(eps).sqrt().to(x_dtype)
return std.expand(x.shape).reshape(B, C, H, W)
def group_std_tpu(x, groups: int = 32, eps: float = 1e-5, diff_sqm: bool = False, flatten: bool = False):
# This is a workaround for some stability / odd behaviour of .var and .std
# running on PyTorch XLA w/ TPUs. These manual var impl are producing much better results
B, C, H, W = x.shape
_assert(C % groups == 0, '')
if flatten:
x = x.reshape(B, groups, -1) # FIXME simpler shape causing TPU / XLA issues
var = manual_var(x, dim=-1, diff_sqm=diff_sqm)
else:
x = x.reshape(B, groups, C // groups, H, W)
var = manual_var(x, dim=(2, 3, 4), diff_sqm=diff_sqm)
return var.add(eps).sqrt().expand(x.shape).reshape(B, C, H, W)
#group_std = group_std_tpu # FIXME TPU temporary
def group_rms(x, groups: int = 32, eps: float = 1e-5):
B, C, H, W = x.shape
_assert(C % groups == 0, '')
x_dtype = x.dtype
x = x.reshape(B, groups, C // groups, H, W)
rms = x.float().square().mean(dim=(2, 3, 4), keepdim=True).add(eps).sqrt_().to(x_dtype)
return rms.expand(x.shape).reshape(B, C, H, W)
class EvoNorm2dB0(nn.Module):
def __init__(self, num_features, apply_act=True, momentum=0.1, eps=1e-3, **_):
super().__init__()
self.apply_act = apply_act # apply activation (non-linearity)
self.momentum = momentum
self.eps = eps
self.weight = nn.Parameter(torch.ones(num_features))
self.bias = nn.Parameter(torch.zeros(num_features))
self.v = nn.Parameter(torch.ones(num_features)) if apply_act else None
self.register_buffer('running_var', torch.ones(num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.ones_(self.weight)
nn.init.zeros_(self.bias)
if self.v is not None:
nn.init.ones_(self.v)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
if self.v is not None:
if self.training:
var = x.float().var(dim=(0, 2, 3), unbiased=False)
# var = manual_var(x, dim=(0, 2, 3)).squeeze()
n = x.numel() / x.shape[1]
self.running_var.copy_(
self.running_var * (1 - self.momentum) +
var.detach() * self.momentum * (n / (n - 1)))
else:
var = self.running_var
left = var.add(self.eps).sqrt_().to(x_dtype).view(v_shape).expand_as(x)
v = self.v.to(x_dtype).view(v_shape)
right = x * v + instance_std(x, self.eps)
x = x / left.max(right)
return x * self.weight.to(x_dtype).view(v_shape) + self.bias.to(x_dtype).view(v_shape)
class EvoNorm2dB1(nn.Module):
def __init__(self, num_features, apply_act=True, momentum=0.1, eps=1e-5, **_):
super().__init__()
self.apply_act = apply_act # apply activation (non-linearity)
self.momentum = momentum
self.eps = eps
self.weight = nn.Parameter(torch.ones(num_features))
self.bias = nn.Parameter(torch.zeros(num_features))
self.register_buffer('running_var', torch.ones(num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.ones_(self.weight)
nn.init.zeros_(self.bias)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
if self.apply_act:
if self.training:
var = x.float().var(dim=(0, 2, 3), unbiased=False)
n = x.numel() / x.shape[1]
self.running_var.copy_(
self.running_var * (1 - self.momentum) +
var.detach().to(self.running_var.dtype) * self.momentum * (n / (n - 1)))
else:
var = self.running_var
var = var.to(x_dtype).view(v_shape)
left = var.add(self.eps).sqrt_()
right = (x + 1) * instance_rms(x, self.eps)
x = x / left.max(right)
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
class EvoNorm2dB2(nn.Module):
def __init__(self, num_features, apply_act=True, momentum=0.1, eps=1e-5, **_):
super().__init__()
self.apply_act = apply_act # apply activation (non-linearity)
self.momentum = momentum
self.eps = eps
self.weight = nn.Parameter(torch.ones(num_features))
self.bias = nn.Parameter(torch.zeros(num_features))
self.register_buffer('running_var', torch.ones(num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.ones_(self.weight)
nn.init.zeros_(self.bias)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
if self.apply_act:
if self.training:
var = x.float().var(dim=(0, 2, 3), unbiased=False)
n = x.numel() / x.shape[1]
self.running_var.copy_(
self.running_var * (1 - self.momentum) +
var.detach().to(self.running_var.dtype) * self.momentum * (n / (n - 1)))
else:
var = self.running_var
var = var.to(x_dtype).view(v_shape)
left = var.add(self.eps).sqrt_()
