🍨 本文为[🔗365天深度学习训练营]内部限免文章(版权归 *K同学啊* 所有)
🍖 作者:[K同学啊]
前言
ResNeXt是由何凯明团队在2017年CVPR会议上提出来的新型图像分类网络。在论文《Aggregated Residual Transformations for Deep Neural Networks》作者提出了当时普遍存在的问题:如何提高模型的准确率?
常用的方法是提高网络的深度或宽度,但单纯的提高网络的深度或宽度,加大了设计的难度,也加大了计算的开销。由此何团队设计了cardinality的概念。将卷积通道分组,再对分组进行卷积。
对比ResNet跟ResNeXt
ResNeXt采用分组卷积的方式,将特征图分为不同的组,再对每组特征图进行卷积,在分组卷积中,每个卷积核只处理一部分的通道
一、分组卷积
我们在ResNet50的基础上进行修改,先设计组采样的Bottleneck模块
class Bottleneck(nn.Module): # 定义残差块,renet50、resnet101、resnet152使用此残差块
expansion = 4 # 残差操作维度变化倍数
def __init__(self, in_channel, out_channel, stride=1, downsample=None, groups=1, base_width=64): # 初始化方法
super(Bottleneck, self).__init__() # 继承初始化方法
width = int(in_channel * (base_width / 64.0)) * groups # F(x)第二个卷积的通道数
self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=width, kernel_size=1, stride=1,
bias=False) # conv操作
self.bn1 = nn.BatchNorm2d(num_features=width) # bn操作
self.conv2 = nn.Conv2d(in_channels=width, out_channels=width, groups=groups, kernel_size=3, stride=stride,
padding=1, bias=False) # conv操作,若为ResNeXt网络,则这里为group conv操作
self.bn2 = nn.BatchNorm2d(num_features=width) # bn操作
self.conv3 = nn.Conv2d(in_channels=width, out_channels=out_channel * self.expansion, kernel_size=1, stride=1,
bias=False) # conv操作
self.bn3 = nn.BatchNorm2d(num_features=out_channel * self.expansion) # bn操作
self.relu = nn.ReLU(inplace=True) # relu激活函数
self.downsample = downsample # 是否下采样
def forward(self, x): # 前传函数
identity = x # 原始x
if self.downsample: # 如果下采样
identity = self.downsample(x) # 残差边存在conv操作,x-->x'
x = self.conv1(x) # conv操作
x = self.bn1(x) # bn操作
x = self.relu(x) # relu激活函数
x = self.conv2(x) # conv操作
x = self.bn2(x) # bn操作
x = self.relu(x) # relu激活函数
x = self.conv3(x) # conv操作
x = self.bn3(x) # bn操作
x += identity # F(x)+x/x'
x = self.relu(x) # relu激活函数
return x
其中downsample是向下的组,每一层layer都向下采样一次
if stride != 1 or self.channel != channel * block.expansion: # 如果卷积步长不为1或卷积前后通道数不一致,则需要对原始x进行操作
downsample = nn.Sequential(
nn.Conv2d(in_channels=self.channel, out_channels=channel * block.expansion, kernel_size=1, # conv操作
stride=stride, bias=False),
nn.BatchNorm2d(num_features=channel * block.expansion) # bn操作
)
二、ResNeXt50_Model模型
class ResNeXt50_Model(nn.Module):
def __init__(self, in_channel=3, N_classes=1000):
super(ResNeXt50_Model, self).__init__()
self.in_channels = in_channel
self.layers = [2, 3, 5, 2]
# ============= 基础层
# 方法1
self.zeropadding2d = nn.ZeroPad2d(3)
self.cov0 = nn.Conv2d(self.in_channels, out_channels=64, kernel_size=7, stride=2, padding=3)
self.bn0 = nn.BatchNorm2d(num_features=64)
self.relu0 = nn.ReLU(inplace=False)
self.maxpool0 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.channel = 64
self.groups = 1
self.base_width = 64
self.layer1 = self._make_layer(Bottleneck, 64, self.layers[0]) # 第一块残差集合,由基本的残差块组成
self.layer2 = self._make_layer(Bottleneck, 128, self.layers[1], stride=2) # 第二块残差集合,由基本的残差块组成
self.layer3 = self._make_layer(Bottleneck, 256, self.layers[2], stride=2) # 第三块残差集合,由基本的残差块组成
self.layer4 = self._make_layer(Bottleneck, 512, self.layers[3], stride=2) # 第四块残差集合,由基本的残差块组成
# 输出网络
self.avgpool = nn.AvgPool2d((7, 7))
# classfication layer
# 7*7均值后2048个参数
self.fc = nn.Sequential(nn.Linear(2048, N_classes),
