paper: Densely Connected Convolutional Networks
Memory-Efficient Implementation of DenseNets
code: https://github.com/pytorch/vision/blob/main/torchvision/models/densenet.py
在本篇文章中，作者提出了Dense Convolutional Network (DenseNet)，下图是一个5层的dense block结构，可以看出每一层都作为后面所有层的输入。
DenseNet有以下几个优点：
- 缓解了梯度消失问题
- 加强了特征的传播
- 增强了特征重用
- 大幅减少了参数数量
作者指出ResNet的一个优势是梯度可以通过identity函数直接从后面的层传播到前面的层，但是identity函数是通过相加（summation）和卷积结果进行融合的，这可能会阻碍网络中的信息流动。因此作者通过拼接（concatenation）的方式在每一层与后面所有的层之间都建立连接，来进一步改善层之间的信息流动。
实现细节：
- 层之间是一个包含三个连续的运算的复合函数：BN – ReLU – Conv(3×3)
- 在DenseNet中复合函数用的是pre-activation即BN-ReLU-Conv，而不是常见的post-activation即Conv-BN-ReLU，作者在文中提到pre-activation对DenseNet的最终效果影响很大，换成post-activation在cifar10和cifar100数据集上都会掉点
- dense block之间的transition层负责下采样，具体包括：BN – Conv(1×1) – 2×2 average pooling
- 假设一个dense block的输入特征图通道数为 (k_{0})，其中每一层的输出通道数为 (k)，则第 (l) 层的输入通道数为 (k_{0}+k\times (l-1))。这里的超参 (k) 就是网络的 growth rate。比较反直觉的是DenseNet的参数较少，就是因为这里的 (k) 可以设置的较小，文章中设置 (k=32)，而不用像其它网络那样输出通道数达到1024、2048。
- 尽管每一层的输出通道数 (k) 比较小，但是因为输入包含了前面所有层因此输入的通道数很大，因此作者又提出了 DenseNet-B，在每个3×3卷积层前添加1×1卷积作为bottleneck layer来进行降维，进一步提高计算效率。这样每一层的复合函数就变成了BN – ReLU – Conv(1×1) – BN – ReLU – Conv(3×3)，文中1×1卷积的输出通道设置为 (4k)。
- 为了进一步减少参数，可以减少transition layer的输出通道数，如果dense block的输出通道为 (m)，设置transition layer的输出为 (\theta m)，对 (\theta < 1) 的DenseNet称为 DenseNet-C。文中设置 (\theta = 0.5)，对于含有bottleneck layer并且 (\theta < 1) 的DenseNet称为 DenseNet-BC。
- 文中对于ImageNet数据集，作者使用含有4个dense block的DenseNet-BC，在第一个dense block前是一个7×7 stride=2 输出通道为 (2k) 的卷积层，然后是一个3×3 stride=2的max pooling层。在最后一个dense block后是一个全局平均池化，然后是softmax分类器。
不同层数的DenseNet结构如下表所示
代码
这里的代码是torchvision的官方实现
import re
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as cp
from collections import OrderedDict
from .._internally_replaced_utils import load_state_dict_from_url
from torch import Tensor
from typing import Any, List, Tuple
__all__ = ['DenseNet', 'densenet121', 'densenet169', 'densenet201', 'densenet161']
model_urls = {
'densenet121': 'https://download.pytorch.org/models/densenet121-a639ec97.pth',
'densenet169': 'https://download.pytorch.org/models/densenet169-b2777c0a.pth',
'densenet201': 'https://download.pytorch.org/models/densenet201-c1103571.pth',
'densenet161': 'https://download.pytorch.org/models/densenet161-8d451a50.pth',
}
class _DenseLayer(nn.Module):
def __init__(
self,
num_input_features: int,
growth_rate: int,
bn_size: int,
drop_rate: float,
memory_efficient: bool = False
) -> None:
super(_DenseLayer, self).__init__()
self.norm1: nn.BatchNorm2d
self.add_module('norm1', nn.BatchNorm2d(num_input_features))
self.relu1: nn.ReLU
self.add_module('relu1', nn.ReLU(inplace=True))
self.conv1: nn.Conv2d
self.add_module('conv1', nn.Conv2d(num_input_features, bn_size *
growth_rate, kernel_size=1, stride=1,
bias=False))
self.norm2: nn.BatchNorm2d
self.add_module('norm2', nn.BatchNorm2d(bn_size * growth_rate))
self.relu2: nn.ReLU
self.add_module('relu2', nn.ReLU(inplace=True))
self.conv2: nn.Conv2d
self.add_module('conv2', nn.Conv2d(bn_size * growth_rate, growth_rate,
kernel_size=3, stride=1, padding=1,
bias=False))
self.drop_rate = float(drop_rate)
self.memory_efficient = memory_efficient
def bn_function(self, inputs: List[Tensor]) -> Tensor:
concated_features = torch.cat(inputs, 1)
bottleneck_output = self.conv1(self.relu1(self.norm1(concated_features))) # noqa: T484
return bottleneck_output
# todo: rewrite when torchscript supports any
def any_requires_grad(self, input: List[Tensor]) -> bool:
for tensor in input:
if tensor.requires_grad:
return True
return False
@torch.jit.unused # noqa: T484
def call_checkpoint_bottleneck(self, input: List[Tensor]) -> Tensor:
def closure(*inputs):
return self.bn_function(inputs)
return cp.checkpoint(closure, *input)
@torch.jit._overload_method # noqa: F811
def forward(self, input: List[Tensor]) -> Tensor:
pass
@torch.jit._overload_method # noqa: F811
def forward(self, input: Tensor) -> Tensor:
