EasyCV/easycv/models/utils/norm.py

114 lines
3.7 KiB
Python

# Copyright (c) Alibaba, Inc. and its affiliates.
import torch
import torch.nn as nn
from easycv.framework.errors import KeyError, NotImplementedError
class SyncIBN(nn.Module):
r"""Instance-Batch Normalization layer from
`"Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net"
<https://arxiv.org/pdf/1807.09441.pdf>`
Args:
planes (int): Number of channels for the input tensor
ratio (float): Ratio of instance normalization in the IBN layer
"""
def __init__(self, planes, ratio=0.5, eps=1e-5):
super(SyncIBN, self).__init__()
self.half = int(planes * ratio)
self.IN = nn.InstanceNorm2d(self.half, affine=True)
self.BN = nn.SyncBatchNorm(planes - self.half, eps=eps)
def forward(self, x):
split = torch.split(x, self.half, 1)
out1 = self.IN(split[0].contiguous())
out2 = self.BN(split[1].contiguous())
out = torch.cat((out1, out2), 1)
return out
class IBN(nn.Module):
r"""Instance-Batch Normalization layer from
`"Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net"
<https://arxiv.org/pdf/1807.09441.pdf>`
Args:
planes (int): Number of channels for the input tensor
ratio (float): Ratio of instance normalization in the IBN layer
"""
def __init__(self, planes, ratio=0.5, eps=1e-5):
super(IBN, self).__init__()
self.half = int(planes * ratio)
self.IN = nn.InstanceNorm2d(self.half, affine=True)
self.BN = nn.BatchNorm2d(planes - self.half, eps=eps)
def forward(self, x):
split = torch.split(x, self.half, 1)
out1 = self.IN(split[0].contiguous())
out2 = self.BN(split[1].contiguous())
out = torch.cat((out1, out2), 1)
return out
norm_cfg = {
# format: layer_type: (abbreviation, module)
'BN': ('bn', nn.BatchNorm2d),
'SyncBN': ('bn', nn.SyncBatchNorm),
'GN': ('gn', nn.GroupNorm),
# and potentially 'SN'
'IBN': ('ibn', IBN),
'SyncIBN': ('ibn', SyncIBN),
'IN': ('in', nn.InstanceNorm2d),
'LN': ('ln', nn.LayerNorm)
}
def build_norm_layer(cfg, num_features, postfix=''):
""" Build normalization layer
Args:
cfg (dict): cfg should contain:
type (str): identify norm layer type.
layer args: args needed to instantiate a norm layer.
requires_grad (bool): [optional] whether stop gradient updates
num_features (int): number of channels from input.
postfix (int, str): appended into norm abbreviation to
create named layer.
Returns:
name (str): abbreviation + postfix
layer (nn.Module): created norm layer
"""
assert isinstance(cfg, dict) and 'type' in cfg
cfg_ = cfg.copy()
layer_type = cfg_.pop('type')
if layer_type not in norm_cfg:
raise KeyError('Unrecognized norm type {}'.format(layer_type))
else:
abbr, norm_layer = norm_cfg[layer_type]
if norm_layer is None:
raise NotImplementedError
assert isinstance(postfix, (int, str))
name = abbr + str(postfix)
requires_grad = cfg_.pop('requires_grad', True)
cfg_.setdefault('eps', 1e-5)
if layer_type != 'GN':
layer = norm_layer(num_features, **cfg_)
if layer_type == 'SyncBN' and hasattr(layer, '_specify_ddp_gpu_num'):
layer._specify_ddp_gpu_num(1)
elif layer_type == 'SyncIBN' and hasattr(layer,
'_specify_ddp_gpu_num'):
layer.BN._specify_ddp_gpu_num(1)
else:
assert 'num_groups' in cfg_
layer = norm_layer(num_channels=num_features, **cfg_)
for param in layer.parameters():
param.requires_grad = requires_grad
return name, layer