为完成自己的科研任务,当前我需要基于VGG16做fine-tuning。于是写下这一节笔记。
我使用的是torch1.0,因此本博客主要基于这篇博客——pytorch finetuning 自己的图片进行行训练做调整
import torch
import torchvision
# prepare model
mode1_vgg16 = torchvision.models.vgg16(pretrained=True)
pretrained 设置为 True,会自动下载模型所对应权重,并加载到模型中
也可以自己下载权重,然后 load 到 模型中,源码中有权重的地址:https://github.com/aaron-xichen/pytorch-playground。
假设我们的类别数为2(即需要修改最后一层全连接层的神经元数量)
先来看看vgg16最后几层
与Resnet(分类器只有一个名为fc的全连接层)不同,vgg16的分类器(classifier)由多层组成,而我们要修改的仅为最后一层的输出1000
因此在,第一步的基础上添加以下代码
#读取输入特征的维度(因为这一维度不需要修改)
num_fc = mode1_vgg16.classifier[6].in_features
#修改最后一层的输出维度,即分类数
mode1_vgg16.classifier[6] = torch.nn.Linear(num_fc,2)
在先前的基础上添加以下代码
#对于模型的每个权重,使其不进行反向传播,即固定参数
for param in mode1_vgg16.parameters():
param.requires_grad = False
#但是参数全部固定了,也没法进行学习,所以我们不固定最后一层,即全连接层
for param in mode1_vgg16.classifier[6].parameters():
param.requires_grad = True
注:可以通过以下代码查看权值数值(可加可不加)
for child in mode1_vgg16.children():
print(child)
for param in child.parameters():
print(param)
未固定的权重会有requires_grad=True
,而被固定的权值为requires_grad=False
默认不显示
先在代码开头补上接下来需要使用的模块
import numpy as np
from torchvision import transforms,utils
from torch.utils.data import DataLoader
from torch.autograd import Variable
1) 这种读取图片的方式用的是torch自带的 ImageFolder,读取的文件夹必须在一个大的子文件下,按类别归好类。就像我现在要区分2个类别
2) 随后用代码导入数据
#train data
train_data=torchvision.datasets.ImageFolder('/disk2/lockonlxf/pin/trainData',transform=transforms.Compose(
[
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor()
]))
train_loader = DataLoader(train_data,batch_size=20,shuffle=True)
#test data
test_data = torchvision.datasets.ImageFolder('/disk2/lockonlxf/pin/testData',transform=transforms.Compose(
[
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor()
]))
test_loader = DataLoader(test_data,batch_size=20,shuffle=True)
1) 定义自己的优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(mode1_vgg16.parameters(),lr=0.001)
2) 训练模型
EPOCH = 1
for epoch in range(EPOCH):
train_loss = 0.
train_acc = 0.
for step, data in enumerate(train_loader):
batch_x, batch_y = data
batch_x, batch_y = Variable(batch_x), Variable(batch_y)
# batch_y not one hot
# out is the probability of eatch class
# such as one sample[-1.1009 0.1411 0.0320],need to calculate the max index
# out shape is batch_size * class
out = mode1_vgg16(batch_x)
loss = criterion(out, batch_y)
train_loss += loss.item()
# pred is the expect class
# batch_y is the true label
pred = torch.max(out, 1)[1]
train_correct = (pred == batch_y).sum()
train_acc += train_correct.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
if step % 100 == 0:
print('Epoch: ', epoch, 'Step', step,
'Train_loss: ', train_loss / ((step + 1) * 20), 'Train acc: ', train_acc / ((step + 1) * 20))
print('Epoch: ', epoch, 'Train_loss: ', train_loss / len(train_data), 'Train acc: ', train_acc / len(train_data))
mode1_vgg16.eval()
eval_loss=0
eval_acc=0
for step ,data in enumerate(test_loader):
batch_x,batch_y=data
batch_x,batch_y=Variable(batch_x),Variable(batch_y)
out = mode1_vgg16(batch_x)
loss = criterion(out, batch_y)
eval_loss += loss.item()
# pred is the expect class
# batch_y is the true label
pred = torch.max(out, 1)[1]
test_correct = (pred == batch_y).sum()
eval_acc += test_correct.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
print( 'Test_loss: ', eval_loss / len(test_data), 'Test acc: ', eval_acc / len(test_data))