pass
# torchscript does not yet support *args, so we overload method
# allowing it to take either a List[Tensor] or single Tensor
def forward(self, input: Tensor) -> Tensor: # noqa: F811
if isinstance(input, Tensor):
prev_features = [input]
else:
prev_features = input
if self.memory_efficient and self.any_requires_grad(prev_features):
if torch.jit.is_scripting():
raise Exception("Memory Efficient not supported in JIT")
bottleneck_output = self.call_checkpoint_bottleneck(prev_features)
else:
bottleneck_output = self.bn_function(prev_features)
new_features = self.conv2(self.relu2(self.norm2(bottleneck_output)))
if self.drop_rate > 0:
new_features = F.dropout(new_features, p=self.drop_rate,
training=self.training)
return new_features
class _DenseBlock(nn.ModuleDict):
_version = 2
def __init__(
self,
num_layers: int,
num_input_features: int,
bn_size: int,
growth_rate: int,
drop_rate: float,
memory_efficient: bool = False
) -> None:
super(_DenseBlock, self).__init__()
for i in range(num_layers):
layer = _DenseLayer(
num_input_features + i * growth_rate,
growth_rate=growth_rate,
bn_size=bn_size,
drop_rate=drop_rate,
memory_efficient=memory_efficient,
)
self.add_module('denselayer%d' % (i + 1), layer)
def forward(self, init_features: Tensor) -> Tensor:
features = [init_features]
for name, layer in self.items():
new_features = layer(features)
features.append(new_features)
return torch.cat(features, 1)
class _Transition(nn.Sequential):
def __init__(self, num_input_features: int, num_output_features: int) -> None:
super(_Transition, self).__init__()
self.add_module('norm', nn.BatchNorm2d(num_input_features))
self.add_module('relu', nn.ReLU(inplace=True))
self.add_module('conv', nn.Conv2d(num_input_features, num_output_features,
kernel_size=1, stride=1, bias=False))
self.add_module('pool', nn.AvgPool2d(kernel_size=2, stride=2))
class DenseNet(nn.Module):
r"""Densenet-BC model class, based on
"Densely Connected Convolutional Networks" _.
Args:
growth_rate (int) - how many filters to add each layer (k in paper)
block_config (list of 4 ints) - how many layers in each pooling block
num_init_features (int) - the number of filters to learn in the first convolution layer
bn_size (int) - multiplicative factor for number of bottle neck layers
(i.e. bn_size * k features in the bottleneck layer)
drop_rate (float) - dropout rate after each dense layer
num_classes (int) - number of classification classes
memory_efficient (bool) - If True, uses checkpointing. Much more memory efficient,
but slower. Default: *False*. See "paper" _.
"""
def __init__(
self,
growth_rate: int = 32,
block_config: Tuple[int, int, int, int] = (6, 12, 24, 16),
num_init_features: int = 64,
bn_size: int = 4,
drop_rate: float = 0,
num_classes: int = 1000,
memory_efficient: bool = False
) -> None:
super(DenseNet, self).__init__()
# First convolution
self.features = nn.Sequential(OrderedDict([
('conv0', nn.Conv2d(3, num_init_features, kernel_size=7, stride=2,
padding=3, bias=False)),
('norm0', nn.BatchNorm2d(num_init_features)),
('relu0', nn.ReLU(inplace=True)),
('pool0', nn.MaxPool2d(kernel_size=3, stride=2, padding=1)),
]))
# Each denseblock
num_features = num_init_features
for i, num_layers in enumerate(block_config):
block = _DenseBlock(
num_layers=num_layers,
num_input_features=num_features,
bn_size=bn_size,
growth_rate=growth_rate,
drop_rate=drop_rate,
memory_efficient=memory_efficient
)
self.features.add_module('denseblock%d' % (i + 1), block)
num_features = num_features + num_layers * growth_rate
if i != len(block_config) - 1:
trans = _Transition(num_input_features=num_features,
num_output_features=num_features // 2)
self.features.add_module('transition%d' % (i + 1), trans)
num_features = num_features // 2
# Final batch norm
self.features.add_module('norm5', nn.BatchNorm2d(num_features))
# Linear layer
self.classifier = nn.Linear(num_features, num_classes)
