Ross Wightman
3 years ago
commit
cad170e494
@ -1,82 +1,332 @@
|
||||
"""EvoNormB0 (Batched) and EvoNormS0 (Sample) in PyTorch
|
||||
""" 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 currently faster than other impl, still quite a ways off the built-in BN
|
||||
in terms of memory usage and throughput (roughly 5x mem, 1/2 - 1/3x speed).
|
||||
While faster than other PyTorch impl, still quite a ways off the built-in BatchNorm
|
||||
in terms of memory usage and throughput on GPUs.
|
||||
|
||||
Still very much a WIP, fiddling with buffer usage, in-place/jit optimizations, and layouts.
|
||||
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
|
||||
"""
|
||||
|
||||
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
|
||||
|
||||
|
||||
class EvoNormBatch2d(nn.Module):
|
||||
def __init__(self, num_features, apply_act=True, momentum=0.1, eps=1e-5, drop_block=None):
|
||||
super(EvoNormBatch2d, self).__init__()
|
||||
def instance_std(x, eps: float = 1e-5):
|
||||
rms = x.float().var(dim=(2, 3), unbiased=False, keepdim=True).add(eps).sqrt().to(x.dtype)
|
||||
return rms.expand(x.shape)
|
||||
|
||||
|
||||
def instance_rms(x, eps: float = 1e-5):
|
||||
rms = x.square().float().mean(dim=(2, 3), keepdim=True).add(eps).sqrt().to(dtype=x.dtype)
|
||||
return rms.expand(x.shape)
|
||||
|
||||
|
||||
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, '')
|
||||
# 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()
|
||||
x = x.reshape(B, groups, C // groups, H, W)
|
||||
std = x.float().var(dim=(2, 3, 4), unbiased=False, keepdim=True).add(eps).sqrt()
|
||||
return std.expand(x.shape).reshape(B, C, H, W).to(x_dtype)
|
||||
|
||||
|
||||
def group_std_tpu(x, groups: int = 32, eps: float = 1e-5, diff_sqm: 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, '')
|
||||
x_dtype = x.dtype
|
||||
x = x.float().reshape(B, groups, C // groups, H, W)
|
||||
xm = x.mean(dim=(2, 3, 4), keepdim=True)
|
||||
if diff_sqm:
|
||||
# difference of squared mean and mean squared, faster on TPU
|
||||
var = (x.square().mean(dim=(2, 3, 4), keepdim=True) - xm.square()).clamp(0)
|
||||
else:
|
||||
var = (x - xm).square().mean(dim=(2, 3, 4), keepdim=True)
|
||||
return var.add(eps).sqrt().expand(x.shape).reshape(B, C, H, W).to(x_dtype)
|
||||
# group_std = group_std_tpu # temporary, for TPU / PT XLA
|
||||
|
||||
|
||||
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)
|
||||
sqm = x.square().mean(dim=(2, 3, 4), keepdim=True).add(eps).sqrt_().to(dtype=x_dtype)
|
||||
return sqm.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-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), requires_grad=True)
|
||||
self.bias = nn.Parameter(torch.zeros(num_features), requires_grad=True)
|
||||
self.v = nn.Parameter(torch.ones(num_features), requires_grad=True) if apply_act else None
|
||||
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.apply_act:
|
||||
if self.v is not None:
|
||||
nn.init.ones_(self.v)
|
||||
|
||||
def forward(self, x):
|
||||
assert x.dim() == 4, 'expected 4D input'
|
||||
x_type = x.dtype
|
||||
running_var = self.running_var.view(1, -1, 1, 1)
|
||||
if self.training:
|
||||
var = x.var(dim=(0, 2, 3), unbiased=False, keepdim=True)
|
||||
n = x.numel() / x.shape[1]
|
||||
running_var = var.detach() * self.momentum * (n / (n - 1)) + running_var * (1 - self.momentum)
|
||||
self.running_var.copy_(running_var.view(self.running_var.shape))
|
||||
else:
|
||||
var = running_var
|
||||
|
||||
_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.to(dtype=x_type).reshape(1, -1, 1, 1)
|
||||
d = x * v + (x.var(dim=(2, 3), unbiased=False, keepdim=True) + self.eps).sqrt().to(dtype=x_type)
|
||||
d = d.max((var + self.eps).sqrt().to(dtype=x_type))
|
||||
x = x / d
|
||||
return x * self.weight.view(1, -1, 1, 1) + self.bias.view(1, -1, 1, 1)
|
||||
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() * 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 EvoNormSample2d(nn.Module):
|
||||
def __init__(self, num_features, apply_act=True, groups=32, eps=1e-5, drop_block=None):
|
||||
super(EvoNormSample2d, self).__init__()
|
||||
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.groups = groups
|
||||
self.momentum = momentum
|
||||
self.eps = eps
|
||||
self.weight = nn.Parameter(torch.ones(num_features), requires_grad=True)
|
||||
self.bias = nn.Parameter(torch.zeros(num_features), requires_grad=True)
|
||||
self.v = nn.Parameter(torch.ones(num_features), requires_grad=True) if apply_act else None
|
||||
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(dtype=self.running_var.dtype) * self.momentum * (n / (n - 1)))
|
||||
else:
|
||||
var = self.running_var
|
||||
var = var.to(dtype=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(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=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(dtype=self.running_var.dtype) * self.momentum * (n / (n - 1)))
