2024年夏季《深度学习》学习报告 -第四周
姓名 | *** |
---|---|
学号 | *** |
课程 | 深度学习基础-算法与实践 |
内容 | |
博客地址 |
一、学习目标
●MobileNet V1 & V2 &V3
● ShuffleNet+SENet&CBAM
二、学习内容
1.理论内容
● MobileNet
问题:由于传统卷积神经网络内存需求大,运算量大,导致无法在移动设备以及嵌入式设备上运行。
解决:MobileNet网络是轻量级CNN网络,相比于传统CNN,在准确率小幅降低的前提下大大减少模型参数与运算量。
V1版本:(1)通过DW卷积+PW卷积操作来大大减少参数和运算量。经过这两个操作得到的特征矩阵的个数与传统卷积是相等的,但参数和运算量减少,理论上普通卷积计算量是DW+PW的9倍。
(2)引入超参数α与β,α目的是让每一层的卷积更瘦,β是用来减少每一层输出的特征图大小的,通过减小特征图的分辨率来降低模型所需要的计算量。
V2版本:(1)Inverted Residuals(倒残差结构),具体是先用11的卷积进行升维,接着用33的卷积DW,最后用1*1的卷积降维。与传统的残差结构刚好相反,同时该结构中的激活函数采用ReLU6,减少低维特征的信息损失。(2)Linear Bottlenecks,ReLU会损失很多低维信息,所以在输入中用线性的激活函数来代替ReLU拿到更多的低维特征向量,避免信息的丢失。
V3版本:(1)更新Block,引入了SE模块,即注意力机制,同时更新了激活函数(2)使用NAS搜素参数(3)重新设计耗时层结构
● ShuffleNet
虽然分组卷积能够减少参数与计算量,但分组卷积中不同组之间没有信息交流,神经网络的表达能力大大降低。ShuffleNet引入通道洗牌操作,来解决该问题。
通道洗牌允许分组卷积从不同组中获取输入数据,从而实现输入通道和输出通道相关联。
● SENet & CBAM
注意力机制:理解为资源分配机制,对重要的特征施加更大的权重,使得网络关注到更重要的特征,提高了人处理图像的效率,准确性和速度。
SENet:(1)Sqneeze:全剧平均池化用来压缩特征矩阵
(2)Excitation:两层全连接+Sigmoid激活函数,得到每 个通道的权重
(3)特征重标定:将权重与原先的特征矩阵分别相乘
CBAM: 对SENet的改进,增加了通道和空间维度,进一步加强了注意力的表示
2.实验部分
HybridSN 高光谱分类
! wget http://www.ehu.eus/ccwintco/uploads/6/67/Indian_pines_corrected.mat
! wget http://www.ehu.eus/ccwintco/uploads/c/c4/Indian_pines_gt.mat
! pip install spectral
//导入基本库
import numpy as np
import matplotlib.pyplot as plt
import scipy.io as sio
from sklearn.decomposition import PCA
from sklearn.model_selection import train_test_split
from sklearn.metrics import confusion_matrix, accuracy_score, classification_report, cohen_kappa_score
import spectral
import torch
import torchvision
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
class_num = 16
windowSize = 25
K = 30 #参考Hybrid-Spectral-Net
rate = 16
class HybridSN(nn.Module):
#定义各个层的部分
def __init__(self):
super(HybridSN, self).__init__()
self.S = windowSize
self.L = K;
#self.conv_block = nn.Sequential()
## convolutional layers
self.conv1 = nn.Conv3d(in_channels=1, out_channels=8, kernel_size=(7, 3, 3))
self.conv2 = nn.Conv3d(in_channels=8, out_channels=16, kernel_size=(5, 3, 3))
self.conv3 = nn.Conv3d(in_channels=16, out_channels=32, kernel_size=(3, 3, 3))
#inputX经过三重3d卷积的大小
inputX = self.get2Dinput()
inputConv4 = inputX.shape[1] * inputX.shape[2]
# conv4 (24*24=576, 19, 19),64个 3x3 的卷积核 ==>((64, 17, 17)
self.conv4 = nn.Conv2d(inputConv4, 64, kernel_size=(3, 3))
#self-attention
self.sa1 = nn.Conv2d(64, 64//rate, kernel_size=1)
self.sa2 = nn.Conv2d(64//rate, 64, kernel_size=1)
# 全连接层(256个节点) # 64 * 17 * 17 = 18496
self.dense1 = nn.Linear(18496, 256)
# 全连接层(128个节点)
self.dense2 = nn.Linear(256, 128)
# 最终输出层(16个节点)
self.dense3 = nn.Linear(128, class_num)
#让某个神经元的激活值以一定的概率p,让其停止工作,这次训练过程中不更新权值,也不参加神经网络的计算。
#但是它的权重得保留下来(只是暂时不更新而已),因为下次样本输入时它可能又得工作了
#参考: https://blog.csdn.net/yangfengling1023/article/details/82911306
#self.drop = nn.Dropout(p = 0.4)
#改成0.43试试
self.drop = nn.Dropout(p = 0.43)
self.soft = nn.Softmax(dim=1)
pass
#辅助函数,求经历过三重卷积后二维的一个大小
def get2Dinput(self):
#torch.no_grad(): 做运算,但不计入梯度记录
with torch.no_grad():
x = torch.zeros((1, 1, self.L, self.S, self.S))
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
return x
pass
#必须重载的部分,X代表输入
def forward(self, x):
#F在上文有定义torch.nn.functional,是已定义好的一组名称
out = F.relu(self.conv1(x))
out = F.relu(self.conv2(out))
out = F.relu(self.conv3(out))
# 进行二维卷积,因此把前面的 32*18 reshape 一下,得到 (576, 19, 19)
out = out.view(-1, out.shape[1] * out.shape[2], out.shape[3], out.shape[4])
