英文:
Transforming an image classification model into a hierarchical model
问题
我正在使用ResNet50来从图像中提取特征。我如何移除网络的分类头部，并且因为我需要拆分网络以获得中间特征，我如何将网络转化为这样一个分层形式，以便我可以像下面这样访问中间网络架构：
import torch
from torchvision.models import resnet50, ResNet50_Weights
import torch.nn as nn
class model(nn.Module):
def __init__(self, pretrained=True):
super(model, self).__init__()
self.featureExtractor = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
def forward(self, x):
x1 = self.featureExtractor.layer1(x) #(number of feature maps:256)
x2 = self.featureExtractor.layer2(x1) # (number of feature maps:512)
x3 = self.featureExtractor.layer3(x2) # (number of feature maps:1024)
x4 = self.featureExtractor.layer4(x3) # (number of feature maps:2048)
return x1, x2, x3, x4
虽然我知道如何使用hook方法提取网络的中间特征，但我不知道如何将网络分割成这样的分层结构。有什么想法吗？
英文:
I am using ResNet50 for extracting features from images. How I can remove the classification head of the network and also because I need to split the network to have intermediate features how I can transform the network into such a hierarchical form that I can access to the intermediate network architecture as follows:
import torch
from torchvision.models import resnet50, ResNet50_Weights
import torch.nn as nn
class model(nn.Module):
def __init__(self, pretrained=True):
super(model, self).__init__()
self.featureExtractor = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
def forward(self, x):
x1= self.featureExtractor_1(x) #(number of feature maps:256)
x2= self.featureExtractor_2(x1) # (number of feature maps:512)
x3= self.featureExtractor_3(x2) # (number of feature maps:1024)
x4= self.featureExtractor_4(x3) # (number of feature maps:2048)
return x1, x2, x3, x4
Although I know how to extract the intermediate features of the network by using hook method I do not know to to split the network into such a hierarchy?
Any idea?
答案1
得分: 1
以下是翻译好的部分：
您可以按照以下方式操作：
import torch
from torchvision.models import resnet50, ResNet50_Weights
import torch.nn as nn
class model(nn.Module):
def __init__(self):
super(model, self).__init__()
self.model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
self.block1 = nn.Sequential(*list(self.model.children())[:5])
self.block2 = nn.Sequential(*list(this.model.children())[5:6])
self.block3 = nn.Sequential(*list(this.model.children())[6:7])
self.block4 = nn.Sequential(*list(this.model.children())[7:8])
def forward(self, x):
x1 = self.block1(x)
x2 = self.block2(x1)
x3 = self.block3(x2)
x4 = self.block4(x3)
return(x1, x2, x3, x4)
x = torch.randn(1, 3, 256, 256)
model = model()
x1, x2, x3, x4 = model(x)
print(x1.shape, x2.shape, x3.shape, x4.shape)
这将产生：
torch.Size([1, 256, 64, 64]) torch.Size([1, 512, 32, 32]) torch.Size([1, 1024, 16, 16]) torch.Size([1, 2048, 8, 8])
英文:
You can do as follows:
import torch
from torchvision.models import resnet50, ResNet50_Weights
import torch.nn as nn
class model(nn.Module):
def __init__(self):
super(model, self).__init__()
self.model = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
self.block1 = nn.Sequential(*list(self.model.children())[:5])
self.block2 = nn.Sequential(*list(self.model.children())[5:6])
self.block3 = nn.Sequential(*list(self.model.children())[6:7])
self.block4 = nn.Sequential(*list(self.model.children())[7:8])
def forward(self, x):
x1 = self.block1(x)
x2 = self.block2(x1)
x3 = self.block3(x2)
x4 = self.block4(x3)
return(x1, x2, x3, x4)
x = torch.randn(1, 3, 256, 256)
model = model()
x1, x2, x3, x4 = model(x)
print(x1.shape, x2.shape, x3.shape, x4.shape)
Which gives:
torch.Size([1, 256, 64, 64]) torch.Size([1, 512, 32, 32]) torch.Size([1, 1024, 16, 16]) torch.Size([1, 2048, 8, 8])
通过集体智慧和协作来改善编程学习和解决问题的方式。致力于成为全球开发者共同参与的知识库，让每个人都能够通过互相帮助和分享经验来进步。
评论