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.
84 lines
3.3 KiB
84 lines
3.3 KiB
"""EvoNormB0 (Batched) and EvoNormS0 (Sample) in PyTorch
|
|
|
|
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).
|
|
|
|
Still very much a WIP, fiddling with buffer usage, in-place/jit optimizations, and layouts.
|
|
|
|
Hacked together by / Copyright 2020 Ross Wightman
|
|
"""
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
|
|
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__()
|
|
self.apply_act = apply_act # apply activation (non-linearity)
|
|
self.momentum = momentum
|
|
self.eps = eps
|
|
param_shape = (1, num_features, 1, 1)
|
|
self.weight = nn.Parameter(torch.ones(param_shape), requires_grad=True)
|
|
self.bias = nn.Parameter(torch.zeros(param_shape), requires_grad=True)
|
|
if apply_act:
|
|
self.v = nn.Parameter(torch.ones(param_shape), requires_grad=True)
|
|
self.register_buffer('running_var', torch.ones(1, num_features, 1, 1))
|
|
self.reset_parameters()
|
|
|
|
def reset_parameters(self):
|
|
nn.init.ones_(self.weight)
|
|
nn.init.zeros_(self.bias)
|
|
if self.apply_act:
|
|
nn.init.ones_(self.v)
|
|
|
|
def forward(self, x):
|
|
assert x.dim() == 4, 'expected 4D input'
|
|
x_type = x.dtype
|
|
if self.training:
|
|
var = x.var(dim=(0, 2, 3), unbiased=False, keepdim=True)
|
|
n = x.numel() / x.shape[1]
|
|
self.running_var.copy_(
|
|
var.detach() * self.momentum * (n / (n - 1)) + self.running_var * (1 - self.momentum))
|
|
else:
|
|
var = self.running_var
|
|
|
|
if self.apply_act:
|
|
v = self.v.to(dtype=x_type)
|
|
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 + self.bias
|
|
|
|
|
|
class EvoNormSample2d(nn.Module):
|
|
def __init__(self, num_features, apply_act=True, groups=8, eps=1e-5, drop_block=None):
|
|
super(EvoNormSample2d, self).__init__()
|
|
self.apply_act = apply_act # apply activation (non-linearity)
|
|
self.groups = groups
|
|
self.eps = eps
|
|
param_shape = (1, num_features, 1, 1)
|
|
self.weight = nn.Parameter(torch.ones(param_shape), requires_grad=True)
|
|
self.bias = nn.Parameter(torch.zeros(param_shape), requires_grad=True)
|
|
if apply_act:
|
|
self.v = nn.Parameter(torch.ones(param_shape), requires_grad=True)
|
|
self.reset_parameters()
|
|
|
|
def reset_parameters(self):
|
|
nn.init.ones_(self.weight)
|
|
nn.init.zeros_(self.bias)
|
|
if self.apply_act:
|
|
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
|
|
if self.apply_act:
|
|
n = x * (x * self.v).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 + self.bias
|