[转]12 个常见 CNN 模型论文集锦与 PyTorch 实现

最近发现了一份不错的源代码，作者使用 PyTorch 实现了如今主流的卷积神经网络 CNN 框架，包含了 12 中模型架构。所有代码使用的数据集是 CIFAR。

项目地址：

https://github.com/BIGBALLON/CIFAR-ZOO

CNN 经典论文

该项目实现的是主流的 CNN 模型，涉及的论文包括：

1. CNN 模型（12 篇）

• (lenet) LeNet-5, convolutional neural networks
  论文地址：http://yann.lecun.com/exdb/lenet/
• (alexnet) ImageNet Classification with Deep Convolutional Neural Networks
  论文地址：https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks
• (vgg) Very Deep Convolutional Networks for Large-Scale Image Recognition
  论文地址：https://arxiv.org/abs/1409.1556
• (resnet) Deep Residual Learning for Image Recognition
  论文地址：https://arxiv.org/abs/1512.03385
• (preresnet) Identity Mappings in Deep Residual Networks
  论文地址：https://arxiv.org/abs/1603.05027
• (resnext) Aggregated Residual Transformations for Deep Neural Networks
  论文地址：https://arxiv.org/abs/1611.05431
• (densenet) Densely Connected Convolutional Networks
  论文地址：https://arxiv.org/abs/1608.06993
• (senet) Squeeze-and-Excitation Networks
  论文地址：https://arxiv.org/abs/1709.01507
• (bam) BAM: Bottleneck Attention Module
  论文地址：https://arxiv.org/abs/1807.06514
• (cbam) CBAM: Convolutional Block Attention Module
  论文地址：https://arxiv.org/abs/1807.06521
• (genet) Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks
  论文地址：https://arxiv.org/abs/1810.12348
• (sknet) SKNet: Selective Kernel Networks
  论文地址：https://arxiv.org/abs/1903.06586

2. 正则化（3 篇）

• (shake-shake) Shake-Shake regularization
  论文地址：https://arxiv.org/abs/1705.07485
• (cutout) Improved Regularization of Convolutional Neural Networks with Cutout
  论文地址：https://arxiv.org/abs/1708.04552
• (mixup) mixup: Beyond Empirical Risk Minimization
  论文地址：https://arxiv.org/abs/1710.09412

3. 学习速率调度器（2 篇）

• (cos_lr) SGDR: Stochastic Gradient Descent with Warm Restarts
  论文地址：https://arxiv.org/abs/1608.03983
• (htd_lr) Stochastic Gradient Descent with Hyperbolic-Tangent Decay on Classification
  论文地址：https://arxiv.org/abs/1806.01593

需求和使用

1. 需求

运行所有代码的开发环境需求为：

• Python >= 3.5
• PyTorch >= 0.4
• TensorFlow/Tensorboard
• 其它依赖项 (pyyaml, easydict, tensorboardX)