nn.Softmax(dim=1))
for m in self.modules(): # 遍历模型结构
if isinstance(m, nn.Conv2d): # 如果当前结构是卷积操作
nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu") # 使用kaiming初始化方法
def basic_layer1(self, x):
'''
input: x = tensor(3, 224, 224).unsqueeze(0)
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 112, 112] 9,408
BatchNorm2d-2 [-1, 64, 112, 112] 128
ReLU-3 [-1, 64, 112, 112] 0
MaxPool2d-4 [-1, 64, 56, 56] 0
================================================================
'''
x = self.zeropadding2d(x)
x = self.cov0(x)
x = self.bn0(x)
x = self.relu0(x)
x = self.maxpool0(x)
return x
def _make_layer(self, block, channel, blocks, stride=1): # 定义函数,用于生成模型结构
downsample = None # 默认不对原始x进行操作
if stride != 1 or self.channel != channel * block.expansion: # 如果卷积步长不为1或卷积前后通道数不一致,则需要对原始x进行操作
downsample = nn.Sequential(
nn.Conv2d(in_channels=self.channel, out_channels=channel * block.expansion, kernel_size=1, # conv操作
stride=stride, bias=False),
nn.BatchNorm2d(num_features=channel * block.expansion) # bn操作
)
layers = [] # 列表用于存放模型结构
layers.append(block(self.channel, channel, downsample=downsample, stride=stride, groups=self.groups,
base_width=self.base_width)) # 模型追加block结构
self.channel = channel * block.expansion # 通道数转换为卷积后输出通道数
for _ in range(1, blocks): # 进行blocks次循环
layers.append(block(self.channel, channel, groups=self.groups, base_width=self.base_width)) # 模型追加block结构
return nn.Sequential(*layers) # 返回模型结构
def forward(self, x):
x = self.basic_layer1(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
x = torch.flatten(x, 1)
x = self.fc(x)
return x
三、测试模型
设定epochs=100, 早停为early_stop=10
import torch
from torchvision import datasets, transforms
import torch.nn as nn
import time
import numpy as np
import matplotlib.pyplot as plt
import torch.nn.functional as F
import torchsummary as summary
import copy
import os
data_dir = './J3-data'
def random_split_imagefolder(data_dir, transforms, random_split_rate=0.8):
_total_data = datasets.ImageFolder(data_dir, transform=transforms)
train_size = int(random_split_rate * len(_total_data))
test_size = len(_total_data) - train_size
_train_datasets, _test_datasets = torch.utils.data.random_split(_total_data, [train_size, test_size])
return _total_data, _train_datasets, _test_datasets
N_classes=2
batch_size = 32
mean = [0.4958, 0.4984, 0.4068]
std = [0.2093, 0.2026, 0.2170]
# 真实均值-标准差重新读取数据
real_transforms = transforms.Compose(
[
transforms.Resize([224, 224]),#中心裁剪到224*224
transforms.ToTensor(),#转化成张量
transforms.Normalize(mean, std)
])
total_data, train_datasets, test_datasets = random_split_imagefolder(data_dir, real_transforms, 0.8)
# 批读取文件
train_data = torch.utils.data.DataLoader(train_datasets, batch_size=batch_size, shuffle=True, num_workers=8)
test_data = torch.utils.data.DataLoader(test_datasets, batch_size=batch_size, shuffle=True, num_workers=8)
train_data_size = len(train_datasets)
test_data_size = len(test_datasets)
def train_and_test(model, loss_func, optimizer, epochs=100, early_stop=10):
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
summary.summary(model, (3, 224, 224))
record = []
best_acc = 0.0
best_epoch = 0
stop_steps = 0
for epoch in range(epochs):#训练epochs轮
epoch_start = time.time()
