You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
353 lines
14 KiB
353 lines
14 KiB
3 years ago
|
""" 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}
|
||
|
}
|
||
5 years ago
|
|
||
|
An attempt at getting decent performing EvoNorms running in PyTorch.
|
||
3 years ago
|
While faster than other PyTorch impl, still quite a ways off the built-in BatchNorm
|
||
|
in terms of memory usage and throughput on GPUs.
|
||
5 years ago
|
|
||
3 years ago
|
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.
|
||
5 years ago
|
|
||
4 years ago
|
Hacked together by / Copyright 2020 Ross Wightman
|
||
5 years ago
|
"""
|
||
3 years ago
|
from typing import Sequence, Union
|
||
5 years ago
|
|
||
|
import torch
|
||
|
import torch.nn as nn
|
||
3 years ago
|
import torch.nn.functional as F
|
||
5 years ago
|
|
||
3 years ago
|
from .create_act import create_act_layer
|
||
3 years ago
|
from .trace_utils import _assert
|
||
|
|
||
5 years ago
|
|
||
3 years ago
|
def instance_std(x, eps: float = 1e-5):
|
||
3 years ago
|
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
|
||
3 years ago
|
|
||
|
|
||
|
def instance_rms(x, eps: float = 1e-5):
|
||
3 years ago
|
rms = x.float().square().mean(dim=(2, 3), keepdim=True).add(eps).sqrt().to(x.dtype)
|
||
3 years ago
|
return rms.expand(x.shape)
|
||
|
|
||
|
|
||
3 years ago
|
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
|
||
|
|
||
|
|
||
3 years ago
|
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, '')
|
||
3 years ago
|
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)
|
||
3 years ago
|
|
||
|
|
||
3 years ago
|
def group_std_tpu(x, groups: int = 32, eps: float = 1e-5, diff_sqm: bool = False, flatten: bool = False):
|
||
3 years ago
|
# 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, '')
|
||
3 years ago
|
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)
|
||
3 years ago
|
else:
|
||
3 years ago
|
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
|
||
3 years ago
|
|
||
|
|
||
|
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)
|
||
3 years ago
|
rms = x.float().square().mean(dim=(2, 3, 4), keepdim=True).add(eps).sqrt_().to(x_dtype)
|
||
3 years ago
|
return rms.expand(x.shape).reshape(B, C, H, W)
|
||
3 years ago
|
|
||
|
|
||
|
class EvoNorm2dB0(nn.Module):
|
||
3 years ago
|
def __init__(self, num_features, apply_act=True, momentum=0.1, eps=1e-3, **_):
|
||
3 years ago
|
super().__init__()
|
||
5 years ago
|
self.apply_act = apply_act # apply activation (non-linearity)
|
||
5 years ago
|
self.momentum = momentum
|
||
|
self.eps = eps
|
||
3 years ago
|
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
|
||
3 years ago
|
self.register_buffer('running_var', torch.ones(num_features))
|
||
5 years ago
|
self.reset_parameters()
|
||
|
|
||
|
def reset_parameters(self):
|
||
|
nn.init.ones_(self.weight)
|
||
|
nn.init.zeros_(self.bias)
|
||
3 years ago
|
if self.v is not None:
|
||
5 years ago
|
nn.init.ones_(self.v)
|
||
|
|
||
|
def forward(self, x):
|
||
3 years ago
|
_assert(x.dim() == 4, 'expected 4D input')
|
||
3 years ago
|
x_dtype = x.dtype
|
||
|
v_shape = (1, -1, 1, 1)
|
||
3 years ago
|
if self.v is not None:
|
||
3 years ago
|
if self.training:
|
||
3 years ago
|
var = x.float().var(dim=(0, 2, 3), unbiased=False)
|
||
3 years ago
|
# var = manual_var(x, dim=(0, 2, 3)).squeeze()
|
||
3 years ago
|
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)
|
||
3 years ago
|
n = x.numel() / x.shape[1]
|
||
3 years ago
|
self.running_var.copy_(
|
||
|
self.running_var * (1 - self.momentum) +
|
||
3 years ago
|
var.detach().to(self.running_var.dtype) * self.momentum * (n / (n - 1)))
|
||
3 years ago
|
else:
|
||
3 years ago
|
var = self.running_var
|
||
3 years ago
|
var = var.to(x_dtype).view(v_shape)
|
||
3 years ago
|
left = var.add(self.eps).sqrt_()
|
||
|
right = (x + 1) * instance_rms(x, self.eps)
|
||
|
x = x / left.max(right)
|
||
3 years ago
|