作者提供了一键安装、配置开发环境的方法：

pip install -r requirements.txt

2. 模型代码

作者将所有的模型都存放在 model 文件夹下，我们来看一下 PyTorch 实现的 ResNet 网络结构：

# -*-coding:utf-8-*-
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
__all__ = ['resnet20', 'resnet32', 'resnet44',
'resnet56', 'resnet110', 'resnet1202']
def conv3x3(in_planes, out_planes, stride=1):
"3x3 convolution with padding"
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=1, bias=False)
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, inplanes, planes, stride=1, downsample=None):
super(BasicBlock, self).__init__()
self.conv_1 = conv3x3(inplanes, planes, stride)
self.bn_1 = nn.BatchNorm2d(planes)
self.relu = nn.ReLU(inplace=True)
self.conv_2 = conv3x3(planes, planes)
self.bn_2 = nn.BatchNorm2d(planes)
self.downsample = downsample
self.stride = stride
def forward(self, x):
residual = x
out = self.conv_1(x)
out = self.bn_1(out)
out = self.relu(out)
out = self.conv_2(out)
out = self.bn_2(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.relu(out)
return out
class Bottleneck(nn.Module):
expansion = 4
def __init__(self, inplanes, planes, stride=1, downsample=None):
super(Bottleneck, self).__init__()
self.conv_1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
self.bn_1 = nn.BatchNorm2d(planes)
self.conv_2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride,
padding=1, bias=False)
self.bn_2 = nn.BatchNorm2d(planes)
self.conv_3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
self.bn_3 = nn.BatchNorm2d(planes * 4)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
def forward(self, x):
residual = x
out = self.conv_1(x)
out = self.bn_1(out)
out = self.relu(out)
out = self.conv_2(out)
out = self.bn_2(out)
out = self.relu(out)
out = self.conv_3(out)
out = self.bn_3(out)
if self.downsample is not None:
residual = self.downsample(x)
out += residual
out = self.relu(out)
return out
class ResNet(nn.Module):
def __init__(self, depth, num_classes, block_name='BasicBlock'):
super(ResNet, self).__init__()
# Model type specifies number of layers for CIFAR-10 model
if block_name == 'BasicBlock':
assert (
depth - 2) % 6 == 0, 'depth should be 6n+2, e.g. 20, 32, 44, 56, 110, 1202'
n = (depth - 2) // 6
block = BasicBlock
elif block_name == 'Bottleneck':
assert (
depth - 2) % 9 == 0, 'depth should be 9n+2, e.g. 20, 29, 47, 56, 110, 1199'
n = (depth - 2) // 9
block = Bottleneck
else:
raise ValueError('block_name shoule be Basicblock or Bottleneck')
self.inplanes = 16
self.conv_1 = nn.Conv2d(3, 16, kernel_size=3, padding=1,
bias=False)
self.bn_1 = nn.BatchNorm2d(16)
self.relu = nn.ReLU(inplace=True)
self.stage_1 = self._make_layer(block, 16, n)
self.stage_2 = self._make_layer(block, 32, n, stride=2)
self.stage_3 = self._make_layer(block, 64, n, stride=2)
self.avgpool = nn.AvgPool2d(8)
self.fc = nn.Linear(64 * block.expansion, num_classes)
for m in self.modules():
if isinstance(m, nn.Conv2d):
# nn.init.xavier_normal(m.weight.data)
nn.init.kaiming_normal_(m.weight.data)
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def _make_layer(self, block, planes, blocks, stride=1):
downsample = None
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(self.inplanes, planes * block.expansion,
kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample))
self.inplanes = planes * block.expansion
for i in range(1, blocks):
layers.append(block(self.inplanes, planes))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv_1(x)
x = self.bn_1(x)
x = self.relu(x) # 32x32
x = self.stage_1(x) # 32x32
x = self.stage_2(x) # 16x16
x = self.stage_3(x) # 8x8
x = self.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x
def resnet20(num_classes):
return ResNet(depth=20, num_classes=num_classes)
def resnet32(num_classes):
return ResNet(depth=32, num_classes=num_classes)
def resnet44(num_classes):
return ResNet(depth=44, num_classes=num_classes)
def resnet56(num_classes):
return ResNet(depth=56, num_classes=num_classes)
def resnet110(num_classes):
return ResNet(depth=110, num_classes=num_classes)
def resnet1202(num_classes):
return ResNet(depth=1202, num_classes=num_classes)

其它模型也一并能找到。

3. 使用

简单运行下面的命令就可以运行程序了：

## 1 GPU for lenet
CUDA_VISIBLE_DEVICES=0 python -u train.py --work-path ./experiments/cifar10/lenet
## resume from ckpt
CUDA_VISIBLE_DEVICES=0 python -u train.py --work-path ./experiments/cifar10/lenet --resume
## 2 GPUs for resnet1202
CUDA_VISIBLE_DEVICES=0,1 python -u train.py --work-path ./experiments/cifar10/preresnet1202
## 4 GPUs for densenet190bc
CUDA_VISIBLE_DEVICES=0,1,2,3 python -u train.py --work-path ./experiments/cifar10/densenet190bc

我们使用 yaml 文件 config.yaml 保存参数，查看 ./experimets 中的任何文件以了解更多详细信息。您可以通过 tensorboard 中 tensorboard —logdir path-to-event —port your-port 查看训练曲线。培训日志将通过日志转储，请检查您工作路径中的 log.txt。

模型在 CIFAR 数据集上的结果

1. 12 种 CNN 模型：

2. 正则化

默认的数据扩充方法是 RandomCrop+RandomHorizontalLip+Normalize，而 √ 表示采用哪种附加方法。

PS：Shake_Resnet26_2X64d 通过剪切和混合达到 97.71% 的测试精度！很酷，对吧？

3. 不同的学习速率调度器

最后，再附上项目地址：

[https://github.com/BIGBALLON/CIFAR-ZOO]https://link.zhihu.com/?target=https%3A//github.com/BIGBALLON/CIFAR-ZOO)
https://zhuanlan.zhihu.com/p/64693337