out = F.relu(self.conv4(out))
# Squeeze 第三维卷成1了
weight = F.avg_pool2d(out, out.size(2)) #参数为输入,kernel
#参考: https://blog.csdn.net/qq_21210467/article/details/81415300
#参考: https://blog.csdn.net/u013066730/article/details/102553073
# Excitation: sa(压缩到16分之一)--Relu--fc(激到之前维度)--Sigmoid(保证输出为0至1之间)
weight = F.relu(self.sa1(weight))
weight = F.sigmoid(self.sa2(weight))
out = out * weight
# flatten: 变为 18496 维的向量,
out = out.view(out.size(0), -1)
out = F.relu(self.dense1(out))
out = self.drop(out)
out = F.relu(self.dense2(out))
out = self.drop(out)
out = self.dense3(out)
#添加此语句后出现LOSS不下降的情况,参考:https://www.e-learn.cn/topic/3733809
#原因是CrossEntropyLoss()=softmax+负对数损失(已经包含了softmax)。如果多写一次softmax,则结果会发生错误
#out = self.soft(out)
#out = F.log_softmax(out)
return out
pass
import numpy as np
from sklearn.decomposition import PCA
# 对高光谱数据 X 应用 PCA 变换
def applyPCA(X, numComponents):
newX = np.reshape(X, (-1, X.shape[2]))
pca = PCA(n_components=numComponents, whiten=True)
newX = pca.fit_transform(newX)
newX = np.reshape(newX, (X.shape[0], X.shape[1], numComponents))
return newX
# 对单个像素周围提取 patch 时,边缘像素就无法取了,因此,给这部分像素进行 padding 操作
def padWithZeros(X, margin=2):
newX = np.zeros((X.shape[0] + 2 * margin, X.shape[1] + 2* margin, X.shape[2]))
x_offset = margin
y_offset = margin
newX[x_offset:X.shape[0] + x_offset, y_offset:X.shape[1] + y_offset, :] = X
return newX
# 在每个像素周围提取 patch ,然后创建成符合 keras 处理的格式
def createImageCubes(X, y, windowSize=5, removeZeroLabels = True):
# 给 X 做 padding
margin = int((windowSize - 1) / 2)
zeroPaddedX = padWithZeros(X, margin=margin)
# split patches
patchesData = np.zeros((X.shape[0] * X.shape[1], windowSize, windowSize, X.shape[2]))
patchesLabels = np.zeros((X.shape[0] * X.shape[1]))
patchIndex = 0
for r in range(margin, zeroPaddedX.shape[0] - margin):
for c in range(margin, zeroPaddedX.shape[1] - margin):
patch = zeroPaddedX[r - margin:r + margin + 1, c - margin:c + margin + 1]
patchesData[patchIndex, :, :, :] = patch
patchesLabels[patchIndex] = y[r-margin, c-margin]
patchIndex = patchIndex + 1
if removeZeroLabels:
patchesData = patchesData[patchesLabels>0,:,:,:]
patchesLabels = patchesLabels[patchesLabels>0]
patchesLabels -= 1
return patchesData, patchesLabels
def splitTrainTestSet(X, y, testRatio, randomState=345):
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=testRatio, random_state=randomState, stratify=y)
return X_train, X_test, y_train, y_test
import scipy.io as sio
from sklearn.model_selection import train_test_split
# 地物类别
class_num = 16
X = sio.loadmat('/content/drive/MyDrive/app/week5/Indian_pines_corrected.mat')['indian_pines_corrected']
y = sio.loadmat('/content/drive/MyDrive/app/week5/Indian_pines_gt.mat')['indian_pines_gt']
# 用于测试样本的比例
test_ratio = 0.90
# 每个像素周围提取 patch 的尺寸
patch_size = 25
# 使用 PCA 降维,得到主成分的数量
pca_components = 30
print('Hyperspectral data shape: ', X.shape)
print('Label shape: ', y.shape)
print('\n... ... PCA tranformation ... ...')
X_pca = applyPCA(X, numComponents=pca_components)
print('Data shape after PCA: ', X_pca.shape)
print('\n... ... create data cubes ... ...')
X_pca, y = createImageCubes(X_pca, y, windowSize=patch_size)
print('Data cube X shape: ', X_pca.shape)
print('Data cube y shape: ', y.shape)
print('\n... ... create train & test data ... ...')