|
||||
else:
|
||||
var = self.running_var
|
||||
var = var.to(dtype=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(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=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')
|
||||
B, C, H, W = x.shape
|
||||
_assert(C % self.groups == 0, '')
|
||||
x_dtype = x.dtype
|
||||
v_shape = (1, -1, 1, 1)
|
||||
if self.v is not None:
|
||||
n = x * (x * self.v.view(1, -1, 1, 1)).sigmoid()
|
||||
x = x.reshape(B, self.groups, -1)
|
||||
x = n.reshape(B, self.groups, -1) / (x.var(dim=-1, unbiased=False, keepdim=True) + self.eps).sqrt()
|
||||
x = x.reshape(B, C, H, W)
|
||||
return x * self.weight.view(1, -1, 1, 1) + self.bias.view(1, -1, 1, 1)
|
||||
v = self.v.view(v_shape).to(dtype=x_dtype)
|
||||
x = x * (x * v).sigmoid() / group_std(x, self.groups, self.eps)
|
||||
return x * self.weight.view(v_shape).to(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=x_dtype)
|
||||
|
||||
|
||||
class EvoNorm2dS0a(EvoNorm2dS0):
|
||||
def __init__(self, num_features, groups=32, group_size=None, apply_act=True, eps=1e-5, **_):
|
||||
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(dtype=x_dtype)
|
||||
x = x * (x * v).sigmoid_()
|
||||
x = x / d
|
||||
return x * self.weight.view(v_shape).to(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=x_dtype)
|
||||
|
||||
|
||||
class EvoNorm2dS1(nn.Module):
|
||||
def __init__(
|
||||
self, num_features, groups=32, group_size=None,
|
||||
apply_act=True, act_layer=nn.SiLU, eps=1e-5, **_):
|
||||
super().__init__()
|
||||
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(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=x_dtype)
|
||||
|
||||
|
||||
class EvoNorm2dS1a(EvoNorm2dS1):
|
||||
def __init__(
|
||||
self, num_features, groups=32, group_size=None,
|
||||
apply_act=True, act_layer=nn.SiLU, eps=1e-5, **_):
|
||||
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(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=x_dtype)
|
||||
|
||||
|
||||
class EvoNorm2dS2(nn.Module):
|
||||
def __init__(
|
||||
self, num_features, groups=32, group_size=None,
|
||||
apply_act=True, act_layer=nn.SiLU, eps=1e-5, **_):
|
||||
super().__init__()
|
||||
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(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=x_dtype)
|
||||
|
||||
|
||||
class EvoNorm2dS2a(EvoNorm2dS2):
|
||||
def __init__(
|
||||
self, num_features, groups=32, group_size=None,
|
||||
apply_act=True, act_layer=nn.SiLU, eps=1e-5, **_):
|
||||
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(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=x_dtype)
|
||||
|
||||
@ -0,0 +1,68 @@
|
||||
""" Filter Response Norm in PyTorch
|
||||
|
||||
Based on `Filter Response Normalization Layer` - https://arxiv.org/abs/1911.09737
|
||||
|
||||
Hacked together by / Copyright 2021 Ross Wightman
|
||||
"""
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
|
||||
from .create_act import create_act_layer
|
||||
from .trace_utils import _assert
|
||||
|
||||
|
||||
def inv_instance_rms(x, eps: float = 1e-5):
|
||||
rms = x.square().float().mean(dim=(2, 3), keepdim=True).add(eps).rsqrt().to(x.dtype)
|
||||
return rms.expand(x.shape)
|
||||
|
||||
|
||||
class FilterResponseNormTlu2d(nn.Module):
|
||||
def __init__(self, num_features, apply_act=True, eps=1e-5, rms=True, **_):
|
||||
super(FilterResponseNormTlu2d, self).__init__()
|
||||
self.apply_act = apply_act # apply activation (non-linearity)
|
||||
self.rms = rms
|
||||
self.eps = eps
|
||||
self.weight = nn.Parameter(torch.ones(num_features))
|
||||
self.bias = nn.Parameter(torch.zeros(num_features))
|
||||
self.tau = nn.Parameter(torch.zeros(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.tau is not None:
|
||||
nn.init.zeros_(self.tau)
|
||||
|
||||
def forward(self, x):
|
||||
_assert(x.dim() == 4, 'expected 4D input')
|
||||
x_dtype = x.dtype
|
||||
v_shape = (1, -1, 1, 1)
|
||||
x = x * inv_instance_rms(x, self.eps)
|
||||
x = x * self.weight.view(v_shape).to(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=x_dtype)
|
||||
return torch.maximum(x, self.tau.reshape(v_shape).to(dtype=x_dtype)) if self.tau is not None else x
|
||||
|
||||
|
||||
class FilterResponseNormAct2d(nn.Module):
|
||||
def __init__(self, num_features, apply_act=True, act_layer=nn.ReLU, inplace=None, rms=True, eps=1e-5, **_):
|
||||
super(FilterResponseNormAct2d, self).__init__()
|
||||
if act_layer is not None and apply_act:
|
||||
self.act = create_act_layer(act_layer, inplace=inplace)
|
||||
else:
|
||||
self.act = nn.Identity()
|
||||
self.rms = rms
|
||||
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)
|
||||
x = x * inv_instance_rms(x, self.eps)
|
||||
x = x * self.weight.view(v_shape).to(dtype=x_dtype) + self.bias.view(v_shape).to(dtype=x_dtype)
|
||||
return self.act(x)
|
||||
Loading…
Reference in new issue