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|
||
5 years ago
|
|
||
|
|
||
3 years ago
|
class EvoNorm2dB2(nn.Module):
|
||
|
def __init__(self, num_features, apply_act=True, momentum=0.1, eps=1e-5, **_):
|
||
|
super().__init__()
|
||
5 years ago
|
self.apply_act = apply_act # apply activation (non-linearity)
|
||
3 years ago
|
self.momentum = momentum
|
||
5 years ago
|
self.eps = eps
|
||
3 years ago
|
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))
|
||
5 years ago
|
self.reset_parameters()
|
||
|
|
||
|
def reset_parameters(self):
|
||
|
nn.init.ones_(self.weight)
|
||
|
nn.init.zeros_(self.bias)
|
||
3 years ago
|
|
||
|
def forward(self, x):
|
||
|
_assert(x.dim() == 4, 'expected 4D input')
|
||
|
x_dtype = x.dtype
|
||
|
v_shape = (1, -1, 1, 1)
|
||
5 years ago
|
if self.apply_act:
|
||
3 years ago
|
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) +
|
||
3 years ago
|
var.detach().to(self.running_var.dtype) * self.momentum * (n / (n - 1)))
|
||
3 years ago
|
else:
|
||
|
var = self.running_var
|
||
3 years ago
|
var = var.to(x_dtype).view(v_shape)
|
||
3 years ago
|
left = var.add(self.eps).sqrt_()
|
||
|
right = instance_rms(x, self.eps) - x
|
||
|
x = x / left.max(right)
|
||
3 years ago
|
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|
||
3 years ago
|
|
||
|
|
||
|
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:
|
||
5 years ago
|
nn.init.ones_(self.v)
|
||
|
|
||
|
def forward(self, x):
|
||
3 years ago
|
_assert(x.dim() == 4, 'expected 4D input')
|
||
3 years ago
|
x_dtype = x.dtype
|
||
|
v_shape = (1, -1, 1, 1)
|
||
|
if self.v is not None:
|
||
3 years ago
|
v = self.v.view(v_shape).to(x_dtype)
|
||
3 years ago
|
x = x * (x * v).sigmoid() / group_std(x, self.groups, self.eps)
|
||
3 years ago
|
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|
||
3 years ago
|
|
||
|
|
||
|
class EvoNorm2dS0a(EvoNorm2dS0):
|
||
3 years ago
|
def __init__(self, num_features, groups=32, group_size=None, apply_act=True, eps=1e-3, **_):
|
||
3 years ago
|
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)
|
||
3 years ago
|
if self.v is not None:
|
||
3 years ago
|
v = self.v.view(v_shape).to(x_dtype)
|
||
3 years ago
|
x = x * (x * v).sigmoid()
|
||
3 years ago
|
x = x / d
|
||
3 years ago
|
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|
||
3 years ago
|
|
||
|
|
||
|
class EvoNorm2dS1(nn.Module):
|
||
|
def __init__(
|
||
|
self, num_features, groups=32, group_size=None,
|
||
3 years ago
|
apply_act=True, act_layer=None, eps=1e-5, **_):
|
||
3 years ago
|
super().__init__()
|
||
3 years ago
|
act_layer = act_layer or nn.SiLU
|
||
3 years ago
|
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)
|
||
3 years ago
|
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|
||
3 years ago
|
|
||
|
|
||
|
class EvoNorm2dS1a(EvoNorm2dS1):
|
||
|
def __init__(
|
||
|
self, num_features, groups=32, group_size=None,
|
||
3 years ago
|
apply_act=True, act_layer=None, eps=1e-3, **_):
|
||
3 years ago
|
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)
|
||
3 years ago
|
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|
||
3 years ago
|
|
||
|
|
||
|
class EvoNorm2dS2(nn.Module):
|
||
|
def __init__(
|
||
|
self, num_features, groups=32, group_size=None,
|
||
3 years ago
|
apply_act=True, act_layer=None, eps=1e-5, **_):
|
||
3 years ago
|
super().__init__()
|
||
3 years ago
|
act_layer = act_layer or nn.SiLU
|
||
3 years ago
|
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)
|
||
3 years ago
|
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|
||
3 years ago
|
|
||
|
|
||
|
class EvoNorm2dS2a(EvoNorm2dS2):
|
||
|
def __init__(
|
||
|
self, num_features, groups=32, group_size=None,
|
||
3 years ago
|
apply_act=True, act_layer=None, eps=1e-3, **_):
|
||
3 years ago
|
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)
|
||
3 years ago
|
return x * self.weight.view(v_shape).to(x_dtype) + self.bias.view(v_shape).to(x_dtype)
|