mirror of https://github.com/alibaba/EasyCV.git
591 lines
19 KiB
Python
591 lines
19 KiB
Python
# Copyright (c) Alibaba, Inc. and its affiliates.
|
|
import torch.nn as nn
|
|
import torch.utils.checkpoint as cp
|
|
from mmcv.cnn import constant_init, kaiming_init
|
|
from torch.nn.modules.batchnorm import _BatchNorm
|
|
|
|
from ..modelzoo import resnet as model_urls
|
|
from ..registry import BACKBONES
|
|
from ..utils import FReLU, build_conv_layer, build_norm_layer
|
|
|
|
|
|
class BasicBlock(nn.Module):
|
|
expansion = 1
|
|
|
|
def __init__(self,
|
|
inplanes,
|
|
planes,
|
|
stride=1,
|
|
dilation=1,
|
|
downsample=None,
|
|
style='pytorch',
|
|
with_cp=False,
|
|
conv_cfg=None,
|
|
norm_cfg=dict(type='BN'),
|
|
frelu=False):
|
|
super(BasicBlock, self).__init__()
|
|
|
|
self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
|
|
self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
|
|
|
|
self.conv1 = build_conv_layer(
|
|
conv_cfg,
|
|
inplanes,
|
|
planes,
|
|
3,
|
|
stride=stride,
|
|
padding=dilation,
|
|
dilation=dilation,
|
|
bias=False)
|
|
self.add_module(self.norm1_name, norm1)
|
|
self.conv2 = build_conv_layer(
|
|
conv_cfg, planes, planes, 3, padding=1, bias=False)
|
|
self.add_module(self.norm2_name, norm2)
|
|
|
|
self.frelu = frelu
|
|
if frelu:
|
|
self.frelu_a = FReLU(planes)
|
|
self.frelu_b = FReLU(planes)
|
|
else:
|
|
self.relu = nn.ReLU(inplace=True)
|
|
self.downsample = downsample
|
|
self.stride = stride
|
|
self.dilation = dilation
|
|
assert not with_cp
|
|
|
|
@property
|
|
def norm1(self):
|
|
return getattr(self, self.norm1_name)
|
|
|
|
@property
|
|
def norm2(self):
|
|
return getattr(self, self.norm2_name)
|
|
|
|
def forward(self, x):
|
|
identity = x
|
|
|
|
out = self.conv1(x)
|
|
out = self.norm1(out)
|
|
if self.frelu:
|
|
out = self.frelu_a(out)
|
|
else:
|
|
out = self.relu(out)
|
|
|
|
out = self.conv2(out)
|
|
out = self.norm2(out)
|
|
|
|
if self.downsample is not None:
|
|
identity = self.downsample(x)
|
|
|
|
out += identity
|
|
if self.frelu:
|
|
out = self.frelu_b(out)
|
|
else:
|
|
out = self.relu(out)
|
|
return out
|
|
|
|
|
|
class Bottleneck(nn.Module):
|
|
expansion = 4
|
|
|
|
def __init__(self,
|
|
inplanes,
|
|
planes,
|
|
stride=1,
|
|
dilation=1,
|
|
downsample=None,
|
|
style='pytorch',
|
|
with_cp=False,
|
|
conv_cfg=None,
|
|
norm_cfg=dict(type='BN'),
|
|
frelu=False):
|
|
"""Bottleneck block for ResNet.
|
|
If style is "pytorch", the stride-two layer is the 3x3 conv layer,
|
|
if it is "caffe", the stride-two layer is the first 1x1 conv layer.