print("Epoch: {}/{}".format(epoch + 1, epochs))
model.train()#训练
train_loss = 0.0
train_acc = 0.0
valid_loss = 0.0
valid_acc = 0.0
for i, (inputs, labels) in enumerate(train_data):
inputs = inputs.to(device)
labels = labels.to(device)
#print(labels)
# 记得清零
optimizer.zero_grad()
outputs = model(inputs)
loss = loss_func(outputs, labels)
loss.backward()
optimizer.step()
train_loss += loss.item() * inputs.size(0)
if i%10==0:
print("train data: {:01d} / {:03d} outputs: {}".format(i, len(train_data), outputs.data[0]))
ret, predictions = torch.max(outputs.data, 1)
correct_counts = predictions.eq(labels.data.view_as(predictions))
acc = torch.mean(correct_counts.type(torch.FloatTensor))
train_acc += acc.item() * inputs.size(0)
with torch.no_grad():
model.eval()#验证
for j, (inputs, labels) in enumerate(test_data):
inputs = inputs.to(device)
labels = labels.to(device)
outputs = model(inputs)
loss = loss_func(outputs, labels)
valid_loss += loss.item() * inputs.size(0)
if j%10==0:
print("val data: {:01d} / {:03d} outputs: {}".format(j, len(test_data), outputs.data[0]))
ret, predictions = torch.max(outputs.data, 1)
correct_counts = predictions.eq(labels.data.view_as(predictions))
acc = torch.mean(correct_counts.type(torch.FloatTensor))
valid_acc += acc.item() * inputs.size(0)
avg_train_loss = train_loss / train_data_size
avg_train_acc = train_acc / train_data_size
avg_valid_loss = valid_loss / test_data_size
avg_valid_acc = valid_acc / test_data_size
record.append([avg_train_loss, avg_valid_loss, avg_train_acc, avg_valid_acc])
if avg_valid_acc > best_acc :#记录最高准确性的模型
best_acc = avg_valid_acc
stop_steps = 0
best_epoch = epoch + 1
best_param = copy.deepcopy(model.state_dict())
else:
stop_steps += 1
if stop_steps >= early_stop:
break
epoch_end = time.time()
print("Epoch: {:03d}, Training: Loss: {:.4f}, Accuracy: {:.4f}%, \n\t\tValidation: Loss: {:.4f}, Accuracy: {:.4f}%, Time: {:.4f}s".format(
epoch + 1, avg_valid_loss, avg_train_acc * 100, avg_valid_loss, avg_valid_acc * 100,
epoch_end - epoch_start))
print("Best Accuracy for validation : {:.4f} at epoch {:03d}".format(best_acc, best_epoch))
model.load_state_dict(best_param)
return model, record
#%%
if __name__=='__main__':
early_stop=10
epochs = 100
model = ResNeXt50_Model(3, N_classes)
loss_func = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(),lr=0.0001)
model, record = train_and_test(model, loss_func, optimizer, epochs, early_stop)
torch.save(model.state_dict(), './Best_ResNeXt50.pth')
record = np.array(record)
plt.plot(record[:, 0:2])
plt.legend(['Train Loss', 'Valid Loss'])
plt.xlabel('Epoch Number')
plt.ylabel('Loss')
plt.ylim(0, 1.5)
plt.savefig('Loss_J3_1.png')
plt.show()
plt.plot(record[:, 2:4])
plt.legend(['Train Accuracy', 'Valid Accuracy'])
plt.xlabel('Epoch Number')
plt.ylabel('Accuracy')
plt.ylim(0, 1)
plt.savefig('Accuracy_J3_1.png')
plt.show()
四、运行结果
模型打印
Bottleneck-49 [-1, 512, 29, 29] 0
Conv2d-50 [-1, 512, 29, 29] 262,144
BatchNorm2d-51 [-1, 512, 29, 29] 1,024
ReLU-52 [-1, 512, 29, 29] 0
Conv2d-53 [-1, 512, 29, 29] 2,359,296
BatchNorm2d-54 [-1, 512, 29, 29] 1,024
ReLU-55 [-1, 512, 29, 29] 0
Conv2d-56 [-1, 512, 29, 29] 262,144
BatchNorm2d-57 [-1, 512, 29, 29] 1,024
ReLU-58 [-1, 512, 29, 29] 0
