Resnet与残差网络
增加层的缺陷
在到达一定深度后加入更多层,模型可能产生梯度消失或爆炸问题
可以通过更好的初始化权重,添加BN层等解决
现代架构,试图通过引入不同的技术来解决这些问题,如残差连接
ResNet
ResNet是一种残差网络,可以理解为是一个模块,这个模块经过堆叠可以构成一个很深的网络
ResNet通过增加残差链接(shortcut connection),显示地址让网络中的层拟合残差映射(residual mapping)
[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-Vo01I4e5-1665630372555)(C:\Users\Doctor Jin\AppData\Roaming\Typora\typora-user-images\image-20221012171527395.png)]
ResNet不再尝试学习x到H(x)的潜在映射,而是学习两者之间的不同,或者说是残差(residual)
然后,为了计算H(x),可以将残差加到输入上。假设参残差是F(x) = H(x)-x,我们将尝试学习F(x)+x,而不是直接学习H(x)。
** 由于ResNet最后一步是由加法操作进行的,是在元素级别执行的,所以输入和输出的大小要一致,如果不同的话要进行填充
ResNet中没有polling层,降采样是通过conv的stride实现的(卷积核滑动的距离)
ResNet网络结构的特点
通过Average Poo得到最终的特征,而不是通过全连接层
每隔卷积层之后都紧接着BatchNorm层
class ResnetbasicBlock(nn.Module):
def __init__(self,in_channels,out_channels):
super(),__init__()
self.conv1 = nn.Conv2d(in_channels,
out_channels,
kernel_size = 3,
padding=1,#padding在图片上下左右都会进行填充
#因为最后要进行加法操作
bias = False)#bias是偏置,当作批标准化过程时偏置会被抵消
self.bn1 = nn.BatchNorm2d(out_channels)
self.conv2 = nn.Conv2d(out_channels,
out_channels,
kernel_size = 3,
padding=1,
bias = False)
self.bn2 = nn.BatchNorm2d(out_channels)
def forward(self,x):
residual = x
out = self.conv1(x)
out = F.relu(self.bn(out),inplace = True)
out = self.conv2(out)
out = self.bn2(out)
out+=residual #最终要的加法操作
return F.relu(out)
Densenet
Densent模型 引入了每层与所有后续层的连接,即每一层都接受所有前置层的特征平面作为输入
需要特征图大小保持一致,在Densent网络中使用DenseBlock+Transition结构 (其结果就是多个Densenet中间通过poling连接降低特征图大小)
DenseBlock
是包含很多层的模块,每个层的特征图大小相同,层与层采用密集连接方式
Transition
Transition模块是连接两个相邻的DenseBlock,并通过Pooling是特征图大小降低
实例 利用DenseNet提取特征(小鸟数据集)
如果数据都在一个文件夹中可以使用
torchvision.datasets.ImageFolder
但是当数据分散在不同问价夹时就需要用glob来提取路径信息
imgs_path = glob.glob(r'birds/*/*.jpg')
对路径数据进行切片,观察其命名规律选择labels标签
imgs_path[:2]
利用列表split切片推导式获取all_labels_names
all_labels_names = [img_p.split('\\')[1].split('.')[1] for img_p in imgs_path]
unique方法获取类的种类
unique_labels = np.unique(all_labels_names)
len(unique_labels)
使用字典方法对路径创建index
label_to_index = dict((v,k)for k,v in enumerate(unique_labels))
创建all_labels(使所有的图片都有index,上面的方法仅仅是针对不同的类创建了index)
all_labels = [label_to_index.get(name)for name in all_labels_names]
利用随机数据区分train和test数据
np.random.seed(2021) #将随机数确定
random_index = np.random.permutation(len(imgs_path))
imgs_path = np.array(imgs_path)[random_index]
all_labels = np.array(all_labels)[random_index]
i = int(len(imgs_path)*0.8)
train_path = imgs_path[:i]
train_label = all_labels[:i]
test_path = imgs_path[i:]
test_label = all_labels[i:]
创建dataset类
transform = transforms.Compose([
transforms.Resize((224,224)),
transforms.ToTensor()
])
class BirdsDataset(data.Dataset):
def __init__(self,imgs_path,labels):
self.imgs = imgs_path
self.labels = labels
def __getitem__(self,index):
img = self.imgs[index]
label = self.labels[index]
pil_img = Image.open(img)
pil_img = pil_img.convert('RGB')
img_tensor = transform(pil_img)
return img_tensor,label
def __len__(self):
return len(self.imgs)
画图
img_batch,label_batch = next(iter(train_dl))
plt.Figure(figsize=(30,30))
for i, (img,label) in enumerate(zip(img_batch[:6],label_batch[:6])):
img = img.permute(1,2,0).numpy()
plt.subplot(3,2,i+1)
plt.title(index_to_label.get(label.item()))
plt.imshow(img)