|
|
"""
|
|
super(Bottleneck, self).__init__()
|
|
assert style in ['pytorch', 'caffe']
|
|
|
|
self.inplanes = inplanes
|
|
self.planes = planes
|
|
self.stride = stride
|
|
self.dilation = dilation
|
|
self.style = style
|
|
self.with_cp = with_cp
|
|
self.conv_cfg = conv_cfg
|
|
self.norm_cfg = norm_cfg
|
|
|
|
if self.style == 'pytorch':
|
|
self.conv1_stride = 1
|
|
self.conv2_stride = stride
|
|
else:
|
|
self.conv1_stride = stride
|
|
self.conv2_stride = 1
|
|
|
|
self.norm1_name, norm1 = build_norm_layer(norm_cfg, planes, postfix=1)
|
|
self.norm2_name, norm2 = build_norm_layer(norm_cfg, planes, postfix=2)
|
|
self.norm3_name, norm3 = build_norm_layer(
|
|
norm_cfg, planes * self.expansion, postfix=3)
|
|
|
|
self.conv1 = build_conv_layer(
|
|
conv_cfg,
|
|
inplanes,
|
|
planes,
|
|
kernel_size=1,
|
|
stride=self.conv1_stride,
|
|
bias=False)
|
|
self.add_module(self.norm1_name, norm1)
|
|
self.conv2 = build_conv_layer(
|
|
conv_cfg,
|
|
planes,
|
|
planes,
|
|
kernel_size=3,
|
|
stride=self.conv2_stride,
|
|
padding=dilation,
|
|
dilation=dilation,
|
|
bias=False)
|
|
self.add_module(self.norm2_name, norm2)
|
|
self.conv3 = build_conv_layer(
|
|
conv_cfg,
|
|
planes,
|
|
planes * self.expansion,
|
|
kernel_size=1,
|
|
bias=False)
|
|
self.add_module(self.norm3_name, norm3)
|
|
|
|
self.frelu = frelu
|
|
if self.frelu:
|
|
self.relu_a = FReLU(planes)
|
|
self.relu_b = FReLU(planes)
|
|
self.relu_c = FReLU(planes * self.expansion)
|
|
else:
|
|
self.relu = nn.ReLU(inplace=True)
|
|
self.downsample = downsample
|
|
|
|
@property
|
|
def norm1(self):
|
|
return getattr(self, self.norm1_name)
|
|
|
|
@property
|
|
def norm2(self):
|
|
return getattr(self, self.norm2_name)
|
|
|
|
@property
|
|
def norm3(self):
|
|
return getattr(self, self.norm3_name)
|
|
|
|
def forward(self, x):
|
|
|
|
def _inner_forward(x):
|
|
identity = x
|
|
|
|
out = self.conv1(x)
|
|
out = self.norm1(out)
|
|
if self.frelu:
|
|
out = self.relu_a(out)
|
|
else:
|
|
out = self.relu(out)
|
|
|
|
out = self.conv2(out)
|
|
out = self.norm2(out)
|
|
if self.frelu:
|
|
out = self.relu_b(out)
|
|
else:
|
|
out = self.relu(out)
|
|
out = self.conv3(out)
|
|
out = self.norm3(out)
|
|
|
|
if self.downsample is not None:
|
|
identity = self.downsample(x)
|
|
|
|
out += identity
|
|
|
|
return out
|
|
|
|
if self.with_cp and x.requires_grad:
|
|
out = cp.checkpoint(_inner_forward, x)
|
|
else:
|
|
out = _inner_forward(x)
|
|
|
|
if self.frelu:
|
|
out = self.relu_c(out)
|
|
else:
|
|
out = self.relu(out)
|
|
|
|
return out
|
|
|
|
|
|
def make_res_layer(
|
|
block,
|
|
inplanes,
|
|
planes,
|
|
blocks,
|
|
stride=1,
|
|
dilation=1,
|
|
style='pytorch',
|
|
avg_down=False,
|
|
with_cp=False,
|
|
conv_cfg=None,
|
|
norm_cfg=dict(type='BN'),
|
|
frelu=False,
|
|
multi_grid=None,
|
|
contract_dilation=False,
|
|
):
|
|
downsample = None
|
|
if stride != 1 or inplanes != planes * block.expansion:
|
|
downsample = []
|
|
conv_stride = stride
|
|
if avg_down:
|
|
conv_stride = 1
|
|
downsample.append(
|
|
nn.AvgPool2d(
|
|
kernel_size=stride,
|
|
stride=stride,
|
|
ceil_mode=True,
|
|
count_include_pad=False))
|
|
downsample.extend([
|
|
build_conv_layer(
|
|
conv_cfg,
|
|
inplanes,
|
|
planes * block.expansion,
|
|
kernel_size=1,
|
|
stride=conv_stride,
|
|
bias=False),
|
|
build_norm_layer(norm_cfg, planes * block.expansion)[1]
|
|
])
|
|
downsample = nn.Sequential(*downsample)
|
|
|
|
if multi_grid is None:
|
|
if dilation > 1 and contract_dilation:
|
|
first_dilation = dilation // 2
|
|
else:
|
|
first_dilation = dilation
|
|
else:
|
|
first_dilation = multi_grid[0]
|
|
layers = []
|
|
layers.append(
|
|
block(
|
|
inplanes=inplanes,
|
|
planes=planes,
|
|
stride=stride,
|
|
dilation=first_dilation,
|
|
downsample=downsample,
|
|
style=style,
|
|
with_cp=with_cp,
|
|
conv_cfg=conv_cfg,
|
|
norm_cfg=norm_cfg,
|
|
frelu=frelu))
|
|
inplanes = planes * block.expansion
|
|
for i in range(1, blocks):
|
|
layers.append(
|
|
block(
|
|
inplanes=inplanes,
|
|
planes=planes,
|
|
stride=1,
|
|
dilation=dilation if multi_grid is None else multi_grid[i],
|
|
style=style,
|
|
with_cp=with_cp,
|
|
conv_cfg=conv_cfg,
|
|
norm_cfg=norm_cfg,
|
|
frelu=frelu))
|
|
|
|
return nn.Sequential(*layers)
|
|
|
|
|
|
@BACKBONES.register_module
|
|
class ResNet(nn.Module):
|
|
"""ResNet backbone.