Bottleneck-59 [-1, 512, 29, 29] 0
Conv2d-60 [-1, 1024, 15, 15] 524,288
BatchNorm2d-61 [-1, 1024, 15, 15] 2,048
Conv2d-62 [-1, 512, 29, 29] 262,144
BatchNorm2d-63 [-1, 512, 29, 29] 1,024
ReLU-64 [-1, 512, 29, 29] 0
Conv2d-65 [-1, 512, 15, 15] 2,359,296
BatchNorm2d-66 [-1, 512, 15, 15] 1,024
ReLU-67 [-1, 512, 15, 15] 0
Conv2d-68 [-1, 1024, 15, 15] 524,288
BatchNorm2d-69 [-1, 1024, 15, 15] 2,048
ReLU-70 [-1, 1024, 15, 15] 0
Bottleneck-71 [-1, 1024, 15, 15] 0
Conv2d-72 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-73 [-1, 1024, 15, 15] 2,048
ReLU-74 [-1, 1024, 15, 15] 0
Conv2d-75 [-1, 1024, 15, 15] 9,437,184
BatchNorm2d-76 [-1, 1024, 15, 15] 2,048
ReLU-77 [-1, 1024, 15, 15] 0
Conv2d-78 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-79 [-1, 1024, 15, 15] 2,048
ReLU-80 [-1, 1024, 15, 15] 0
Bottleneck-81 [-1, 1024, 15, 15] 0
Conv2d-82 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-83 [-1, 1024, 15, 15] 2,048
ReLU-84 [-1, 1024, 15, 15] 0
Conv2d-85 [-1, 1024, 15, 15] 9,437,184
BatchNorm2d-86 [-1, 1024, 15, 15] 2,048
ReLU-87 [-1, 1024, 15, 15] 0
Conv2d-88 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-89 [-1, 1024, 15, 15] 2,048
ReLU-90 [-1, 1024, 15, 15] 0
Bottleneck-91 [-1, 1024, 15, 15] 0
Conv2d-92 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-93 [-1, 1024, 15, 15] 2,048
ReLU-94 [-1, 1024, 15, 15] 0
Conv2d-95 [-1, 1024, 15, 15] 9,437,184
BatchNorm2d-96 [-1, 1024, 15, 15] 2,048
ReLU-97 [-1, 1024, 15, 15] 0
Conv2d-98 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-99 [-1, 1024, 15, 15] 2,048
ReLU-100 [-1, 1024, 15, 15] 0
Bottleneck-101 [-1, 1024, 15, 15] 0
Conv2d-102 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-103 [-1, 1024, 15, 15] 2,048
ReLU-104 [-1, 1024, 15, 15] 0
Conv2d-105 [-1, 1024, 15, 15] 9,437,184
BatchNorm2d-106 [-1, 1024, 15, 15] 2,048
ReLU-107 [-1, 1024, 15, 15] 0
Conv2d-108 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-109 [-1, 1024, 15, 15] 2,048
ReLU-110 [-1, 1024, 15, 15] 0
Bottleneck-111 [-1, 1024, 15, 15] 0
Conv2d-112 [-1, 2048, 8, 8] 2,097,152
BatchNorm2d-113 [-1, 2048, 8, 8] 4,096
Conv2d-114 [-1, 1024, 15, 15] 1,048,576
BatchNorm2d-115 [-1, 1024, 15, 15] 2,048
ReLU-116 [-1, 1024, 15, 15] 0
Conv2d-117 [-1, 1024, 8, 8] 9,437,184
BatchNorm2d-118 [-1, 1024, 8, 8] 2,048
ReLU-119 [-1, 1024, 8, 8] 0
Conv2d-120 [-1, 2048, 8, 8] 2,097,152
BatchNorm2d-121 [-1, 2048, 8, 8] 4,096
ReLU-122 [-1, 2048, 8, 8] 0
Bottleneck-123 [-1, 2048, 8, 8] 0
Conv2d-124 [-1, 2048, 8, 8] 4,194,304
BatchNorm2d-125 [-1, 2048, 8, 8] 4,096
ReLU-126 [-1, 2048, 8, 8] 0
Conv2d-127 [-1, 2048, 8, 8] 37,748,736
BatchNorm2d-128 [-1, 2048, 8, 8] 4,096
ReLU-129 [-1, 2048, 8, 8] 0
Conv2d-130 [-1, 2048, 8, 8] 4,194,304
BatchNorm2d-131 [-1, 2048, 8, 8] 4,096
ReLU-132 [-1, 2048, 8, 8] 0
Bottleneck-133 [-1, 2048, 8, 8] 0
AvgPool2d-134 [-1, 2048, 1, 1] 0
Linear-135 [-1, 2] 4,098
Softmax-136 [-1, 2] 0
================================================================
Total params: 118,185,090
Trainable params: 118,185,090
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 363.40
Params size (MB): 450.84
Estimated Total Size (MB): 814.81
----------------------------------------------------------------
在25层达到最优模型,35层停止