# Official init from torch repo.
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Linear):
nn.init.constant_(m.bias, 0)
def forward(self, x: Tensor) -> Tensor:
features = self.features(x)
out = F.relu(features, inplace=True)
out = F.adaptive_avg_pool2d(out, (1, 1))
out = torch.flatten(out, 1)
out = self.classifier(out)
return out
def _load_state_dict(model: nn.Module, model_url: str, progress: bool) -> None:
# '.'s are no longer allowed in module names, but previous _DenseLayer
# has keys 'norm.1', 'relu.1', 'conv.1', 'norm.2', 'relu.2', 'conv.2'.
# They are also in the checkpoints in model_urls. This pattern is used
# to find such keys.
pattern = re.compile(
r'^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$')
state_dict = load_state_dict_from_url(model_url, progress=progress)
for key in list(state_dict.keys()):
res = pattern.match(key)
if res:
new_key = res.group(1) + res.group(2)
state_dict[new_key] = state_dict[key]
del state_dict[key]
model.load_state_dict(state_dict)
def _densenet(
arch: str,
growth_rate: int,
block_config: Tuple[int, int, int, int],
num_init_features: int,
pretrained: bool,
progress: bool,
**kwargs: Any
) -> DenseNet:
model = DenseNet(growth_rate, block_config, num_init_features, **kwargs)
if pretrained:
_load_state_dict(model, model_urls[arch], progress)
return model
def densenet121(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> DenseNet:
r"""Densenet-121 model from
"Densely Connected Convolutional Networks" _.
The required minimum input size of the model is 29x29.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
memory_efficient (bool) - If True, uses checkpointing. Much more memory efficient,
but slower. Default: *False*. See "paper" _.
"""
return _densenet('densenet121', 32, (6, 12, 24, 16), 64, pretrained, progress,
**kwargs)
def densenet161(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> DenseNet:
r"""Densenet-161 model from
"Densely Connected Convolutional Networks" _.
The required minimum input size of the model is 29x29.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
memory_efficient (bool) - If True, uses checkpointing. Much more memory efficient,
but slower. Default: *False*. See "paper" _.
"""
return _densenet('densenet161', 48, (6, 12, 36, 24), 96, pretrained, progress,
**kwargs)
def densenet169(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> DenseNet:
r"""Densenet-169 model from
"Densely Connected Convolutional Networks" _.
The required minimum input size of the model is 29x29.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
memory_efficient (bool) - If True, uses checkpointing. Much more memory efficient,
but slower. Default: *False*. See "paper" _.
"""
return _densenet('densenet169', 32, (6, 12, 32, 32), 64, pretrained, progress,
**kwargs)
def densenet201(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> DenseNet:
r"""Densenet-201 model from
"Densely Connected Convolutional Networks" _.
The required minimum input size of the model is 29x29.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
progress (bool): If True, displays a progress bar of the download to stderr
memory_efficient (bool) - If True, uses checkpointing. Much more memory efficient,
but slower. Default: *False*. See "paper" _.
"""
return _densenet('densenet201', 32, (6, 12, 48, 32), 64, pretrained, progress,
**kwargs)
Memory-Efficient Implementation
DenseNet虽然参数很少，但因为每一层的输入都要对之前所有层的输出进行拼接，在大多数框架如Tensorflow和Pytorch中，每进行一次拼接操作，都会开辟新的内存用于保存拼接后的结果，因此对于一个 (L) 层的DenseNet，这就要占用 (L(L+1)/2) 层的内存。对于一个普通的卷积网络如vgg、resnet，中间特征图的数量随着网络深度的加深呈线性增长，因此将这些特征图保存在内存中不会带来显著的内存负担。但对于一个有 (m) 层的dense block，保存中间特征图会导致 (O(m^2)) 的内存使用，这种指数级的增长会导致显存不够的情况。
解决方法：深度学习框架中之所以要保存中间结果，是因为反向传播过程中需要用到这些中间结果来计算梯度。具体解决方法是前向过程中不保存中间结果，而是保存输入和具体的运算函数（比如卷积、BN、池化等），这样在反向过程中需要用到前向的中间结果时，通过保存的输入和函数重新计算，这样就大大减小了显存的占用，但同时也导致了训练时间的增加，算是一种用时间换空间的方法。
在上面torchvision的实现中，通过 torch.utils.checkpoint.checkpoint来实现，并且只对复合函数BN – ReLU – Conv(1×1) – BN – ReLU – Conv(3×3)中的前半部分进行该操作，因为当 (growth_rate=k) 时，1×1卷积的输出通道数是 (4k)，3×3卷积的输出是 (k)，前半部分的显存占用更多。
Original: https://blog.csdn.net/ooooocj/article/details/124060354
Author: 00000cj
Title: DenseNet
原创文章受到原创版权保护。转载请注明出处：https://www.johngo689.com/663421/
转载文章受原作者版权保护。转载请注明原作者出处！