|
|
|
|
Args:
|
|
depth (int): Depth of resnet, from {18, 34, 50, 101, 152}.
|
|
in_channels (int): Number of input image channels. Normally 3.
|
|
num_stages (int): Resnet stages, normally 4.
|
|
strides (Sequence[int]): Strides of the first block of each stage.
|
|
dilations (Sequence[int]): Dilation of each stage.
|
|
out_indices (Sequence[int]): Output from which stages.
|
|
style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
|
|
layer is the 3x3 conv layer, otherwise the stride-two layer is
|
|
the first 1x1 conv layer.
|
|
deep_stem (bool): Replace 7x7 conv in input stem with 3 3x3 conv.
|
|
Default: False.
|
|
avg_down (bool): Use AvgPool instead of stride conv when
|
|
downsampling in the bottleneck. Default: False.
|
|
frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
|
|
-1 means not freezing any parameters.
|
|
norm_cfg (dict): dictionary to construct and config norm layer.
|
|
norm_eval (bool): Whether to set norm layers to eval mode, namely,
|
|
freeze running stats (mean and var). Note: Effect on Batch Norm
|
|
and its variants only.
|
|
with_cp (bool): Use checkpoint or not. Using checkpoint will save some
|
|
memory while slowing down the training speed.
|
|
original_inplanes: start channel for first block, default=64
|
|
stem_channels (int): Number of stem channels. Default: 64.
|
|
zero_init_residual (bool): whether to use zero init for last norm layer
|
|
in resblocks to let them behave as identity.
|
|
multi_grid (Sequence[int]|None): Multi grid dilation rates of last
|
|
stage. Default: None.
|
|
contract_dilation (bool): Whether contract first dilation of each layer
|
|
Default: False.
|
|
|
|
Example:
|
|
>>> from easycv.models import ResNet
|
|
>>> import torch
|
|
>>> self = ResNet(depth=18)
|
|
>>> self.eval()
|
|
>>> inputs = torch.rand(1, 3, 32, 32)
|
|
>>> level_outputs = self.forward(inputs)
|
|
>>> for level_out in level_outputs:
|
|
... print(tuple(level_out.shape))
|
|
(1, 64, 8, 8)
|
|
(1, 128, 4, 4)
|
|
(1, 256, 2, 2)
|
|
(1, 512, 1, 1)
|
|
"""
|
|
|
|
arch_settings = {
|
|
10: (BasicBlock, (1, 1, 1, 1)),
|
|
18: (BasicBlock, (2, 2, 2, 2)),
|
|
34: (BasicBlock, (3, 4, 6, 3)),
|
|
50: (Bottleneck, (3, 4, 6, 3)),
|
|
101: (Bottleneck, (3, 4, 23, 3)),
|
|
152: (Bottleneck, (3, 8, 36, 3))
|
|
}
|
|
|
|
def __init__(self,
|
|
depth,
|
|
in_channels=3,
|
|
num_stages=4,
|
|
strides=(1, 2, 2, 2),
|
|
dilations=(1, 1, 1, 1),
|
|
out_indices=(0, 1, 2, 3, 4),
|
|
style='pytorch',
|
|
deep_stem=False,
|
|
avg_down=False,
|
|
num_classes=0,
|
|
frozen_stages=-1,
|
|
conv_cfg=None,
|
|
norm_cfg=dict(type='BN', requires_grad=True),
|
|
norm_eval=False,
|
|
with_cp=False,
|
|
frelu=False,
|
|
original_inplanes=64,
|
|
stem_channels=64,
|
|
zero_init_residual=False,
|
|
multi_grid=None,
|
|
contract_dilation=False):
|
|
super(ResNet, self).__init__()
|
|
if depth not in self.arch_settings:
|
|
raise KeyError('invalid depth {} for resnet'.format(depth))
|
|
self.depth = depth
|
|
self.num_stages = num_stages
|
|
assert num_stages >= 1 and num_stages <= 4