Xtrain, Xtest, ytrain, ytest = splitTrainTestSet(X_pca, y, test_ratio)
print('Xtrain shape: ', Xtrain.shape)
print('Xtest shape: ', Xtest.shape)
# 改变 Xtrain, Ytrain 的形状,以符合 keras 的要求
Xtrain = Xtrain.reshape(-1, patch_size, patch_size, pca_components, 1)
Xtest = Xtest.reshape(-1, patch_size, patch_size, pca_components, 1)
print('before transpose: Xtrain shape: ', Xtrain.shape)
print('before transpose: Xtest shape: ', Xtest.shape)
# 为了适应 pytorch 结构,数据要做 transpose
Xtrain = Xtrain.transpose(0, 4, 3, 1, 2)
Xtest = Xtest.transpose(0, 4, 3, 1, 2)
print('after transpose: Xtrain shape: ', Xtrain.shape)
print('after transpose: Xtest shape: ', Xtest.shape)
""" Training dataset"""
class TrainDS(torch.utils.data.Dataset):
def __init__(self):
self.len = Xtrain.shape[0]
self.x_data = torch.FloatTensor(Xtrain)
self.y_data = torch.LongTensor(ytrain)
def __getitem__(self, index):
# 根据索引返回数据和对应的标签
return self.x_data[index], self.y_data[index]
def __len__(self):
# 返回文件数据的数目
return self.len
""" Testing dataset"""
class TestDS(torch.utils.data.Dataset):
def __init__(self):
self.len = Xtest.shape[0]
self.x_data = torch.FloatTensor(Xtest)
self.y_data = torch.LongTensor(ytest)
def __getitem__(self, index):
# 根据索引返回数据和对应的标签
return self.x_data[index], self.y_data[index]
def __len__(self):
# 返回文件数据的数目
return self.len
# 创建 trainloader 和 testloader
trainset = TrainDS()
testset = TestDS()
train_loader = torch.utils.data.DataLoader(dataset=trainset, batch_size=128, shuffle=True, num_workers=2)
test_loader = torch.utils.data.DataLoader(dataset=testset, batch_size=128, shuffle=False, num_workers=2)
import torch.nn.functional as F
# 使用GPU训练,可以在菜单 "代码执行工具" -> "更改运行时类型" 里进行设置
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# 网络放到GPU上
net = HybridSN().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=0.00037)
# 开始训练
net.train()
total_loss = 0
for epoch in range(200):
for i, (inputs, labels) in enumerate(train_loader):
inputs = inputs.to(device)
labels = labels.to(device)
# 优化器梯度归零
optimizer.zero_grad()
# 正向传播 + 反向传播 + 优化
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
total_loss += loss.item()
print('[Epoch: %d] [loss avg: %.4f] [current loss: %.4f]' %(epoch + 1, total_loss/(epoch+1), loss.item()))
print('Finished Training')
from sklearn.metrics import classification_report
count = 0
net.eval()
# 模型测试
for inputs, _ in test_loader:
inputs = inputs.to(device)
outputs = net(inputs)
outputs = np.argmax(outputs.detach().cpu().numpy(), axis=1)
if count == 0:
y_pred_test = outputs
count = 1
else:
y_pred_test = np.concatenate( (y_pred_test, outputs) )
# 生成分类报告
classification = classification_report(ytest, y_pred_test, digits=4)
print(classification)
三、问题总结与体会
● 训练HybridSN,然后多测试⼏次,会发现每次分类的结果都不⼀样,请思考为什么?
网络采用dropout技术,其在学习过程中随机丢弃一部分神经元,防止过拟合,同时也增加了不确定性。
● 如果想要进⼀步提升⾼光谱图像的分类性能,可以如何改进?
将注意力机制加在第三个三维卷积后,保留更多的光谱信息,进一步提升分类性能
● depth-wise conv 和 分组卷积有什么区别与联系?
分组卷积:将卷积核分为G组,则该层的参数量减少为原来的1/G
深度分离卷积:本质上就是分组卷积,其在MobileNet中联合点卷积使用
● SENet 的注意⼒是不是可以加在空间位置上?
● 在 ShuffleNet 中,通道的 shuffle 如何⽤代码实现?
def channel_shuffle(x: Tensor, groups: int) -> Tensor:
batchsize, num_channels, height, width = x.size()
channels_per_group = num_channels // groups
# reshape
x = x.view(batchsize, groups, channels_per_group, height, width)
x = torch.transpose(x, 1, 2).contiguous()
# flatten
x = x.view(batchsize, -1, height, width)
return x
● 在 ShuffleNet 中,通道的 shuffle 如何⽤代码实现?
def channel_shuffle(x: Tensor, groups: int) -> Tensor:
batchsize, num_channels, height, width = x.size()
channels_per_group = num_channels // groups
# reshape
x = x.view(batchsize, groups, channels_per_group, height, width)
x = torch.transpose(x, 1, 2).contiguous()
# flatten
x = x.view(batchsize, -1, height, width)
return x