right = instance_rms(x, self.eps) - x
x = x / left.max(right)
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
class EvoNorm2dS0(nn.Module):
def __init__(self, num_features, groups=32, group_size=None, apply_act=True, eps=1e-5, **_):
super().__init__()
self.apply_act = apply_act # apply activation (non-linearity)
if group_size:
assert num_features % group_size == 0
self.groups = num_features // group_size
else:
self.groups = groups
self.eps = eps
self.weight = nn.Parameter(torch.ones(num_features))
self.bias = nn.Parameter(torch.zeros(num_features))
self.v = nn.Parameter(torch.ones(num_features)) if apply_act else None
self.reset_parameters()
def reset_parameters(self):
nn.init.ones_(self.weight)
nn.init.zeros_(self.bias)
if self.v is not None:
nn.init.ones_(self.v)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
if self.v is not None:
v = self.v.view(v_shape).to(x_dtype)
x = x * (x * v).sigmoid() / group_std(x, self.groups, self.eps)
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
class EvoNorm2dS0a(EvoNorm2dS0):
def __init__(self, num_features, groups=32, group_size=None, apply_act=True, eps=1e-3, **_):
super().__init__(
num_features, groups=groups, group_size=group_size, apply_act=apply_act, eps=eps)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
d = group_std(x, self.groups, self.eps)
if self.v is not None:
v = self.v.view(v_shape).to(x_dtype)
x = x * (x * v).sigmoid()
x = x / d
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
class EvoNorm2dS1(nn.Module):
def __init__(
self, num_features, groups=32, group_size=None,
apply_act=True, act_layer=None, eps=1e-5, **_):
super().__init__()
act_layer = act_layer or nn.SiLU
self.apply_act = apply_act # apply activation (non-linearity)
if act_layer is not None and apply_act:
self.act = create_act_layer(act_layer)
else:
self.act = nn.Identity()
if group_size:
assert num_features % group_size == 0
self.groups = num_features // group_size
else:
self.groups = groups
self.eps = eps
self.pre_act_norm = False
self.weight = nn.Parameter(torch.ones(num_features))
self.bias = nn.Parameter(torch.zeros(num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.ones_(self.weight)
nn.init.zeros_(self.bias)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
if self.apply_act:
x = self.act(x) / group_std(x, self.groups, self.eps)
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
class EvoNorm2dS1a(EvoNorm2dS1):
def __init__(
self, num_features, groups=32, group_size=None,
apply_act=True, act_layer=None, eps=1e-3, **_):
super().__init__(
num_features, groups=groups, group_size=group_size, apply_act=apply_act, act_layer=act_layer, eps=eps)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
x = self.act(x) / group_std(x, self.groups, self.eps)
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
class EvoNorm2dS2(nn.Module):
def __init__(
self, num_features, groups=32, group_size=None,
apply_act=True, act_layer=None, eps=1e-5, **_):
super().__init__()
act_layer = act_layer or nn.SiLU
self.apply_act = apply_act # apply activation (non-linearity)
if act_layer is not None and apply_act:
self.act = create_act_layer(act_layer)
else:
self.act = nn.Identity()
if group_size:
assert num_features % group_size == 0
self.groups = num_features // group_size
else:
self.groups = groups
self.eps = eps
self.weight = nn.Parameter(torch.ones(num_features))
self.bias = nn.Parameter(torch.zeros(num_features))
self.reset_parameters()
def reset_parameters(self):
nn.init.ones_(self.weight)
nn.init.zeros_(self.bias)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
if self.apply_act:
x = self.act(x) / group_rms(x, self.groups, self.eps)
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
class EvoNorm2dS2a(EvoNorm2dS2):
def __init__(
self, num_features, groups=32, group_size=None,
apply_act=True, act_layer=None, eps=1e-3, **_):
super().__init__(
num_features, groups=groups, group_size=group_size, apply_act=apply_act, act_layer=act_layer, eps=eps)
def forward(self, x):
_assert(x.dim() == 4, 'expected 4D input')
x_dtype = x.dtype
v_shape = (1, -1, 1, 1)
x = self.act(x) / group_rms(x, self.groups, self.eps)
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|