|
|
self.strides = strides
|
|
self.dilations = dilations
|
|
assert len(strides) == len(dilations) == num_stages
|
|
self.out_indices = out_indices
|
|
assert max(out_indices) < num_stages + 1
|
|
self.style = style
|
|
self.deep_stem = deep_stem
|
|
self.avg_down = avg_down
|
|
self.frozen_stages = frozen_stages
|
|
self.conv_cfg = conv_cfg
|
|
self.norm_cfg = norm_cfg
|
|
self.with_cp = with_cp
|
|
self.norm_eval = norm_eval
|
|
self.zero_init_residual = zero_init_residual
|
|
self.block, stage_blocks = self.arch_settings[depth]
|
|
self.stage_blocks = stage_blocks[:num_stages]
|
|
self.original_inplanes = original_inplanes
|
|
self.stem_channels = stem_channels
|
|
self.inplanes = stem_channels
|
|
self.frelu = frelu
|
|
self.multi_grid = multi_grid
|
|
self.contract_dilation = contract_dilation
|
|
|
|
self._make_stem_layer(in_channels, stem_channels)
|
|
|
|
self.res_layers = []
|
|
for i, num_blocks in enumerate(self.stage_blocks):
|
|
stride = strides[i]
|
|
dilation = dilations[i]
|
|
# multi grid is applied to last layer only
|
|
stage_multi_grid = multi_grid if i == len(
|
|
self.stage_blocks) - 1 else None
|
|
planes = self.original_inplanes * 2**i
|
|
res_layer = make_res_layer(
|
|
self.block,
|
|
self.inplanes,
|
|
planes,
|
|
num_blocks,
|
|
stride=stride,
|
|
dilation=dilation,
|
|
style=self.style,
|
|
avg_down=self.avg_down,
|
|
with_cp=with_cp,
|
|
conv_cfg=conv_cfg,
|
|
norm_cfg=norm_cfg,
|
|
frelu=self.frelu,
|
|
multi_grid=stage_multi_grid,
|
|
contract_dilation=contract_dilation,
|
|
)
|
|
self.inplanes = planes * self.block.expansion
|
|
layer_name = 'layer{}'.format(i + 1)
|
|
self.add_module(layer_name, res_layer)
|
|
self.res_layers.append(layer_name)
|
|
|
|
self._freeze_stages()
|
|
|
|
self.feat_dim = self.block.expansion * self.original_inplanes * 2**(
|
|
len(self.stage_blocks) - 1)
|
|
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
|
|
if num_classes > 0:
|
|
self.fc = nn.Linear(self.feat_dim, num_classes)
|
|
|
|
self.default_pretrained_model_path = model_urls.get(
|
|
self.__class__.__name__ + str(depth), None)
|
|
|
|
@property
|
|
def norm1(self):
|
|
return getattr(self, self.norm1_name)
|
|
|
|
def _make_stem_layer(self, in_channels, stem_channels):
|
|
if self.frelu:
|
|
relu = FReLU(stem_channels)
|
|
else:
|
|
relu = nn.ReLU(inplace=True)
|
|
|
|
if self.deep_stem:
|
|
self.stem = nn.Sequential(
|
|
build_conv_layer(
|
|
self.conv_cfg,
|
|
in_channels,
|
|
stem_channels // 2,
|
|
kernel_size=3,
|
|
stride=2,
|
|
padding=1,
|
|
bias=False),
|
|
build_norm_layer(self.norm_cfg, stem_channels // 2)[1], relu,
|
|
build_conv_layer(
|
|
self.conv_cfg,
|
|
stem_channels // 2,
|
|
stem_channels // 2,
|
|
kernel_size=3,
|
|
stride=1,
|
|
padding=1,
|
|
bias=False),
|
|
build_norm_layer(self.norm_cfg, stem_channels // 2)[1], relu,
|
|
build_conv_layer(
|
|
self.conv_cfg,
|
|
stem_channels // 2,
|
|
stem_channels,
|
|
kernel_size=3,
|
|
stride=1,
|
|
padding=1,
|
|
bias=False),
|
|
build_norm_layer(self.norm_cfg, stem_channels)[1], relu)