使用DenseNet提取特征
my_densenet = torchvision.models.densenet121(pretrained=True).features #仅仅使用features部分对数据的特征进行提取
if torch.cuda.is_available():
my_densenet = my_densenet.cuda()
不改变DenseNet中已经训练好的parameters
for p in my_densenet.parameters():
p.requires_grad = False
提取特征
train_features = []
train_feat_labels = []
for im,la in train_dl:
out = my_densenet(im.cuda())
out = out.view(out.size(0),-1)
train_features.extend(out.cpu().data)#append 和 extend的区别
train_feat_labels.extend(la)
test_features = []
test_feat_labels = []
for im,la in test_dl:
out = my_densenet(im.cuda())
out = out.view(out.size(0),-1)
test_features.extend(out.cpu().data)# 数据要在cpu上进行加减(数据类型的不同)
test_feat_labels.extend(la)
将得到的特征和label作为新的dataset再用线性全裂阶层进行拟合
class FeatureDataset(data.Dataset):
def __init__(self,feat_list,label_list):
self.feat_list = feat_list
self.label_list = label_list
def __getitem__(self,index):
return self.feat_list[index],self.label_list[index]
def __len__(self):
return len(self.feat_list)
train_feat_ds = FeatureDataset(train_features,train_label)
test_feat_ds = FeatureDataset(test_features,test_label)
train_feat_dl = data.DataLoader(train_feat_ds,batch_size=32,shuffle=True)
test_feat_dl = data.DataLoader(test_feat_ds,batch_size=32)
看train_features的维度,因为要做线性拟合
in_feat_size = train_features[0].shape[0]
in_feat_size
nn.Linear的创建
class FCModel(torch.nn.Module):
def __init__(self,in_size,out_size):
super().__init__()
self.lin1 = nn.Linear(in_size,out_size)
def forward(self,input):
return self.lin1(input)
初始化
net = FCModel(50176,200)
if torch.cuda.is_available():
net.to('cuda')
fit函数
def fit(epoch, model, trainloader, testloader):
correct = 0
total = 0
running_loss = 0
model.train()
for x, y in train_feat_dl:
if torch.cuda.is_available():
y = torch.tensor(y,dtype = torch.long)
x, y = x.to('cuda'), y.to('cuda')
y_pred = net(x)
loss = loss_fn(y_pred, y)
optim.zero_grad()
loss.backward()
optim.step()
with torch.no_grad():
y_pred = torch.argmax(y_pred, dim=1)
correct += (y_pred == y).sum().item()
total += y.size(0)
running_loss += loss.item()
epoch_loss = running_loss / len(trainloader)
epoch_acc = correct / total
test_correct = 0
test_total = 0
test_running_loss = 0
model.eval()
with torch.no_grad():
for x, y in test_feat_dl:
if torch.cuda.is_available():
y = torch.tensor(y,dtype = torch.long)
x, y = x.to('cuda'), y.to('cuda')
y_pred = net(x)
loss = loss_fn(y_pred, y)
y_pred = torch.argmax(y_pred, dim=1)
test_correct += (y_pred == y).sum().item()
test_total += y.size(0)
test_running_loss += loss.item()
epoch_test_loss = test_running_loss / len(testloader)
epoch_test_acc = test_correct / test_total
print('epoch: ', epoch,
'loss: ', round(epoch_loss, 3),
'accuracy:', round(epoch_acc, 3),
'test_loss: ', round(epoch_test_loss, 3),
'test_accuracy:', round(epoch_test_acc, 3)
)
return epoch_loss, epoch_acc, epoch_test_loss, epoch_test_acc
epoches = 30
train_loss = []
train_acc = []
test_loss = []
test_acc = []
for epoch in range(epoches):
epoch_loss,epoch_acc,epoch_test_loss,epoch_test_acc = fit(epoch,net,train_dl,test_dl)
train_loss.append(epoch_loss)
train_acc.append(epoch_acc)
test_loss.append(epoch_test_loss)
test_acc.append(epoch_test_acc)