|
|
else:
|
|
self.conv1 = build_conv_layer(
|
|
self.conv_cfg,
|
|
in_channels,
|
|
stem_channels,
|
|
kernel_size=7,
|
|
stride=2,
|
|
padding=3,
|
|
bias=False)
|
|
self.norm1_name, norm1 = build_norm_layer(
|
|
self.norm_cfg, stem_channels, postfix=1)
|
|
self.add_module(self.norm1_name, norm1)
|
|
self.relu = relu
|
|
|
|
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
|
|
|
|
def _freeze_stages(self):
|
|
if self.frozen_stages >= 0:
|
|
if self.deep_stem:
|
|
self.stem.eval()
|
|
for param in self.stem.parameters():
|
|
param.requires_grad = False
|
|
else:
|
|
self.norm1.eval()
|
|
for m in [self.conv1, self.norm1]:
|
|
for param in m.parameters():
|
|
param.requires_grad = False
|
|
|
|
for i in range(1, self.frozen_stages + 1):
|
|
m = getattr(self, 'layer{}'.format(i))
|
|
m.eval()
|
|
for param in m.parameters():
|
|
param.requires_grad = False
|
|
|
|
def init_weights(self):
|
|
for m in self.modules():
|
|
if isinstance(m, nn.Conv2d):
|
|
kaiming_init(m, mode='fan_in', nonlinearity='relu')
|
|
elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
|
|
constant_init(m, 1)
|
|
|
|
if self.zero_init_residual:
|
|
for m in self.modules():
|
|
if isinstance(m, Bottleneck):
|
|
constant_init(m.norm3, 0)
|
|
elif isinstance(m, BasicBlock):
|
|
constant_init(m.norm2, 0)
|
|
|
|
def forward(self, x):
|
|
outs = []
|
|
if self.deep_stem:
|
|
x = self.stem(x)
|
|
else:
|
|
x = self.conv1(x)
|
|
x = self.norm1(x)
|
|
x = self.relu(x) # r50: 64x128x128
|
|
if 0 in self.out_indices:
|
|
outs.append(x)
|
|
x = self.maxpool(x) # r50: 64x56x56
|
|
for i, layer_name in enumerate(self.res_layers):
|
|
res_layer = getattr(self, layer_name)
|
|
x = res_layer(x)
|
|
if i + 1 in self.out_indices:
|
|
outs.append(x)
|
|
# r50: 1-256x56x56; 2-512x28x28; 3-1024x14x14; 4-2048x7x7
|
|
|
|
if hasattr(self, 'fc'):
|
|
bs = x.size(0)
|
|
x = self.avgpool(x).view(bs, -1)
|
|
x = self.fc(x)
|
|
# outs.append(x)
|
|
outs = [x]
|
|
|
|
return outs
|
|
|
|
def train(self, mode=True):
|
|
super(ResNet, self).train(mode)
|
|
self._freeze_stages()
|
|
if mode and self.norm_eval:
|
|
for m in self.modules():
|
|
# trick: eval have effect on BatchNorm only
|
|
if isinstance(m, _BatchNorm):
|
|
m.eval()
|
|
|
|
|
|
@BACKBONES.register_module()
|
|
class ResNetV1c(ResNet):
|
|
"""Compared to ResNet, ResNetV1c replaces the 7x7 conv in the input stem with three 3x3 convs.
|
|
For more details please refer to <https://arxiv.org/abs/1812.01187>.
|
|
"""
|
|
|
|
def __init__(self, **kwargs):
|
|
super(ResNetV1c, self).__init__(
|
|
deep_stem=True, avg_down=False, **kwargs)
|
|
|
|
|
|
@BACKBONES.register_module()
|
|
class ResNetV1d(ResNet):
|
|
"""Compared to ResNet, ResNetV1d replaces the 7x7 conv in the input stem with three 3x3 convs.
|
|
And in the downsampling block, a 2x2 avg_pool with stride 2 is added before conv, whose stride is changed to 1.
|
|
"""
|
|
|
|
def __init__(self, **kwargs):
|
|
super(ResNetV1d, self).__init__(
|
|
deep_stem=True, avg_down=True, **kwargs)
|